An Industry Primer on AI Agents for Investment Professionals

1. Executive Summary

1.1. Why AI Agents Matter to Investors

Generative AI chatbots have amazed many with their writing skills. Now, the focus shifts to AI agents that can act independently. These systems can handle complex, multi-step tasks online. For investors, this isn't just a software upgrade; it's a whole new layer of technology. AI agents can greatly enhance productivity, automate complex workflows, and create new business models. Investors must understand this emerging field. It’s also crucial to distinguish AI agents from simple applications to spot companies that will generate significant value.

1.2. The AI Agent Industry in One Paragraph

The AI agent industry includes companies that build, deploy, and manage software capable of sensing its environment, making decisions, and taking action to achieve goals. In venture capital, this market is rapidly growing. Demand for hyper-automation, advances in foundational AI models, and strong venture capital investment drive this growth. Projections estimate the industry will expand from $5.4 billion in 2024 to over $50 billion by 2030, with some estimates reaching $236 billion by 2034. The ecosystem includes foundational model providers, agent development platforms, and application-specific agents in finance, software development, and customer service. The goal is clear: shift from human-assisted processes to fully automated execution for greater efficiency and scale.

1.3. Key Metrics in a Snapshot

The AI agent industry has strong potential to disrupt many sectors. This drives high growth expectations for the market. Market research firms agree on an annual compound growth rate (CAGR) of over 40% until the decade ends. Estimates may differ, but the trend shows fast growth from a small base. By early 2025, the market size is projected between $5.3 billion and $7.9 billion, with a CAGR nearing 46%. A key indicator of this growth is the rapid adoption; by 2028, one-third of all enterprise software applications are expected to include AI capabilities.

1.4. Three Key Structural Characteristics

1. Dependence on Foundational Models and Compute: The industry relies on a concentrated upstream supply chain. Agents' capabilities depend on foundational models from companies like OpenAI, Google, and Anthropic. Also, operational costs are mainly driven by compute expenses, especially from firms like NVIDIA.
2. The Autonomy-Control Paradox: An agent’s value comes from its autonomy, which also poses risks. Companies must balance allowing agents to act independently while ensuring enough supervision and safety measures to prevent mistakes and reputational harm.
3. The Rise of Task-Based Economic Models: Most software firms offer Software as a Service (SaaS) to charge for product access. AI agents introduce new economic models focused on task completion and value delivery. A task-based model bases unit economics on resource consumption, like compute used per API call, unlike the traditional software model for legacy applications.

1.5. Section Roadmap

This primer helps investment professionals understand the AI agent industry. Section 2, Industry Fundamentals, defines key terms and offers background information. Section 3, How the Industry Works, explains the value chain and business models. Section 4, Industry Economics, covers revenue, costs, and profitability. Section 5, Competitive Landscape, highlights key players and examines competitive dynamics. Section 6, External Forces, looks at regulatory, technology, and ESG factors. Section 7, Investment Framework, presents tools for valuation and due diligence. Finally, Section 8, What Drives Returns, discusses major catalysts and risks for investors.

2. Industry Fundamentals

2.1. Definition and Scope

To invest wisely, it's key to define an AI Agent clearly. This helps you tell it apart from similar technologies like chatbots and basic automation tools. An AI Agent is software that works on its own. It understands its digital environment, thinks about tasks, creates a step-by-step plan, and executes it with little human help. What sets an AI Agent apart is its ability to take action towards a goal. In contrast, a chatbot is designed for conversation and answering questions. An AI Agent acts for the user, handling tasks like arranging travel, managing calendars, analysing data, drafting reports, or even coding.

The AI Agent industry can be broken down into three layers:

1. Infrastructure and Foundational Models: This includes those providing intelligence, like OpenAI and Google, as well as the computational power from companies like NVIDIA and AWS.
2. Agentic Platforms and Frameworks: These are the companies that create tools and platforms for building and managing enterprise-ready agents, such as LangChain and Microsoft's framework.
3. Application-Specific Agents: This includes companies that develop agents for specific roles, like customer-service bots or marketing agents, or for certain industries, such as finance, healthcare, or legal.

This primer focuses mainly on the last two categories. We expect to see innovative application models and economic value emerge here. We will not cover basic systems, like rule-based bots or simple conversational agents, as they lack the adaptive intelligence needed for decision-making and do not qualify as "agents."

2.2. Industry Terminology Glossary

• AI Agent: A self-operating software program that plans, perceives its environment, and takes actions to meet specific goals. It can do more than just provide information. 
• Autonomy: The level at which an agent can work without human control. It ranges from human-in-the-loop (requiring approval) to semi-autonomous (with human oversight) to fully autonomous. 
• Foundational Model: A large AI model, like a Large Language Model or Vision-Language Model, trained on extensive datasets. This model acts as the "brain" for an agent (e.g., OpenAI GPT series, Google's Gemini). 
• Orchestration: The process where an agent coordinates its internal parts (like planning, memory, and reasoning) and external tools (like APIs and databases) to complete tasks that may involve multiple steps. This orchestration layer is often the largest part of an agent's proprietary IP. 
• Human-in-the-Loop (HITL): A system where an agent pauses at key points to ask a human for approval before moving forward. This ensures safety and quality in high-stakes situations. 
• Multi-Agent System (MAS): A system where several AI agents work together to solve problems that one agent alone cannot handle. This allows for teamwork, negotiation, and task division. 
• Inference: The process of using a trained AI model to make predictions or generate outputs, like deciding the next action for an agent. Inference costs, based on compute usage, are a major expense of running an agent.
• Tool Use: An agent's ability to select and use external software tools, such as APIs, web browsers, calculators, and code interpreters, to carry out its plans.

2.3. Historical Context: From Chatbots to Autonomous Agents

The idea of autonomous agents has been a long-standing goal for AI researchers. Earlier efforts focused on simple reflex or rule-based agents. However, the industry is now at a turning point, thanks to major advances in large language models that began to gain traction in the early 2020s. Before this shift, mainstream AI applications were mainly advanced chatbots, which were conversational and reactive.

A key moment for the industry came with the launch and rapid adoption of systems like OpenAI's ChatGPT, Google's Bard (now Gemini), and Anthropic's Claude. By early 2025, these platforms had amassed vast user bases, with ChatGPT alone attracting hundreds of millions of weekly users. This remarkable adoption introduced the public to the power of generative AI, setting the stage for even more powerful developments.

The turning point happened when developers began using these generative models not just as text generators or conversational AI, but as reasoning engines within a broader autonomous framework. This approach, referred to as Plan-Act-Observe, allows agents to create plans to achieve goals. For instance, an agent can execute the first step of its plan using available tools, observe the outcome, and adjust its plan until it reaches the goal. Recent advancements from major labs, including OpenAI's autonomous features and Baidu's GenFlow, indicate a clear shift in the industry from conversational AI to agentic AI. This change marks a transition from systems that 'chat' to those that 'act', redefining AI from a mere information tool to a productivity tool.

2.4. Industry Lifecycle Position

The AI agents industry is in the embryonic, or early growth stage. This can be highlighted by several key points:

• Explosive growth from a low base: The market shows extremely high forecast CAGRs, usually between 41% and 46%. This reflects a shift from initial development to early expansion. 
• Rapid technology evolution: The technology foundation, especially foundational models and agent architectures, is changing quickly. There are no established technical standards across the industry. Best practices for building, deploying, and securing agents are also not yet defined. 
• High levels of capital investment and M&A activity: The sector is attracting significant venture capital. For example, Anthropic raised $13 billion in its 2025 funding round. This influx of funding boosts R&D and sparks competition for market share. 
• Emphasis on Product-Market Fit: Many companies are still working to identify commercially viable use cases. Adoption is increasing, but it remains largely experimental as enterprises learn to integrate agents into existing workflows. Gartner predicts that by the end of 2026, 40% of enterprise applications will include task-specific agents, indicating a fast but still early adoption phase. 
• Fragmented and Evolving Competitive Landscape: The market features established incumbents and numerous VC-backed startups. Competitive advantages are still forming, and market share is fiercely contested across the value chain, from computing infrastructure to specific applications. The costs for training and deploying large models can exceed $100 million, creating a significant, though not impossible, barrier to entry in the foundational layer.

3. How the Industry Works

The AI agent industry turns advances in artificial intelligence into economic value by automating complex digital tasks. At the heart of a C-suite is a clear value chain, various evolving business models, and a complex economic profile shaped by a few key suppliers. For investors, grasping this C-suite engine—from computing resources to end-user application features—is vital for spotting solid business models and adding value. The AI agent industry resembles a complex ecosystem rather than a simple software market. Here, an agent's ability to act autonomously drives its value, while reliability serves as the currency of trust.

3.1. The AI Agent Value Chain

The AI agent value chain outlines the full lifecycle of creating, deploying, and maintaining autonomous agents. It can be divided into three layers: Foundational Inputs, Core Agent Development, and Application & Distribution. Each layer has unique technical challenges, economic traits, and competitive dynamics.

Layer 1: Foundational Inputs This upstream layer supplies the essential elements for building AI agents. It involves high capital intensity and power dynamics among a few dominant players. Foundational inputs include:

• Compute Infrastructure: The most crucial and contested resource, relating to hardware (mainly GPU providers like NVIDIA), data centre capacity, and energy use. Access to and cost of compute are vital for training, fine-tuning, and running models at scale.
• Foundational Models: Large-scale models (LLMs, VLMs) act as the "cognitive engine" for agents, offering baseline abilities in natural language processing, understanding, reasoning, and generation. This area is dominated by well-funded labs like OpenAI, Google, and Anthropic, often accessed via APIs.
• Data: Quality, domain-specific data fuels the training and fine-tuning of agents for specialized tasks. Data sources include structured databases, unstructured text, and real-time streams from public and private sources.

Layer 2: Core Agent Development This layer is the "factory floor" where agents are developed and where most agent-native organisations build their intellectual property. Core agent development integrates foundational inputs with proprietary logic to enable autonomous actions. Key activities include:

• Orchestration and Planning: This core logic breaks high-level goals into specific tasks. Developers create a planning engine that decides which tools to use, what information to gather, and the order of tasks. Orchestrating multiple sub-agents for complex workflows offers significant efficiencies.
• Tool Integration and Actions: Agents interact with the world through APIs, browsers, and various software environments. This development builds strong integrations that allow agents to perform actions like sending emails, accessing databases, or making purchases.
• Memory and Learning: This component helps agents retain context, learn from past interactions, and improve over time. Memory and learning features enable personalisation, allowing agents to move beyond simple, stateless tasks.

Layer 3: Application & Distribution This downstream layer focuses on how agents’ functionality reaches users and integrates into existing workflows. Here, the agents' value is realised.

• Enterprise Integration: For agents to work effectively, they must fit into business processes and software platforms, like customer relationship management (CRM) and enterprise resource planning (ERP). This often requires significant changes to existing systems, shifting from static API-based setups to dynamic, event-driven architectures.
• User Interface (UI) and User Experience (UX): This is how users assign tasks to agents, track their status, and intervene when needed. The UI/UX can range from a simple chat window to a complex dashboard managing multiple agents.
• Governance and Safety: A critical aspect of this layer is implementing controls to ensure the safe, reliable, and ethical use of agents. These controls must be integrated throughout the design, build, and operation stages to reduce risk and build customer trust.

Value in this sector is expected to concentrate in the Core Agent Development layer. Companies that create proprietary orchestration and reasoning engines will have a competitive edge, as they can deliver agents that reliably perform complex actions, rather than simply generating content from prompts.

3.2. Business Model Archetypes

As the AI agent sector grows, it shifts from traditional software as a service (SaaS) pricing. Now, it focuses on models that reflect the value of autonomous work. The main change is moving from charging per human user to pricing based on the agent's output or efficiency. There are three main business model types: subscription, usage-based, and outcome-based. A hybrid approach is the most common.

1. Subscription Models 

This pricing structure comes from SaaS, where customers pay a recurring fee (monthly or annually) for access to an agent platform or a specific number of agents. Per-Seat Pricing: Customers pay for each human user managing the agents. This model is simple but does not reflect the value of agent output, as one agent can replace hundreds of users. Per-Agent Pricing: Customers pay a flat fee for each active AI agent. This approach aligns better with agent value but doesn’t account for varying levels of agent activity. Platform-based solutions often use tiered subscriptions, offering different access levels and higher capacity at premium tiers.

2. Usage-Based Models

 These models charge based on consumption, providing more flexibility and scalability. They are common in API-based offerings where agents are accessed via programming code. Pay-Per-Call: Customers are charged for each API call made to the agent service. This model is transparent but may lead to unpredictable costs. Per-Workflow/Task: Customers pay each time an agent successfully completes an agreed workflow (e.g., processing an invoice or onboarding a new employee). This structure aligns price more closely with discrete units of work. Output: Pricing is based on specific outputs completed by the agent, such as a report or a block of code.

3. Outcome-Based Models 

These models tie price directly to measurable business results from the agent. They represent the ultimate form of "paying for work done" and highlight a clear ROI for customers. For instance, a customer service agent's price might be a percentage of costs saved by deflecting support tickets. A sales agent could earn a commission for leads generated. While outcome-based models can be appealing, they bring complexities, like the need for measurements and attribution regarding the outcomes created by the agent.

4. Hybrid Models

Most organizations are landing on hybrid approaches, that marry the predictability of subscriptions with the value-alignment of variable pricing. A typical structure is a flat subscription fee for access platform access, which includes a certain allowance of usage or outcomes, but with charges for overages. For example, a software company might charge a base per-seat license for its human-users blind, and charge on a per-resolution basis on its AI agent that completes support tickets . This gives vendors a revenue floor while allowing for elastic scaling for customers. The correct model is a key strategic decision based on the agent's purpose, workload patterns, and governance model .

3.3. Unit Economics Deep Dive

The unit economics of AI agents differ greatly from traditional software. While a software product may have almost no marginal cost for each new user, AI agents incur significant variable costs for every task they perform. Misunderstanding these unit economics is a major risk. By 2027, around 40% of AI agent projects may be cancelled due to unclear project value. As an investor, it’s crucial to examine "per-unit" profitability.

At the centre is the outcome-unit. Outcome-units represent one task or workflow successfully completed by an agent. The definition of outcomes is influenced by the business model. Profitability is thus the difference between revenues from outcome-units and total costs.

Cost Side: Total Cost per Outcome The marginal cost of operating an agent includes direct costs from LLM API calls. However, variable costs run deeper and are often higher than expected. A thorough understanding of the cost structure is essential.

• Orchestration Costs: These are direct, variable costs incurred during task completion. Examples include fees per API call to foundational models, charges for specialist tools or sub-agents, and the infrastructure costs for processing. These costs can vary and may lead to unexpected cloud bills if not managed well.
• Operational Infrastructure Costs: Beyond direct compute costs, there are ongoing expenses for the infrastructure supporting the agent's performance. This covers data storage, high-throughput networking for low-latency implementation, caching systems for frequently accessed information, and vector databases for memory and retrieval.
• Quality and Intervention Costs: No agent is flawless. Thus, the cost of goods sold (COGS) must account for quality assurance. Key metrics to monitor include the rework rate (the percentage of tasks needing re-completion due to errors) and the human intervention rate (the percentage of tasks requiring human input to complete). Each intervention raises labour costs and reduces the ROI of the agent's success.

Revenue Side: Revenue per Outcome The revenue per outcome-unit depends on the pricing model. In an outcome-based model, this is clear (e.g. $1 for each resolved ticket). In a usage-based model, it relates to the workflow price. In a subscription model, it’s derived from the total fee divided by the number of completed outcome-units.

Key Unit Economic Metrics Investors should look beyond standard SaaS metrics and focus on agent-specific KPIs:

• Total Orchestration Cost per Outcome: This is the sum of all direct variable costs for completing one unit of work.
• Time-to-Outcome (Latency): The time taken for the agent to finish its task, which is vital for user satisfaction and efficiency.
• Success / Rework Rate: The percentage of tasks completed correctly the first time. A low success rate doesn’t change the cost per successful item but inflates the overall cost per successful item.
• Gross Margin per Successful Outcome: (Revenue per Successful Outcome) - (Cost per Successful Outcome).

Ultimately, the future of an agent-based business hinges on lowering the cost per successful outcome. This can be achieved by improving the agent's success rate while delivering value to ensure a solid gross margin. Economic value arises not when the model is trained, but when the agent is used for scalable, profitable transactions.

3.4. Customer Dynamics and Adoption Trends

The adoption of AI agents in enterprises is speeding up. This growth comes from the potential boosts in productivity and ROI for key business functions. A Google Cloud study from 2025 revealed that 52% of executives have already started using AI agents. These early adopters can measure their success, seeing higher say-do ROI in core areas like customer service (43% for adopters versus 36% average) and security operations (40% for adopters versus 30% average).

AI agents are most commonly used in areas with high volumes of repetitive, data-driven tasks:

• Customer Service and Tech Support: Almost half (49%) of enterprises use agents for customer service tasks, like answering questions, processing returns, and providing tech support (45%).
• Marketing: 46% of customers employ agents for tasks such as personalising campaigns, analysing marketing data, and managing social media.
• Security Operations & Cybersecurity: Similarly, 46% of users have agents monitor threats, analyse logs, and handle incident responses.

While omni-application is common, vertical applications are growing too. For instance, in financial services, 43% use agents for fraud management. In retail and consumer packaged goods (CPG), 39% use them for quality control. This shows that businesses are moving from basic bots to using agents in critical processes. In supply chains, agents can combine complex data across systems, helping to predict and prevent delays.

However, adoption isn’t easy. Trust is the biggest requirement. Customers need to believe in the AI agent's reliability to understand and automate tasks. They are cautious about using fully autonomous systems in high-stakes areas. This leads to phased adoption, starting with human-in-the-loop agents that enhance employee efficiency. Over time, as the technology proves its value, businesses shift to more autonomous applications. Phased adoption often means investing time and resources to redesign enterprise platforms, ensuring they are safe for agents to operate.

3.5. Supplier Dynamics: Compute is at the Centre

The AI agent market is shaped by a few key suppliers. These include upstream suppliers to agents, with AI holding significant power. This creates both opportunities and dependencies for agent developers, affecting their pricing, innovation, and competition. The two main suppliers are compute infrastructure and foundational models.

Compute Infrastructure - The AI market is facing a bottleneck in high-performance compute. This has turned suppliers of core hardware and infrastructure into key players. The battle for NVIDIA's latest GPUs, along with the necessary data centre capacity and energy, is crucial for everyone from large labs to small startups. The scarcity of hardware gives these providers strong pricing power and influences the industry's development timeline. For agent companies, compute is not just a one-time cost for training; it becomes a major recurring expense for inference, which is the main function of the agent.

Foundational Models - The core intelligence engines for agents exist in an oligopoly. This is primarily a two-horse race between Google’s Gemini models and OpenAI’s GPT models. Anthropic, a well-funded firm, is a strong third player, especially in enterprise safety and reliability. This market structure means that most agent developers depend heavily on these few providers for their agents' "brains." Decisions regarding API access, pricing, capabilities, and safety filters dictate what agent companies can create and how much it will cost. This supplier power pushes agent developers to focus on building defensible value in their proprietary orchestration layers, as the underlying intelligence is becoming commoditised, though still expensive.

This reliance on suppliers means agent companies must excel in partnership strategies and technology optimisation. They need to manage risks from price increases or changes in access while working to reduce dependencies. Tactics include using smaller, fine-tuned open-source solutions for specific tasks or building efficient orchestration engines that leverage powerful models only when necessary. Additionally, the demand for specialised AI talent to optimise these processes adds another layer of supply constraint, intensifying competition in the sector.

4. Industry Economics

The economic landscape of the AI agent industry presents a duality. It offers significant value through automation but also introduces new costs. Investors need to navigate this tension. The route to profitability is uncertain and may require new revenue models, fresh operating costs, and scalable unit economics. This section explores the financial aspects of the AI agent market, including revenue sources, unique cost structures, and margin profiles.

4.1. Revenue Generators

In the AI agent industry, revenue stems from the measurable value agents deliver. Key drivers include increased productivity, reduced operating expenses, and quicker decision-making, which appeal to enterprise customers. As the market matures, companies are moving away from traditional software-as-a-service (SaaS) models. Revenue will increasingly reflect the usage-based dynamics of agents.

The main revenue types are:

• Subscription-based Models: Companies provide tiered access based on agent capabilities for a monthly or annual fee. This model is common in the SaaS world, offering predictable revenue for vendors and costs for customers. Tiers are often based on the number of agents, task volume, or functionality. For instance, a basic plan may focus on task automation, while a premium plan offers advanced reasoning or integration features.
• Utilization/Pay-per-Use Models: This model charges based on consumption, such as the number of API calls, tasks completed, or data processed (e.g., tokens). This pricing aligns with the value customers receive but can lead to cost and revenue variability. High usage periods can quickly escalate costs for customers and pose adoption risks if not managed carefully.
• Performance-Based Models: This emerging model links agent costs to achieving specific outcomes or value delivered. For example, an agent automating sales qualification might earn based on the number of qualified leads generated. While this model aligns well with customer ROI, it requires precise tracking and clear definitions of success.
• Hybrid Models: Many providers are exploring hybrid models to balance predictability with value alignment. Common structures include a flat monthly subscription for platform access, plus usage-based charges for excess consumption. Companies like Intercom use a flat charge per agent, creating a predictable cost structure while recognising the agent as a unique value-creating entity.

Demand is currently strongest in enterprises, where the potential ROI from automating complex workflows is highest. Mid-market companies present a major growth opportunity, especially for agents that integrate with core business systems like CRMs and support dashboards. As costs decrease and usability improves, small and medium-sized businesses (SMBs) and individual professionals will further expand the total addressable market.

4.2. Cost Structure

The AI agent industry has a cost structure that differs from traditional software. While it shares some common SaaS expenses, like sales and marketing (S&M) and research and development (R&D), it also faces significant variable costs tied to computing. These costs fall into two main categories: the one-time expense of training a model and the ongoing cost of inference.

Upfront Investment: Model Training & Development Building a modern foundational model (the "brain" of many advanced agents) is very costly. This high cost is a key barrier to entry in the market's infrastructure layer.

• Training Compute Costs: This is the largest expense. Reports suggest that Google spent around £191 million to train their top-tier model, Gemini Ultra, and GPT-4 training exceeded £100 million. These numbers reflect the costs of running thousands of high-end GPUs, like NVIDIA's H100 chip, for weeks or months. 
• Data Work: Creating and curating the large amounts of high-quality data needed for training is also costly. Estimates for data collection and preparation range from £10,000 to over £50,000 for specialised projects. 
• Human Capital: R&D costs are high due to the salaries of top AI researchers and engineers. Talent is scarce. For example, OpenAI reportedly spent £3 billion on compute in 2024, part of an estimated total spend of £9 billion. This shows that R&D is a major cost beyond raw compute.
• Ongoing Operational Costs: Inference After training, the ongoing cost to run the model(s) for agents (i.e., inference) becomes the most significant operational cost. Unlike traditional software, where output costs approach zero per user, each task an agent performs incurs a real variable compute cost.
• Compute-as-COGS: For AI agent companies, inference costs are part of their Cost of Goods Sold (COGS). GPU compute can account for 40-60% of a new AI startup's technical budget in its first few years. Running a large-scale service may lead to daily expenses in the hundreds of thousands of pounds. 
• Orchestration Cost: The cost of an agent's action includes more than just a single model call; it also covers planning, tool usage, and sub-agent calls. Additional compute expenses from planning and execution add both cost and latency to the agent's action. 
• Ancillary Costs: Beyond basic agent inference, ongoing costs for monitoring, safety alignment, cloud infrastructure management, and engineer efforts to improve models (like distillation and quantization) add both cost and complexity.

Overall, AI agent companies should carefully plan their pricing models. They need to align them with variable cost risks to avoid significant margin loss.

4.3. Margin Analysis and Profitability Metrics

The profitability of AI agent companies depends on their ability to manage high and unpredictable inference costs. Traditional software gross margins, usually 80-90%, are not suitable benchmarks for this industry. Instead, gross margins for AI services face constant pressure from compute costs, which act as variable COGS.

Margins can be very unstable! A small change in user prompts or agent features can raise backend costs significantly. This leads to unpredictable margin forecasts and value compression. Consequently, "FinOps for AI" has become a practice that tracks infrastructure usage by customer and feature to understand the true cost of serving each customer.

At the unit-economic level, a company's profitability hinges on the revenue per task versus the total cost to perform that task. A recent study predicts that up to 40% of AI agent projects could be scrapped by 2027. This is due to companies not grasping the basic economics of value creation. Some firms are running workflows that destroy value by spending more to operate agents than they earn.

Key Profitability Metrics: 

• Gross Margin: Calculated as (Revenue - Inference Costs - Other COGS) / Revenue. This metric shows how well the company prices its services above its marginal delivery costs. 
• Contribution Margin per Task/Customer: Tracks the profitability of a single agent action or customer account after accounting for all variable costs. This is crucial for usage-based pricing strategies.
• Compute Cost as a % of Revenue: This ratio reveals how dependent and efficient the company is regarding its main cost driver. A low, stable percentage indicates a healthier, scalable business model. Many AI services are currently subsidised by providers to encourage user adoption, meaning they don’t reflect the true cost of inference. Sustainable long-term profits require a significant decrease in compute costs, higher prices, or the success of high-value agent applications that can command premium pricing.

4.4. Capital Intensity and Returns

The AI agent sector has high capital intensity, especially at the foundational model layer. Training a large language model incurs nine-figure costs, requiring substantial capital raises. Anthropic's September 2025 funding round raised £13 billion with a post-money valuation of £183 billion. This underscores the capital needed to compete in this space. Capital is essential for both the one-off expense of model training and the ongoing costs of accessing scarce compute resources, often involving multi-billion dollar contracts with cloud providers.

While the capital intensity is lower for application layer companies, it remains significant. They don’t face the same massive upfront R&D costs, but they do incur operational expenses for inference, which rise with customer usage. An early-stage AI startup might spend £10,000-£30,000 per month on GPU infrastructure, while research-focused firms typically spend £50,000 or more monthly.

From the customer’s perspective, a company’s return on investment (ROI) from modifying workflows and integrating AI agents is a key indicator of adoption. A 2022 PwC survey found that 66% of companies using AI agents reported improved productivity, and 57% noted cost savings. These metrics significantly affect the bottom line and justify subscription or usage fees, driving industry growth. An AI agent provider's ability to measure and communicate this ROI is crucial for financial success. High customer returns lead to greater demand, pricing power, and ultimately, better returns for investors in agent product companies.

5. Competitive Landscape

The competitive landscape for AI agents is complex and ever-changing. It's not just one market; it's a system where companies compete on three fronts: foundational models, application-layer workflows, and access to compute infrastructure. This creates a web of competition and cooperation, where firms can be both partners and rivals. Here, advantages come from technology, data, distribution, and domain expertise.

5.1. Market Structure and Concentration

The AI agent market is concentrated at the foundational model and compute layers, showing a clear power-law distribution. It is becoming more fragmented at the application layer.

Foundational Model & Compute Infrastructure Layer: This layer is an oligopoly. A few companies with large capital and research resources control state-of-the-art AI models and the compute needed to run them.

• Model Providers: Top model providers include OpenAI, Google, and Anthropic, delivering the most powerful large language models used in many agents..
• Compute Providers: Nvidia leads as the main supplier of high-performance GPUs for AI training and inference, holding a near-monopoly in this area. The main cloud providers—Amazon AWS, Microsoft Azure, and Google Cloud—distribute this compute capacity. Access to computing power is a key competitive hurdle.

Application Layer: The application layer shows a more fragmented competitive landscape but is moving towards consolidation among strong players in specific niches. Market share figures vary based on the segment measured.

• Consumer Chatbots/Search: In the consumer market, OpenAI's ChatGPT leads with an estimated 61% market share (often including Microsoft Bing and Copilot traffic from OpenAI models) through late 2025 (id:18). Google Gemini follows at around 13-14%, and Claude from Anthropic, though smaller, is growing (id:13, 18).
• Browser-Based Agents: For browser-based automation, Google’s Gemini (Agent Mode) leads with a 57.8% market share, followed by Claude (Computer Use) at 7.0%. This shows Google’s strength in integrating agent capabilities into its user ecosystem.
• Enterprise & Vertical Agents: In the enterprise sector, market share focuses more on key accounts than on public user numbers. Microsoft Copilot has a strong presence, deployed in about 70% of Fortune 500 companies. Many specialized players are carving out niches in consumer and corporate use cases (e.g., customer service, software development) across targeted sectors. North America holds the largest share of the agentic AI market at 46% in 2024, driven by its robust technology sector and significant investment.

5.2. Major Player Profiles

The landscape features Big Tech giants leveraging scale and funding, along with challengers pushing tech boundaries.

• OpenAI: The leader in market share and recognition. Its ChatGPT is the most popular AI assistant. The company is expanding into agentic capabilities with its Assistants API and Agents SDK. OpenAI dominates the consumer AI space and is a strong player in the developer platform sector.
• Google: A significant competitor to OpenAI, many view them as rivals in the AI field. Google boasts strong research assets (DeepMind), the Gemini model family, and unique distribution through Android and Google Search. Gemini Agent Mode has gained traction for browser tasks.
• Microsoft: With a unique strategic position, Microsoft is both a key partner and distributor for OpenAI. It also develops its own AI tools. The Copilot suite is embedded in Microsoft 365 and Azure, giving it extensive access to enterprises.
• Anthropic: A challenger focusing on AI safety and reliability. Its Claude models are popular among professionals for their reasoning skills and consistent results. The company is well-funded, enabling it to compete with top players.
• NVIDIA: The "kingmaker" of the AI industry. As the leading supplier of AI-accelerating GPUs, NVIDIA’s hardware and software ecosystem (CUDA) is essential for the agent industry. Its decisions and product roadmap affect costs and capabilities across the market.
• Specialized and Vertical Players: Beyond the giants, many firms develop agents for specific functions. These include leaders in customer service and sales, as well as niche companies in finance, healthcare, and law. Examples include Adept AI and Leena AI, while tools like Bardeen and AgentGPT target the SMB market with straightforward automation.

5.3. Competitive Dynamics: Models, Applications, and Compute

Competition in the agent AI market unfolds across three connected layers.

1. The Model Layer: At the top, the focus is on who has the best general intelligence. The main competition is between OpenAI's GPT series and Google's Gemini models, which now match or exceed GPT-4 capabilities. They compete on performance benchmarks, multimodality (text, images, etc.), efficiency (cost per token), and safety. While this layer is important, foundational models may become commodities among the top players.
2. The Application Layer: This is where companies build defensible advantages. Success comes from how well they create agent applications tailored to workflows or domains. For instance, an agent for legal contract analysis will compete on its understanding of legal contexts and connections to databases, rather than just its raw intelligence. Winners will solve specific, high-value business problems better than others.
3. The Compute Layer: This layer is a battle for resources. Without sufficient data centre capacity and access to NVIDIA GPUs, companies struggle to compete at the model layer. Many are in a capital arms race to secure compute power, which is vital for training new models and scaling services. This gives larger tech firms an edge due to their resources and cloud infrastructure.

5.4. Barriers to Entry Analysis

Barriers to entry in the AI agent market depend on which stack layer a new entrant targets.

Foundational Model Layer (High to Prohibitive Barriers):

• Capital Investment: As noted in the economics section, building and training a competitive large language model costs hundreds of millions. Only a few entities globally can afford this.
• Access to Compute: Obtaining thousands of specialized GPUs for training poses a logistical and financial challenge due to supply limits and high demand for NVIDIA hardware.
• World-Class Talent: There are very few AI researchers and engineers capable of building and training state-of-the-art models. Their skills are in high demand.
• Data Scale: Pre-training a competitive model requires assembling petabyte-scale, diverse datasets. This task is daunting and favours incumbents with access to costly proprietary data.

Application & Vertical Agent Layer (Moderate to High Barriers): While creating a simple agent wrapper around an existing API is easy, building a market-dominating agent business is much harder.

• Domain Expertise: Creating agents for high-stakes, specialised tasks (e.g., finance or medicine) needs deep domain expertise. This expertise acts as a strong barrier.
• Proprietary Data + Integrations: Access to unique datasets for fine-tuning and deep integration into customer workflows creates high switching costs. Mid-market firms ensure agents fit into existing dashboards and CRMs.
• Distribution + Trust: Gaining access to enterprise customers and building trust for them to delegate critical tasks to AI is a major marketing challenge. This gives incumbents like Microsoft a significant edge.
• Orchestration & Reliability: Ensuring an agent operates reliably in unpredictable environments is a complex engineering challenge. Developing a trustworthy orchestration layer that can plan, execute, and self-correct is a form of defensible intellectual property.

6. External Forces

The AI agent industry will be influenced by external forces, which have the potential to be just as powerful as internal forces. Each investment professional needs to consider the external forces that will affect long-term risk and opportunity: regulation, macroeconomics, technology, and social expectations. External forces will influence the speed of adoption, the constructs of what's acceptable, and will be a contributing factor to which companies and strategies win.

6.1. Regulatory Environment

The regulatory landscape for artificial intelligence (AI) is changing. It is shifting from varied guidelines to a clearer legal framework. This shift is a major force impacting the AI agent industry. The EU is at the forefront with the EU AI Act. Published in the EU Official Journal in July 2024, it will take effect on August 1, 2024. This Act will influence corporate behaviour beyond Europe. A full application is planned for August 2026, but some rules will start sooner. In February 2025, the prohibited AI rules will come into effect. The obligations for "General-Purpose AI" (GPAI) models will begin in August 2025.

The EU AI Act adopts a risk-based approach, creating a tiered system that will shape how AI agents are developed and used. It’s crucial to understand the risks tied to AI applications. For instance, unacceptable risks, like social scoring by public authorities, are banned. In high-risk areas—such as infrastructure, employment, and law enforcement—there will be a "duty of care" for data quality, transparency, human oversight, and robustness. Consequently, AI agents will face rising compliance challenges and costs. For commercial AI agents, transparency is key. Users must be informed when AI is in use. Developers must prioritise safety and compliance; a "fire and forget" approach will not suffice. The GP AI Code of Practice enforces obligations, requiring signatories to submit Safety and Security Model Reports and demonstrate strong internal governance.

For investors, the regulatory impacts are mixed. On one hand, compliance costs and development constraints can delay market entry. This situation may favour larger companies that can navigate the regulatory landscape. Firms will need to establish robust policies. Forming an AI governance committee could help in managing AI use and liability. Regulations will require providers of high-risk systems to report serious incidents to the relevant authorities. On the other hand, clear legislation can reduce investment risks. By laying down clear rules, it can help companies navigate concerns about liability, misuse, and brand damage—barriers to deployment. In this sense, regulation may accelerate the adoption of trusted AI agents. Another key development is the global governance of AI. The US and EU have co-signed the Council of Europe Framework Convention on AI, signalling a move towards global commitment in this area.

6.2. Macroeconomic Sensitivity

The AI agent sector is growing alongside technological advancements. However, we must consider how macroeconomic cycles affect this sector. AI agent engagement is complex, featuring both defensive and cyclical trends. Like other cutting-edge technologies, spending on AI agents is often seen as discretionary. During economic downturns, firms may cut these expenses from their IT budgets to focus on essential operations.

Access to venture capital, which fuels the industry, also depends on macroeconomic conditions and company performance. In 2025, AI startups drew a significant share of VC funding, capturing 34% of global VC capital, even though they represented only 18% of funded firms. This capital-rich environment, characterised by multi-billion funding rounds and "flights to quality," shrinks during recessions. As a result, earlier-stage or less-distinguished companies may struggle to secure growth funding. For example, in Q1-Q3 2025, around 65% of VC funding went to AI, creating a landscape dominated by a few key players.

On the other hand, if AI agents can automate complex tasks, they can boost productivity and reduce costs. This makes them a strong counter-cyclical investment. When firms face pressure to improve efficiency and do more with less, investment in automation technologies will likely increase. AI agents can safeguard their budget allocations, even during austerity measures. Projections indicate robust demand, with enterprise spending expected to grow at a 31.9% CAGR from 2025 to 2029. Enterprises will need to decide whether to view AI agents as essential cost-saving tools or speculative luxuries. This decision will depend on the maturity of AI agents and the return on investment they deliver.

6.3. Technology and Disruption

Technology drives change in the AI agent industry. The fast progress of foundational models shapes the competitive landscape, constantly altering it. As models from developers like OpenAI, Anthropic, and Google improve, the gap between agents that can perform cognitive work and those that cannot shrinks. With easier access to basic agents, the industry is becoming more commoditised. This shifts the focus for developers; a sustainable competitive edge now relies less on powerful models and more on the unique enhancements added to them. Value will lie in the application layer, not in basic frameworks or specialised workflows.

This technological shift poses risks and offers opportunities for solution providers. A major concern is the rise of Artificial General Intelligence (AGI). If AGI develops an all-encompassing intelligence, specialised agent companies could vanish overnight. The timeline for AGI is uncertain, but it looms over long-term investors. In the short term, agent developers face a choice between open-source and proprietary models. Open-source models lower costs and foster healthy ecosystems but provide less protection for solutions providers. Proprietary models offer more control and clearer differentiation but create supplier dependencies and require ongoing R&D investment. Resilient companies will likely adopt model-agnostic technologies, allowing them to leverage the best available models. This focus will help them build proprietary expertise that drives value through unique orchestration, planning, and reasoning engines.

6.4. ESG Considerations: Safety, Ethics, and Governance

Environmental, social, and governance (ESG) considerations for AI agents are crucial for long-term value and risk management. The autonomous design of these agents highlights the importance of these issues as strategic priorities.

Safety and Security: The defining feature of agents is their autonomy, which brings unique safety and security risks. Their complexity can lead to unpredictable behaviour and unintended consequences, known as "over-autonomy." This occurs when an agent acts beyond its intended scope or competence. Such actions can disrupt operations or compromise confidential data. External threats also increase; attackers may use techniques like prompt injection or "agent hijacking" to force agents into taking unauthorised actions, leaking sensitive information, or attacking other systems. Since agents rely on a broad supply chain of services, vulnerabilities from third-party models can quickly affect the agent. For financial regulators like the Bank of England, these risks are real. They worry that highly autonomous agents could exploit weaknesses in algorithmic trading, potentially destabilising the market without human intent.

Ethics and Social Impact: Ethical issues are inherent in AI agent design. AI can create models that unintentionally perpetuate bias, especially in sensitive areas like hiring and lending. Accountability is crucial; who is responsible when an autonomous agent causes financial, physical, or reputational harm? The legal and ethical landscape around ownership and accountability is still evolving. Public accountability is also key, particularly regarding job displacement linked to AI applications.

Governance: In light of these risks, strong governance is essential for sustainability and success. Building trust is vital, and this can only be achieved through consistent safety and ethical practices. Business leaders must implement agents in critical roles with strong guardrails and confidence in their decisions to overcome operational challenges. As proactive examples emerge, companies are forming committees for AI governance and crafting policies that go beyond basic regulatory requirements. For investors, evaluating AI safety measures, model transparency, and ethical governance frameworks is as important as assessing technology and business models. Over time, companies that excel in responsible AI development will gain positive brand equity, reduce regulatory risks, and build trust in deploying agents.

7. Investment Framework

Investing in AI agents needs a special framework. This goes beyond the usual analysis of software companies. The early market stage, foundational technologies, and new economic models add to the complexity of valuing AI agents, doing due diligence, and assessing risk. This framework helps investment professionals navigate the industry's complexity, spot attractive opportunities, and identify major warning signs.

7.1. Valuation Approaches

Valuing companies in the rapidly growing AI agent space is tough. AI agent valuations mix venture capital-style growth forecasting with traditional financial discipline. High valuations stem from a surge in venture funding and excitement about AI's potential. In 2025, AI startups had average valuations 3.2 times higher than traditional tech. A lot of capital is focused on a few perceived leaders, causing valuations to soar. Recently, some companies have doubled or tripled in value within months during funding rounds. For example, a startup in the market research space raised $1 billion in Series A funding, while Sierra, an application layer company, raised $350 million and is now valued at over $10 billion.

Investors need to use a multi-faceted approach:

• Comparable Company and Transaction Analysis. This is the most common valuation method in the private market. Funding rounds often use recent funding of similar companies as benchmarks. While this gives a sense of market sentiment, it can be influenced by hype and may overlook underlying fundamentals. Investors should also examine the deal terms, as structured terms (like liquidity preferences and downside protection) can make headline valuations misleading.
• Total Addressable Market (TAM) and Outcome-Oriented Sizing. A fundamental approach is estimating the economic value of tasks an agent can automate. For example, consider the total global spend on specific workflows targeted by a vertical agent (e.g., legal document review). Valuation depends on projected market share and the take rate on value created. This aligns with the trend of outcome-based pricing strategies.
• Modified SaaS Metrics. AI agents aren’t traditional SaaS companies, but metrics like Annual Recurring Revenue (ARR), customer lifetime value (LTV), and customer acquisition cost (CAC) are good starting points. These metrics need adjusting due to companies' cost structures. For instance, "Gross Margin-Adjusted LTV," which considers variable inference and compute costs, is better for measuring long-term customer profitability.
• Discounted Cash Flow (DCF). For more mature companies or those eyeing the public market, DCF remains the gold standard. However, it carries significant uncertainty. The aim is not to find a single number but to use scenario analyses to see how key variables (like adoption rates and pricing power) affect valuation. DCF encourages discipline in considering the entire model, from revenue sources to free cash flow.

7.2. Industry-Based Analytical Framework

An effective analytical framework for AI agent companies should break down the layers of value creation. This spans from technology and architecture to the company's market strategy. Investors can assess the company using these four pillars:

1. Technology and Architecture. This pillar examines the strength of the company's technical advantage.

• Orchestration and Reasoning Engine. This may be the core IP. Is the ability to plan, manage tasks, or leverage tools a unique asset? Or is it just a basic wrapper around a third-party API? This question is key to understanding technical differentiation.
• Foundational Model Strategy. Are the foundational models dependent on a single provider (e.g., OpenAI) which poses supplier risks? Or does the company use a model-agnostic architecture? This allows for switching providers or using open-source options based on cost and performance.
• Tooling and Integration Ecosystem. An agent’s effectiveness depends on the tools it can use. Assess how well it integrates with existing external APIs, databases, and enterprise applications.
• Learning and Adaptability. Does the agent learn from usage and feedback? A system that adapts over time can offer a strong competitive edge and improve customer retention.

2. Unit Economics and Your Business Model. This pillar examines the financial viability of the company’s solution.

• Cost-per-Outcome Analysis. Consider the total cost to complete a task. This includes orchestration costs, LLM inference fees, external tool charges, vector database queries, and amortised compute costs for proprietary models or fine-tuning.
• Pricing/Commercial Model Alignment. Ensure your pricing model matches the value you provide and your cost structure. Simple per-seat SaaS models can fail if usage varies. Usage-based, outcome-based, or hybrid models can be more sustainable, as they align costs with resource consumption.
• Gross Margin Feasibility. Inference and compute create a new category of COGS, making gross margins different from traditional SaaS. Understanding how to achieve scalable gross margins is vital. This depends on model choice, request batching, and other efficiencies.

3. Go-To-Market and Defensibility. This pillar reviews how the company gains customers and maintains its market position.

• Vertical vs. Horizontal Strategy. Horizontal platforms can address a larger total addressable market (TAM), but face more competition. Vertical-focused companies can build deeper moats through domain expertise, proprietary datasets, and tailored workflows, facing lower replication risks.
• Data Moat. Does the company access unique data generated through its offering? This can create a feedback loop where increased usage boosts agent performance.
• Distribution Channels and Customer Stickiness. How does the company reach its customers? Through direct sales, partners, or larger ecosystems? Once deployed, how integrated is the agent in customer workflows, leading to high switching costs?

4. Governance and Trust. This pillar assesses the company's readiness for the non-technical challenges of deploying autonomous systems.

• Safety and Reliability. What guardrails or validation processes are in place to prevent major errors? Reliability is crucial for enterprise adoption.
• Transparency and Explainability. If the agent makes an error, can the company explain why? Clear visibility into decision-making is vital for debugging, trust, and meeting regulatory needs.
• Regulatory Preparedness. How is the company preparing for regulations like the EU AI Act? A proactive approach to compliance and ethical governance can offer a competitive edge.

7.3. Financial Model Considerations

Creating a financial model for an AI agent company requires understanding its unique revenue and cost drivers.

Revenue Projections Build revenue projections from scratch based on the chosen pricing model. For usage-based models, key drivers include the number of customers, active agents per customer, and task volume processed by each agent. For outcome-based models, revenue hinges on the number of successful outcomes multiplied by the price per outcome. Analysts should consider adoption rates, which are expected to be high; Gartner predicts that 40% of enterprise applications will feature AI agents by 2026.

Cost of Goods Sold (COGS) This area is crucial for the financial model. COGS should be outlined with its variable components, including:

• Third-Party API Costs Fees from foundational model providers, like Anthropic or OpenAI, who usually charge per token.
• Compute Infrastructure Costs for running proprietary models and overall agent orchestration on cloud services (e.g., GPU instances on AWS, Azure, GCP). Distinguish between one-time training costs and ongoing variable inference costs.
• Other Variable Costs Fees for any third-party tools, APIs, or data sources the agent uses to complete tasks.

Margin Analysis Gross Margin (Revenue - COGS) is a key measure of economic viability. Investors should focus on assumptions that could boost gross margin over time, like lower inference costs, increased model efficiency, and better pricing power. Contribution Margin (Revenue - all variable costs) is also vital for assessing potential profitability before fixed overhead. Many agent projects risk cancellation due to misunderstanding these economic factors related to value creation.

Capital Intensity Unlike traditional asset-light SaaS companies, some agent firms may have significant capital needs. This is especially true if they choose to build or pre-purchase substantial compute capacity to ensure performance and secure better pricing. This decision affects returns on invested capital.

7.4. Due Diligence Priorities

Due diligence for investing in an AI agent must be thorough and technically sound. Key priorities include:

1. Technical Deep Dive Go beyond presentations. Conduct live, unscripted product demos where the agent tackles complex, multi-step problems in its target domain. This helps stress-test the agents and assess their planning, reasoning, and tool usage. If possible, have an expert review the agents' architectural design and orchestration layer.
2. Customer Validation Engage with numerous reference customers, especially early adopters. Focus on measurable ROI, including task successes, human intervention frequency, and overall reliability. Also, inquire about the onboarding process and the time to achieve value.
3. Unit Economic Scrutiny This involves detailed financial analysis. Expect breakdowns of cost-per-task or cost-per-customer. Examine contracts and pricing with foundational model and cloud providers. Model unit economics to reflect variations in API pricing and compute costs.
4. Team and Talent Assessment Assess the team’s composition. A successful agent company needs a unique blend of talent, including top AI researchers, strong software engineers for scalable systems, and domain experts in the targeted field.
5. Governance and Risk Review Examine the company’s AI safety, testing, and data governance protocols. Assess their awareness and readiness regarding the evolving regulatory landscape, as this poses significant operational and reputational risks.

7.5. Pattern Recognition Guide

After analysing industry dynamics, investors can create a mental checklist to identify promising opportunities and potential pitfalls.

Bull Case Patterns (Green Lights): 

• Vertical Focus with Proprietary Data: The company targets a high-value vertical (e.g., paralegal support, insurance claims) and effectively uses operational data in a defensible learning loop. 
• Demonstrable and Scalable Unit Economics: Management clearly defines cost-per-outcome and outlines a credible path to profitable scaling. 
• Proprietary Orchestration Layer: The core value is a unique reasoning and planning engine, not just an attractive UI on a generic LLM. 
• High Reliability and Customer Trust: Customer testimonials highlight the agent's reliability and consistent performance with minimal human oversight. 
• Model-Agnostic Architecture: The company is not reliant on a single foundational model provider and can quickly adapt to technological changes.

Bear Case Patterns (Red Flags):

•  "Agent" as Marketing Term: A chatbot or basic workflow tool labelled as an "agent" without true autonomy, reasoning, or planning capabilities. 
• Opaque or Negative Unit Economics: A "grow at all costs" approach with little understanding of profitability at the task level.
•  Competing in Crowded Horizontal Markets: Attempting to create a generic sales or customer service agent without strong, distinct technology or distribution channels. 
• Over-reliance on Single Supplier: The business depends entirely on one foundational model provider, risking exposure to price hikes, API changes, or strategic shifts.
•  Weak Governance and Safety: A dismissive attitude toward safety and ethics places the company at significant reputational and regulatory risk.

8. What is Driving Returns

Investors face challenges in the emerging AI agent market. They must identify key components that drive value and catalysts for growth. Returns depend on both technology and a company's ability to turn that tech into a scalable business model. This section discusses important catalysts, investment theses, and risks affecting financial returns.

8.1. Key Catalysts and Drivers of Stock Price

We expect market-moving events in the AI agent space to emerge at the crossroads of technology, adoption, and regulation. These catalysts will help investors track industry growth and key players.

1. Tech-Related Breakthroughs and Product Cycles. The speed of innovation in foundational models is a key indicator. Competition between top labs, especially between OpenAI and Google, can be measured. Breakthroughs leading to complex agent behaviour, like specialized agents working together, will unlock significant economic value. By 2027, we predict that one in three AI deployments will involve multi-agent scenarios.
2. Enterprise Adoption and ROI Validation. A major indicator of returns will be the widespread use of AI agents in enterprises. Gartner forecasts that by 2026, 40% of enterprise applications will use task-specific AI agents, up from less than 5% in 2025. This rapid adoption shows strong interest in AI within enterprise IT. Additionally, 93% of IT leaders plan to introduce autonomous agents in the next 24 months. Validating ROI through case studies in areas like customer service and marketing will encourage broader market acceptance.
3. Regulatory Clarity and Governance Frameworks. Clear regulations can actually boost adoption. The EU's AI Act, effective from August 2025, will set rules for General-Purpose AI models. This legal framework will give enterprises the confidence to scale. Legal liability, anti-bias measures, and data governance will reduce investment risks and stimulate demand from cautious organisations.

8.2. Common Investment Theories

Investors are diving into the AI agent sector with varied theories, each covering different layers of value and risk.

1. Productivity Supercycle Thesis. This is the broadest thesis. AI agents will replace and enhance complex cognitive tasks, creating a long-term productivity boom. It looks at the wide-ranging effects of agents across industries. Predictions suggest that by 2028, agentic AI will make 15% of daily work decisions. The investment strategy focuses on companies building foundational AI platforms or those excelling at integrating agents into enterprise functions like IT support.
2. Vertical Dominance Thesis. This theory suggests that the highest value will come from companies developing specialized agents for high-value sectors like finance and healthcare. These agents will be trained on proprietary data and integrated into specific workflows, creating competitive advantages. Success stories in fraud management and quality control support this thesis.
3. Infrastructure and Platform Thesis. Like the rise of cloud computing, we believe "picks and shovels" providers will capture significant value in this industry. This includes companies providing essential compute power (e.g., NVIDIA), foundational models (e.g., OpenAI, Google), and development platforms. The thesis hinges on the idea that these infrastructure players will benefit from strong network effects and high switching costs for the ecosystems built on their offerings.

8.3. Risk Factor Inventory

While high returns from the AI agent sector are tempting, there are significant risks that investors must consider.

• Technical and Security Risks: The autonomous nature of agents brings new failure modes. "Over-autonomy" happens when an agent acts beyond its intended scope, leading to harmful outcomes (id:92). Security is a major concern; agents create new attack surfaces, risking prompt injection, data leaks, or hijacking. These issues can lead to unauthorized actions or the exposure of sensitive data (id:97; id:99). The Bank of England warns that advanced agents might exploit market mechanisms, causing financial instability unintentionally (id:92).
• Adoption and Execution Risks: There's a big gap between trying out agents and scaling them for broader use. McKinsey data shows that 62% of organisations are experimenting with agents, but only 23% report scaling them in any business function (id:85). Key barriers include: 1) convincing internal stakeholders, 2) trust in autonomous decisions, 3) complex integration with legacy systems, and 4) fears about liability and brand damage (id:91).
• Regulatory and Compliance Risks: The regulatory landscape is evolving, with various developments worldwide. Fragmentation is common, and it's unclear if regulations will align. For instance, the EU AI Act sorts AI systems into three risk categories (high, limited, low), each with specific obligations (id:62). Navigating different regulations adds costs and slows market entry for newer and smaller firms.
• Economic and Commoditization Risks: AI agents face economic pressures from compute costs. The value created per task must exceed the costs of inference, talent, and safety monitoring. If margins shrink, commoditization risks arise. More foundational model providers may integrate agent capabilities into their platforms, reducing value for independent application-layer companies.

8.4. Risk Monitoring Framework

With all of the risk factors and considerations taken into account, our recommendation is to do a risk analysis of these risks in order to create a systematic approach to monitor our risks as part of managing our portfolio.

9. Reference List

This section presents materials for investment professionals. Some may be familiar, but they can enhance diligence and awareness of the AI agent sector. These resources are selected for deeper exploration in this fast-moving field.

9.1. Data Sources and Market Research

Technology & Venture Capital Trends:

• KPMG Venture Pulse Report: A quarterly report that highlights trends in global VC and funding rounds for AI platform and agent firms.
• PitchBook & CB Insights: Essential for details on private company valuations, funding rounds, investor groups, and M&A activity in the agent and AI sectors.

Market Research & Consulting Firms Reports:

• Gartner: Recognised for insights on AI adoption, market size, and maturity models. Their forecasts help understand enterprise views and timelines.
• McKinsey & Company: Their annual "State of AI" survey offers valuable data on AI adoption, use cases, and challenges across various industries.
• Deloitte: Produces extensive reports, like the "AI Dossier," that compile AI use cases across functions and industries, providing a clear understanding of agent experiences.

Academic & Technical Research:

• arXiv.org: A pre-print repository for academic papers, ideal for assessing new research in agent architecture, algorithms, and safety from top labs.

9.2. Further Reading: Foundational Reports and Papers

Regulatory Frameworks:

• The EU AI Act (Regulation (EU) 2024/1689): The first comprehensive AI law. This document is crucial for understanding compliance and governance that will influence the global market.

AI Safety & Risk Assessment:

• Bank of England Reports on AI Risk: Reports from central banks and financial regulators that discuss systemic risks related to autonomous agents in financial systems.
• McKinsey & Domo Security Reports: These focus on AI security for agents, offering frameworks and insights on new threats like prompt injection and agent hijacking.

9.3. Key Companies and Consortiums to Follow

Foundational Model & Compute Leaders:

• Public: Google (Alphabet), Microsoft, NVIDIA. These firms have significant influence in the ecosystem, providing cloud compute infrastructure and leading research labs like DeepMind.
• Private: OpenAI, Anthropic, Cohere. These well-funded labs drive foundational model development and unique agent technologies.

Venture Backed Agent Specialization Firms:

• Adept: Working on general intelligence to automate software workflows.
• Imbue: Developing large models designed for reasoning and coding to create practical AI agents.
• Sierra: A vertical specialist focused on using AI for customer service agents in enterprises.
• Leena AI: Concentrated on HR and IT service management agents.

Industry and Safety Consortiums:

• Partnership on AI: A non-profit consortium of academia, civil society, and industry partners (including major tech firms) focused on best practices for AI systems.
• AI Safety Institutes: National government-sponsored bodies (in England and the US) tasked with evaluating and setting standards for advanced AI models, shaping future safety specifications.

References

 AI Agents Market Size to Hit USD 236.03 Billion by 2034. https://www.precedenceresearch.com/ai-agents-market

 AI Agents Market Size, Share & Trends - MarketsandMarkets. https://www.marketsandmarkets.com/Market-Reports/ai-agents-market-15761548.html

 Agentic AI Stats 2026: Adoption Rates, ROI, & Market Trends. https://onereach.ai/blog/agentic-ai-adoption-rates-roi-market-trends/

 The AI Power Map: NVIDIA, Google, OpenAI, Anthropic, and … - Reddit. https://www.reddit.com/r/ThinkingDeeplyAI/comments/1pb3x7o/the_ai_power_map_nvidia_google_openai_anthropic/

 AI Agents vs AI Chatbots: Understanding the Difference | CSA. https://cloudsecurityalliance.org/blog/2025/06/16/ai-agents-vs-ai-chatbots-understanding-the-difference

 What are AI agents? Definition, examples, and types | Google Cloud. https://cloud.google.com/discover/what-are-ai-agents

 Understanding the Differences: Chatbots vs. Autonomous Agents in AI. https://www.linkedin.com/pulse/understanding-differences-chatbots-vs-autonomous-ai-kalai-shakrapani-e9zbc

 AI vs Bots vs AI Agents vs Chatbots: Accure Difference. https://www.crescendo.ai/blog/ai-vs-bots-vs-ai-agents-vs-chatbots

 AI Agent Taxonomy: Struggling with AI Agent Classification? Explore … https://www.seriousinsights.net/ai-agent-taxonomy/

 AI Agents: Evolution, Architecture, and Real-World Applications - arXiv. https://arxiv.org/html/2503.12687v1

 The Most Used AI Chatbots in 2025: Global Usage, Trends, and … https://www.datastudios.org/post/the-most-used-ai-chatbots-in-2025-global-usage-trends-and-platform-comparisons-of-chatgpt-gemini

 Top AI Chatbots 2025: ChatGPT, Gemini, Copilot, Claude. https://alphacorp.ai/top-5-ai-chatbots-of-2025-chatgpt-and-the-new-challengers/

 AI Agent Market Forecast to Reach $42.7 Billion by 2030. https://finance.yahoo.com/news/ai-agent-market-forecast-reach-135600682.html

 AI Model Training Costs: How Much Big Tech Spends on AI … https://patentpc.com/blog/ai-model-training-costs-how-much-big-tech-spends-on-ai-development-latest-stats

 Artificial Intelligence in Supply Chain: From Predictive Models to AI … https://www.25madison.com/content/artificial-intelligence-in-supply-chain-from-predictive-models-to-ai-agents

 Agentic AI in the global supply chain - SAP. https://www.sap.com/uk/blogs/agentic-ai-in-global-supply-chain

 How Agentic AI is Transforming Enterprise Platforms | BCG. https://www.bcg.com/publications/2025/how-agentic-ai-is-transforming-enterprise-platforms

 How to Price an AI Agent: Subscription Pricing, Usage-Based … https://saaslogic.io/blog/how-to-price-an-ai-agent

 8 AI Agent Pricing Models Explained - Ema. https://www.ema.co/additional-blogs/addition-blogs/ai-agents-pricing-strategies-models-guide

 How Do API-Based and Platform-Based AI Agent Pricing Models … https://www.getmonetizely.com/articles/how-do-api-based-and-platform-based-ai-agent-pricing-models-differ

 The Future Role of Generative AI in SaaS Pricing | L.E.K. Consulting. https://www.lek.com/insights/tmt/us/ei/future-role-generative-ai-saas-pricing

 Unit Economics Between AI Agents. https://www.linkedin.com/pulse/unit-economics-between-ai-agents-samet-ozkale-pejqf

 Unit Economics of AI. https://charliai.com/wp-content/uploads/2025/05/Charli-AI-Unit-Economics-of-AI-051225.pdf

 Google Cloud Study Reveals 52% of Executives Say Their … https://www.googlecloudpresscorner.com/2025-09-04-Google-Cloud-Study-Reveals-52-of-Executives-Say-Their-Organizations-Have-Deployed-AI-Agents,-Unlocking-a-New-Wave-of-Business-Value,1

 AI agent survey: PwC. https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html

 How SaaS Companies Can Profitably Price AI Agents - CloudZero. https://www.cloudzero.com/blog/ai-agent-pricing-models/

 From Traditional SaaS-Pricing to AI Agent Seats - Research AIMultiple. https://research.aimultiple.com/ai-agent-pricing/

 The Top AI Agents by Market Share – 2025 - First Page Sage. https://firstpagesage.com/seo-blog/the-top-ai-agents-by-market-share/

 What is the cost of training large language models?. https://www.cudocompute.com/blog/what-is-the-cost-of-training-large-language-models

 AI Agent Development Cost: Factors & Pricing Explained. https://appinventiv.com/blog/ai-agent-development-cost/

 The Numbers Behind Training, Inference, and Chat. https://www.linkedin.com/pulse/broken-economics-ai-numbers-behind-training-inference-joy-b--c120c

 How Much Do GPU Cloud Platforms Cost for AI Startups in … https://www.gmicloud.ai/blog/how-much-do-gpu-cloud-platforms-cost-for-ai-startups-in-2025

 The New Economics of AI: Balancing Training Costs and … https://www.finout.io/blog/the-new-economics-of-ai-balancing-training-costs-and-inference-spend

 The Most Overlooked Metric for Sustainable Margins. https://www.mavvrik.ai/cost-to-serve-in-ai/

 I Mapped the Entire AI Agents Market. Here’s What You Need to Know… https://aakashgupta.medium.com/i-mapped-the-entire-ai-agents-market-heres-what-you-need-to-know-0699a79bdc38

 Agentic AI Market Size to Hit USD 199.05 Billion by 2034. https://www.precedenceresearch.com/agentic-ai-market

 17 Best AI Agent Companies Transforming Business in 2025. https://webisoft.com/articles/ai-agent-companies/

 AI Watch: Global regulatory tracker - European Union. https://www.whitecase.com/insight-our-thinking/ai-watch-global-regulatory-tracker-european-union

 EU AI Regulations: Future of AI in Europe | Nemko Digital. https://digital.nemko.com/regulations/eu-ai-act

 AI Act | Shaping Europe’s digital future - European Union. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

 An Introduction to the Code of Practice for General-Purpose AI. https://artificialintelligenceact.eu/introduction-to-code-of-practice/

 White Papers 2024 Understanding the EU AI Act. https://www.isaca.org/resources/white-papers/2024/understanding-the-eu-ai-act

 The Future of AI Agents: Top Predictions and Trends to Watch in 2026. https://www.salesforce.com/uk/news/stories/the-future-of-ai-agents-top-predictions-trends-to-watch-in-2026/

 Are AI Agents Safe? Potential Risks and Challenges - Springs. https://springsapps.com/knowledge/are-ai-agents-safe-potential-risks-and-challenges

 As AI Agents Scale, So Does the Security Risk | Domo. https://www.domo.com/blog/as-ai-agents-scale-so-does-the-security-risk

 Top 100 AI Startup Funding & Investment Statistics . https://www.secondtalent.com/resources/ai-startup-funding-investment/

 AI startup valuations are doubling and tripling within … https://fortune.com/2025/11/29/ai-startup-valuations-are-doubling-and-tripling-within-months-as-back-to-back-funding-rounds-fuel-a-stunning-growth-spurt/

 Aaru hits $1B valuation with multi-tier Series A funding. https://www.techbuzz.ai/articles/aaru-hits-1b-valuation-with-multi-tier-series-a-funding

 Here are the 49 US AI startups that have raised $100M or … https://techcrunch.com/2025/11/26/here-are-the-49-us-ai-startups-that-have-raised-100m-or-more-in-2025/

 Agentic AI security: Risks & governance for enterprises | McKinsey. https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/deploying-agentic-ai-with-safety-and-security-a-playbook-for-technology-leaders

 Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific … https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025

 The State of AI: Global Survey 2025 - McKinsey. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

 AI use cases by industry, function and type. https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/ai-use-cases.html

 Q3’25 Venture Pulse Report — Global trends - KPMG International. https://kpmg.com/sa/en/insights/sector-insights/venture-pulse-q3-2025.html