How to Run Commercial Due Diligence on AI and Machine Learning Companies

Deal activity in 2025 and into 2026 has increasingly clustered around technology-related sectors — AI & Machine Learning, SaaS, Big Data, and CloudTech. At the same time, survey data from Mergermarket shows that 47% of dealmakers say technology due diligence has been their top priority over the past twelve months, and 51% now call it the single most burdensome element of the entire review process.

That burden isn't surprising. AI companies are genuinely harder to diligence than traditional software businesses. The moats are more technical, the revenue models are less proven, the customer switching costs are murkier, and the competitive landscape shifts in months, not years. Generic tech DD frameworks — the ones built for evaluating a mid-market SaaS platform — break down when you're trying to assess whether a company's proprietary model actually matters, or whether their "AI-powered" product is a thin wrapper on a foundation model anyone can access.

This guide is a practical framework for investment professionals — PE deal teams, corporate M&A groups, hedge fund analysts, and consultants supporting them — who need to run commercial due diligence on AI/ML companies and get to conviction faster. We'll cover what to assess, how to structure the work, and the specific questions you should be asking experts and customers to separate signal from noise.

Why AI Companies Require a Different DD Playbook

Before we get into the framework, it's worth naming the specific characteristics that make AI/ML targets different from standard software acquisitions:

• The technology moat is often overstated. Many AI companies describe proprietary models, unique training data, or novel architectures. In practice, the gap between a company's model and what's available open-source or through foundation model APIs (OpenAI, Anthropic, Google) is frequently narrower than the pitch deck suggests. Your DD needs to pressure-test whether the technology actually creates defensibility.
• Revenue quality is harder to read. AI companies often have a mix of recurring SaaS revenue, professional services/implementation revenue, and usage-based revenue that scales with consumption. Telling these apart — and understanding which is durable — requires deeper customer-level diligence.
• Customer dependency on the product varies wildly. Some AI tools are deeply embedded in customer workflows. Others are experimental budget items that could be cut in a single quarter. You need to understand where on this spectrum the target sits — and the answer is different for different customer segments.
• Talent concentration risk is acute. A disproportionate share of the value in an AI company often sits with a small number of ML engineers or researchers. If three people leave post-acquisition, the moat may leave with them.
• The competitive landscape is moving faster than in any other sector. A company's differentiation in Q1 can evaporate by Q3 if a foundation model provider ships a competing capability natively.

None of these mean AI companies are bad investments. They mean you need to ask different questions — and source your answers from people closer to the ground truth than the management team.

The Commercial DD Framework for AI/ML Targets

We structure commercial due diligence on AI companies around six workstreams. Each one maps to a core investment question, and each requires targeted primary research — expert interviews, customer calls, and in some cases competitive benchmarking — to answer properly.

1. Technology Moat & Defensibility

Core question: Is the AI/ML technology genuinely differentiated, or is it a feature that can be replicated?

This is where most AI DD starts — and where most of it stays too shallow. Management will walk you through their model architecture, their training data pipeline, and their accuracy benchmarks. That's useful context. But the real question is whether any of it creates lasting competitive advantage.

What to investigate:

• Is the company's core model proprietary, fine-tuned from an open-source base, or a wrapper on a third-party API? Each has radically different defensibility implications.
• What is the company's data advantage? Do they have access to proprietary training data that competitors cannot replicate? How was it sourced, and are there contractual or regulatory risks to continued access?
• How frequently does the model need to be retrained, and what does that cost? Companies with high retraining costs and narrow data pipelines are more fragile than they appear.
• What happens if a major foundation model (GPT, Claude, Gemini) ships a competing capability natively? Has this already happened in adjacent areas?

Key expert call questions:

• "If you were building this product from scratch today, would you build your own model or use [specific foundation model] as a base? Why?"
• "What would it take for a well-funded competitor to replicate this company's core AI capability within 12-18 months?"
• "How much of the product's value comes from the model itself versus the data pipeline, integrations, and workflow around it?"
• "Where does this company's model meaningfully outperform alternatives you've evaluated — and where does it not?"

2. Revenue Quality & Unit Economics

Core question: Is the revenue durable, recurring, and growing for the right reasons?

AI companies often blend revenue streams in ways that inflate apparent quality. Implementation and professional services revenue gets bundled with SaaS subscriptions. Usage-based pricing creates volatility that looks like growth in an up cycle and creates churn risk in a down cycle. Your job in DD is to decompose the revenue and understand what's actually sticky.

What to investigate:

• What percentage of revenue is truly recurring (annual or multi-year contracts) versus usage-based or transactional?
• What does the net revenue retention look like when you strip out price increases and upsells into new product lines? Is the core product expanding within accounts?
• How much professional services or implementation work is required to land a new customer? What does that do to payback periods?
• Are customers paying for outcomes (e.g., cost savings, revenue lift) or paying for access to a tool? Outcome-based pricing signals higher perceived value but also creates more complex retention dynamics.

Key customer call questions:

• "Walk me through how you budgeted for this product. Is it a line item in your software budget, your data/analytics budget, or somewhere else?"
• "If you had to cut 20% of your vendor spend this year, would this product be in the cut list? Why or why not?"
• "Has your usage of this product increased, decreased, or stayed flat over the past 12 months? What drove that?"
• "What would you estimate the ROI of this product has been for your team — and how did you measure that?"

3. Customer Switching Costs & Workflow Embeddedness

Core question: How deeply embedded is this product in the customer's workflow, and what would it take to rip it out?

This is one of the highest-signal areas in AI DD and one of the most underexplored. An AI product that is embedded in a customer's production workflow — processing real data, feeding real decisions, integrated into real systems — is fundamentally stickier than one that sits in a sandbox or proof-of-concept stage.

What to investigate:

• Is the product being used in production or still in pilot/experimentation?
• What systems does it integrate with? How many internal processes depend on its output?
• If the customer wanted to switch to a competitor, how long would the migration take and what would break in the meantime?
• What percentage of the customer base has moved from pilot to production deployment? What's the conversion rate, and what's blocking the ones that haven't converted?

Key customer call questions:

• "How many people on your team interact with this product daily versus weekly versus monthly?"
• "Have you evaluated any alternatives since you deployed this? What prompted that — and what did you conclude?"
• "If this product disappeared overnight, what would you do? What manual process or alternative tool would you fall back to?"
• "How long did it take to get from signed contract to production deployment? What were the biggest blockers?"

4. Competitive Positioning & Market Dynamics

Core question: Where does this company actually sit in the competitive landscape — and is its position improving or eroding?

AI markets are notoriously hard to map because the boundaries are fluid. A company that started as a "computer vision platform" may now compete with a horizontal data analytics vendor, a vertical SaaS player that bolted on AI features, and a foundation model provider's native capabilities — all simultaneously.

What to investigate:

• Who does the company win deals against, and who do they lose to? What determines the outcome?
• Are customers evaluating the company against AI-native competitors, incumbent software vendors adding AI features, or build-it-themselves options?
• What is the company's go-to-market motion? Enterprise sales with long cycles, product-led growth, channel partnerships?
• How do industry analysts and technical evaluators rank this company relative to competitors?

Key expert call questions:

• "When enterprises in [target vertical] are evaluating solutions in this space, which 3-4 vendors consistently make the shortlist?"
• "What's the most common reason a buyer would choose [target company] over [named competitor] — and vice versa?"
• "Is the trend in this market toward best-of-breed AI point solutions or toward platforms? Where does that leave this company in two to three years?"
• "Have you seen any major foundation model providers start to offer capabilities that directly compete with what this company sells?"

5. Talent & Organisational Risk

Core question: Is the value concentrated in a few key people, and what's the retention risk post-acquisition?

In most software acquisitions, talent risk is a secondary concern. In AI acquisitions, it's often a primary one. The difference between a strong and a weak ML team can determine whether the company can continue to improve its product, retrain models on new data, and respond to competitive shifts.

What to investigate:

• How large is the core ML/AI engineering team? How long have they been with the company?
• What is the company's reputation in the ML talent market? Can they attract and retain top-tier researchers?
• Is the IP held in systems and code, or in the heads of a few individuals? Could the product continue to operate and improve if key personnel left?
• What retention mechanisms are in place, and how do they change post-acquisition?

Key expert call questions:

• "How would you rate this company's ML team relative to peers in the space? Are there specific individuals whose departure would materially impact the product?"
• "If this company were acquired, what's the flight risk for the technical team? What would the acquirer need to do to retain them?"
• "How much of this company's model performance comes from ongoing research and iteration versus the initial architecture and training data?"

6. Market Sizing & Growth Trajectory

Core question: Is the addressable market as large as the company claims — and can this company actually capture a meaningful share of it?

Every AI company pitches a TAM that includes the broadest possible definition of "AI spending." Your DD needs to cut through to the serviceable addressable market: the specific customer segments, use cases, and geographies where this company actually competes and can win.

What to investigate:

• What is the realistic SAM — not the aspirational TAM — based on the company's current product, positioning, and go-to-market capability?
• What's the enterprise adoption curve for this category? Are we in early-adopter territory, early majority, or something else?
• What are the expansion levers? New verticals, new geographies, new use cases within existing accounts?
• What regulatory or compliance dynamics could accelerate or constrain adoption?

Key expert call questions:

• "Based on your experience, what percentage of enterprises in [target vertical] have adopted a solution like this — and what's driving the ones that haven't?"
• "What's the typical budget cycle for a purchase like this? Is budget growing, flat, or being scrutinised more closely?"
• "Are there regulatory changes on the horizon — AI governance, data privacy, sector-specific rules — that would impact demand for this product, either positively or negatively?"

Structuring the Research: A Practical Sequence

You don't run all six workstreams simultaneously. The most effective approach follows a deliberate sequence:

1. Week 1 — Desk research and hypothesis formation. Map the competitive landscape, decompose the revenue model from available data, and draft your initial view of where the risks are. Formulate specific hypotheses you need primary research to validate or reject.
2. Week 2 — Expert interviews (technical and market). Run 6-10 calls with ML engineers, data scientists, industry analysts, and former employees of the target and its competitors. Focus on technology defensibility, competitive positioning, and talent quality. These calls will sharpen your customer call discussion guides.
3. Week 3 — Customer interviews and channel checks. Run 8-15 calls with current customers, churned customers, and prospective customers who evaluated but didn't buy. Focus on revenue quality, embeddedness, and switching costs. Supplement with a short quantitative survey if the customer base is large enough.
4. Week 4 — Synthesis and red team. Consolidate findings, update your market model, and pressure-test your conclusions. Identify the two or three issues that matter most to the investment decision and make sure you have definitive evidence on each.

This sequence isn't fixed — deal timelines compress, and sometimes you're running expert and customer calls in parallel. But the principle holds: talk to technical experts before you talk to customers, because the expert calls will tell you what to listen for in customer conversations.

The Red Flags That Kill AI Deals

After running DD on dozens of AI-related targets, certain patterns reliably signal trouble. Watch for:

• High professional services mix that isn't declining. If services revenue is 30%+ and the ratio hasn't improved over the past eight quarters, the product likely isn't self-serve enough to scale efficiently.
• Pilot-to-production conversion below 40%. This suggests the product demos well but doesn't deliver enough value in real-world conditions to justify full deployment.
• Customer concentration in the AI budget line. If the product is funded out of "innovation" or "AI experimentation" budgets rather than operational or departmental budgets, it's vulnerable to cuts.
• No clear answer to "what happens when GPT can do this?" If neither the management team nor independent experts can articulate why the company's product is defensible against foundation model commoditisation, that's a fundamental risk.
• Key-person dependency with no succession plan. If the CTO or head of ML is the only person who fully understands the model architecture, and there's no documentation or team depth, you're buying a fragile asset.

Where Primary Research Makes or Breaks the Decision

Management presentations and data rooms will give you the financial picture and the company's own narrative. Desk research will give you the competitive landscape at a surface level. But neither will answer the questions that actually determine whether the investment works.

Primary research — structured expert interviews, customer calls, and competitive channel checks — is where you find out whether the moat is real, whether customers actually depend on the product, whether the competitive position is improving or deteriorating, and whether the market is as large as the model assumes.

For AI targets specifically, the gap between the company narrative and ground truth tends to be wider than in other sectors. The technology is genuinely hard to evaluate from the outside, customer sentiment is harder to read from NPS scores alone, and competitive dynamics shift fast enough that even recent secondary research can be outdated.

This is exactly the kind of work we do at Woozle Research. We run end-to-end primary research for deal teams — from designing the discussion guides and sourcing the right experts and customers, to conducting the interviews and delivering a finished, actionable research output. No scheduling calls yourself, no synthesising transcripts at midnight. You brief us on the target, and we deliver the answers.

If you're running DD on an AI or ML acquisition target and need primary research support, get in touch.