Commercial due diligence lives or dies on whether the numbers are real. AI is excellent at synthesis and structure, and dangerous at market sizing, unless every figure is sourced.
Key takeaways
AI is used in commercial due diligence for synthesis and structure — turning twenty interview transcripts into a themed findings deck, restructuring a messy customer file into a cohort view, screening public material for competitor moves, and drafting the report. It is not used, or should not be, to produce market sizes and growth rates, because that is where it invents most fluently.
The buyer is asking four things: is the market real and growing, is this company’s position defensible, will the customers stay, and does the plan hold up. Every answer is either evidenced or it is an opinion dressed as one.
A generative tool asked for the size of the Canadian market for a niche industrial product will produce a number, a growth rate and a currency. It will be well-formatted and it may be entirely fabricated. In a report that a buyer will rely on to commit capital, that is the single largest risk the technology introduces.
The discipline that removes it is simple and non-negotiable: every figure in the deliverable carries the source next to it, and someone on the team has opened that source. Where a figure cannot be sourced, the report says what is known and what is not, rather than filling the gap.
The sourcing rule
For Canadian market and industry work, Statistics Canada is the primary source and its tables are addressable and checkable. Its business performance and ownership statistics collection is the usual entry point, and individual tables can be cited directly — for example the table of experimental estimates for business openings and closures, which is one way to characterise churn in a sector without guessing at it. Note the word experimental in that series title: describing a source accurately is part of the work.
Beyond StatCan, the sourceable material is industry association data, regulatory filings, provincial registries, municipal permit data and the target’s own systems. AI can help you find and read these faster. It cannot substitute for them.
Interview synthesis. Twenty customer and channel interviews produce a great deal of unstructured material. Coding it against the hypotheses, pulling the verbatim quotes that evidence each theme, and flagging where interviewees disagree is genuinely faster with a tool — and every quote is checkable against the transcript.
Customer data structuring. Cohort retention, revenue concentration, churn by segment and price realisation over time, from a transaction extract that arrived as an unhelpful spreadsheet. This is the analysis clients value most and it is arithmetic on the target’s own data, so there is nothing to fabricate.
Competitor and channel screening. Reading a large volume of public material for changes in positioning, pricing pages, hiring patterns and product launches, then summarising with links. Insist on links.
Drafting. First drafts of the market overview, the competitive landscape narrative and the executive summary, written from findings the team has already evidenced.
A private buyer is looking at a business-services company operating across three provinces and wants to know whether the growth is the market or the management.
The customer extract is the centre of the work. Structured into cohorts, it shows whether new revenue is coming from new customers or from price increases on existing ones — two very different stories about durability. That analysis is arithmetic on the target’s own data, and a tool that helps clean and pivot it introduces no fabrication risk.
The market section is where discipline is required. The team sizes the addressable market from published Statistics Canada industry data and names the tables used, notes explicitly the segments where no reliable published figure exists, and presents the interview evidence separately as what participants said rather than as fact. The report is more useful for being candid about its limits, and it survives the buyer’s own review, which is the actual test. Where the deal also involves premises or contracts, the legal workstream runs alongside — for instance whether a commercial lease can be assigned.
Customer and employee interviews collect personal information, and often opinions attributed to named individuals at named companies. PIPEDA’s fair information principles apply to that material, and the Office of the Privacy Commissioner’s generative AI principles recommend using de-identified data where personal information is not needed for the purpose, and knowing whether prompt content is used for further training. In practice: strip names before transcripts go into a tool, and honour whatever you promised the interviewee about attribution.
Can AI do the market sizing if we check it afterwards?
Checking a fabricated number takes as long as building a real one, and the fabricated number anchors the team. Build from sources; use the tool to find and read them.
What about vendor commercial due diligence?
Same method. The advantage is time: findings you generate yourself can be addressed before a buyer raises them.
How do we present interview evidence honestly?
Separate what participants said from what the data shows, and give the counts. Six of twenty is a finding; “customers report” is not.
A 30-minute call is enough to tell you whether AI pays for itself here.
