AI in a four-week FDD: what it changes and what it doesn't

AI now does the data-room triage, the GL testing and the first draft. It does not decide what EBITDA is worth. A note on how we use it, how we control it, and how we diligence 'AI ARR'.

A four-week buy-side FDD on a mid-cap target used to follow a fixed pattern. The first week went on reading the data room and chasing the general ledger. The second and third weeks went on analysis. The last few days went on writing. AI has compressed the first week and the writing. It has not compressed the analysis, and it should not.

Law firms and the Big 4 now write about AI-assisted due diligence as normal practice, not as an experiment. Sullivan & Cromwell's January memo on AI in M&A notes that machine learning has been used for years to extract provisions from data rooms, and that generative tools now produce summaries and first drafts. Virtual data room providers have added AI search, document summaries, automatic indexing and Q&A tracking. Our clients' counterparties use the same tools. The question is no longer whether to use AI, but which parts of the work it can take on and how the output is checked.

What we automate, and how we check it

We use AI where the task is high-volume and the answer can be verified. We do not use it where the task is a judgment.

Full-population testing changes what a red flag looks like. On a sample basis, a €300k manual credit note posted on 30 December could be missed. With every journal entry scored, it shows up in the first pass. On a recent €45m EV distribution target, the GL scan returned around 140 exceptions in two days. Eleven needed a question to management. Two became adjustments: a year-end rebate accrual released without a matching credit note (€0.4m), and recurring consultancy fees booked to capex (€0.25m a year). The tool found the entries. Deciding that the first was a cut-off error and the second a recurring cost took a conversation with the CFO and a look at three years of invoices.

That is the limit. A tool can tell you that an entry is unusual. It cannot tell you whether the EBITDA it affects is sustainable, whether a cost is truly non-recurring, or whether management's run-rate is credible. Quality of earnings is a view on the future built from evidence about the past. The view is ours, and the client pays for it.

Confidentiality: three checks before anything is uploaded

Most of the risk in AI-assisted diligence is not technical. It is contractual.

1. The NDA and the data-room terms. Sellers now often prohibit uploading data-room content into AI tools. Commentary from US and UK law firms since late 2025 describes a sharp rise in such clauses. Some are blanket bans, which are hard to apply when office software itself embeds AI. Others allow "controlled" use: enterprise tools only, no training on the data, no retention by the provider beyond the engagement. We read these terms at the start of every engagement and record what is permitted. Where the terms are silent, we ask the seller's adviser before uploading anything.

2. The tool itself. We use enterprise tools with contractual terms that exclude training on client data and allow deletion at the end of the engagement. A provider's licence over uploaded content, or a retention period longer than the NDA allows, can amount to a prohibited disclosure. It also makes the return-or-destroy clause impossible to comply with.

3. Client consent. Our engagement letter states which tools we use and for which tasks. Some clients, mostly corporates, ask us to switch them off. We accept that, and we price the extra time.

One further point for buyers. A Mayer Brown note from March reports US decisions on whether AI prompts and outputs are protected from disclosure. In United States v. Heppner (SDNY, February 2026), materials an executive prepared with a generative AI tool were held not to be privileged. AI-generated diligence summaries are documents. In a post-closing warranty claim, they may be used to argue that the buyer knew about an issue. The same discipline applies as for any workpaper: write what you mean, and do not leave unreviewed summaries in the file.

In software targets: separating "AI ARR" from recurring revenue

Most software deals we see now include an AI story, and SEG's 2026 SaaS report confirms that buyers pay a premium for clear AI positioning. That makes the revenue labelled "AI" worth testing line by line. Regulators are watching AI claims too: in March 2024 the SEC settled charges against two investment advisers, Delphia and Global Predictions, for describing AI capabilities they did not have. The penalties were small (USD 400k in total). The point for diligence is that "AI" in a pitch deck is a claim, and a claim needs evidence.

The difficulty is that AI features are often priced on usage: tokens, credits, queries. There is no accounting standard for annualising usage revenue into ARR. Some companies take their best month and multiply by twelve. Lenders to software companies have started to use rolling three- or six-month averages, and minimum consumption floors, for that reason.

A worked example. A French vertical software target, €60m EV, reports ARR of €12.4m, of which €2.1m is presented as "AI ARR". Our breakdown:

1. €0.9m: AI module sold as a subscription add-on, on annual contracts with the same renewal terms as the core product. This is recurring. We keep it in ARR.
2. €0.7m: usage-based credits, annualised from March, the highest month of the year. The trailing three-month average gives €0.45m annualised, and usage is concentrated in four customers. We report it separately as usage revenue, at €0.45m, with the concentration shown.
3. €0.5m: paid pilots, three-month contracts with no commitment to convert. This is not ARR. We show it as pipeline.

Restated ARR is €11.2m (€10.3m core plus €0.9m AI subscriptions), with €0.45m of usage revenue shown alongside it rather than inside it. Headline ARR falls by €1.2m, and "AI ARR" falls from €2.1m to €0.9m. At the multiple the buyer had in mind, the difference matters more than any EBITDA adjustment in the report.

We also look at gross margin. Inference costs (model provider fees, compute) are often booked in R&D or in general IT costs rather than in cost of sales. Reclassifying them can move gross margin on the AI product by ten points or more. We ask for the provider invoices and the usage logs, and we reconcile the two.

What we suggest

• Buyers: ask your FDD provider which tasks are automated and how the output is reviewed. Full-population GL testing should be the default, not an option.
• Before uploading: check the NDA and the data-room terms for AI clauses, and record what is permitted. Assume your diligence summaries could be read in a later dispute.
• On software targets: request ARR by revenue type (subscription, usage, pilots), monthly usage data for 12 months, and the inference cost invoices. Value usage revenue on a trailing average, not the best month.
• In the SPA: if AI revenue supports the price, consider a specific warranty on the ARR definition and the revenue split, so the definition is fixed by reference to the DD workpapers.

This note is general information, not advice on a specific transaction.