AI Investment Committee Reporting: What to Trust, What to Verify

AI investment committee reporting works for one part of the pack and should stay out of the rest. The low-risk, genuinely useful part is the cost-and-scope section: pulling committed versus actual advisor spend by workstream, flagging what changed since the committee last sat, and drafting the variance narrative from a data trail someone can trace back to source. The recommendation, the risk framing, and the deal thesis are judgement calls, and Deloitte's 2026 Generative AI in M&A Pulse Study found human review remains the leading requirement organisations attach to high-stakes AI use in M&A, for good reason.

That distinction matters more than it looks, and it maps onto a pattern familiar from diligence budgets generally: they don't blow up because advisors charge too much, they blow up because scope drifts silently and nobody prices the drift until the invoice lands. The IC cost-and-scope section is exactly where that drift either gets caught while there is still time to act on it, or stays hidden until the deal has closed. Most of the anxiety around "AI in the IC pack" is really about the other category, an AI system quietly shading a go/no-go call, and that anxiety is pointed the wrong way. Almost nobody is worried that software adds up a spend table correctly, which means the section actually worth watching gets the least scrutiny of the three.

What Can AI Credibly Do in an IC Cost Update?#

AI can credibly handle the mechanical half of an IC cost update: reconciling committed and actual spend by workstream, surfacing what moved since the previous meeting, and turning a structured variance into readable prose. None of that requires judgement about the deal itself, only accurate arithmetic over data that is already structured, which is precisely the kind of task a language model does well and a spreadsheet does slowly.

Concretely, that means three things. First, assembling the numbers: total approved, committed, and forecast spend per advisor workstream, the same structure our guide to what goes in an IC cost report sets out, but generated from live records rather than reconstructed from invoices the night before. Second, flagging the delta: which workstream moved since the last sitting, by how much, and why, drawing on the audit trail a well-run change-request process should already be producing. Third, drafting the narrative sentence for each material variance, "financial diligence is forecast £38k over approved after the target's second trading entity was brought into scope," because that sentence is a mechanical restatement of a change-request record, not an opinion.

This is the part of the wedge Advilink is building toward for pilot deal teams: live committed-versus-actual tracking by workstream that a deal lead can turn into an IC-ready cost update on demand, rather than reconstructing it from five invoices and an email thread. It is in development and being validated with design partners, not a shipped feature you can point at today.

Why Should the Recommendation Stay Human?#

The recommendation should stay human because it requires weighing evidence the way a diligence pack cannot fully represent: which risks are dealbreakers, how much they trade against price, and whether the target's story survives contact with the numbers. Weighing evidence like that is judgement, and language models are trained to produce fluent, confident text regardless of whether the underlying judgement is sound.

Grant Thornton's 2026 AI Impact Survey, a survey of 950 business leaders across finance, operations and technology roles conducted in February and March 2026, found only 5% of organisations permit AI agents to execute high-stakes decisions without human review; 60% cap agents at moderate-risk task automation instead. As Tom Puthiyamadam, a managing partner at Grant Thornton Advisors, put it in the firm's release of the findings, "AI deployment has outpaced the infrastructure to defend it" at most organisations. An IC recommendation is about as high-stakes as a corporate decision gets, and the infrastructure gap he is describing, audit trails, model validation, someone accountable for the output, is exactly what a deal-thesis recommendation would need before a committee should trust it unreviewed.

Is an AI-Assisted Cost Update More Defensible Than a Generative Summary?#

Yes, and the reason is traceability, not intelligence. A cost-and-scope section built from structured records, approved budgets, logged change requests, invoiced and committed spend, can be traced back to the individual record that produced each number. A generative summary of email threads and PDF invoices produces fluent prose with no equivalent trail; if a committee member asks where a figure came from, the honest answer is "the model inferred it from the documents it was given," which is a materially weaker answer than "line 14 of the legal workstream ledger, change request CR-07, approved 12 August."

The gap is not theoretical. A preregistered evaluation of leading AI legal research tools, published by Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning and Daniel E. Ho in the Journal of Empirical Legal Studies in 2025, found correct-and-grounded answers ranged from 65% for the best-performing tool down to 19% for the weakest, across 202 professionally posed queries, and these are retrieval-augmented tools purpose-built for grounded legal answers, not a general model summarising a loose email thread. If a specialised, retrieval-grounded product still gets a material share of citations wrong on a narrow task, a free-form summary of an inbox is a worse bet for anything a committee will rely on.

Daniel E. Ho, the Stanford law professor who co-authored that study and directs Stanford's Regulation, Evaluation and Governance Lab, has pointed to the deeper cost this creates in practice, not just the error rate itself. Describing why long, citation-heavy AI answers can undercut their own time-saving promise, he said on the Stanford Legal podcast that "it was extraordinarily time consuming to actually fact check every single one of these things, 'cause you were getting long answers with 13 citations, you had to go to every one of the underlying citations to figure out where it was grounded." That is the fact-checking tax a generative summary imposes on whoever has to trust it, and it is precisely the tax a source-linked cost record is designed to avoid: the committee member checks one line, not thirteen citations.

Committees are already living with the consequence of skipping that check. Workiva's 2026 Midyear Executive Benchmark Survey, covering 2,272 finance, risk and legal professionals including 847 C-level executives, fielded across 16 countries in May 2026, found 84% of executives at least somewhat confident in AI output accuracy without human review, while 26% said an internal audit had already caught an AI-driven error that reached external audiences or the board. That 84%-versus-26% gap is the verification gap: confidence running well ahead of the checks that would justify it. An auditable cost trail closes that gap by construction, because the number and its source are the same artefact; a generative summary widens it, because the fluency of the output has nothing to do with whether it is right.

What Should a Committee Demand From AI Investment Committee Reporting?#

A committee should demand four things before it treats an AI-assisted figure as reliable: source traceability, a visible timestamp, a named human sign-off, and the ability to drill from the summary number down to the underlying record. None of these require banning AI from the pack. They require the pack to be built so that AI's contribution is checkable rather than taken on faith.

Auditors already apply this discipline to any control that produces a number for a board: know the source, know the date, know who signed off. IC reporting simply needs that same standard extended to cover a model's output as well as a spreadsheet's, and most organisations have not caught up yet. Grant Thornton's survey found 78% of the executives it polled lack strong confidence they could pass an independent AI governance audit within 90 days, which is a fair proxy for how few IC packs today could show their working if a committee member actually asked.

One honest limitation: none of this eliminates the underlying data-quality problem. If the workstream budgets, change requests, or invoices feeding the cost section are themselves wrong or stale, an auditable trail faithfully reproduces the error with a timestamp attached. A trail lets a bad number get caught; it doesn't guarantee the number was right in the first place. That is still worth having, because a checkable error gets caught, and an untraceable one usually doesn't until the deal has closed and the final invoice disagrees with everyone's memory of what was agreed.

The committees that get this right will not be the ones with the most AI in the pack. They will be the ones who can say, for every number in the cost section, exactly where it came from, and who kept the recommendation on the page that stays a person's to make.

Frequently Asked Questions#

Can AI write the investment committee memo for due diligence?#

AI can credibly assemble and draft the cost-and-scope section, committed versus actual spend by workstream, what changed since the last sitting, and the variance narrative behind it, because that content is a mechanical restatement of structured data. The recommendation, risk framing, and deal thesis should be authored and signed off by the deal lead; Deloitte's 2026 Generative AI in M&A Pulse Study found human review remains the leading condition organisations attach to high-stakes AI use in M&A.

Is AI-generated investment committee reporting reliable?#

It is reliable for the parts drawn from an auditable data trail, structured budgets, logged change requests, invoiced spend, because every figure traces back to a specific record. It is materially less reliable when it comes from a generative summary of unstructured sources like email threads, where a 2025 study in the Journal of Empirical Legal Studies found even purpose-built, retrieval-grounded professional AI tools produced correct-and-grounded answers only 19% to 65% of the time across 202 queries.

What should an investment committee ask before trusting an AI-assisted cost update?#

Four things: where each number traces back to, whether the figures are timestamped to the meeting date rather than carried over from last time, whether a named person signed off on the variance narrative, and whether the cost section is kept structurally separate from the recommendation. Workiva's 2026 Midyear Executive Benchmark Survey found 84% of executives confident in unreviewed AI output even though 26% had already seen an AI-driven error reach the board.

Does using AI for IC cost reporting replace the deal lead's role?#

No. It is designed to remove the manual reconciliation, chasing invoices, checking engagement letters, rebuilding a spreadsheet, that currently eats a half-day before every committee meeting, not to remove the deal lead's judgement about what the numbers mean. The deal lead still owns the narrative sign-off, the risk framing, and the recommendation.

How is an AI-assisted IC cost update different from a static spreadsheet report?#

A spreadsheet report reflects whatever was last typed into it, often days or weeks stale by the time it reaches the committee. An AI-assisted update built on a live, workstream-level cost trail can reflect committed and forecast spend as of the morning of the meeting, and because the underlying data is structured, every figure it produces can be traced back to a source record rather than reconstructed from memory.

Sources#

1. 2026 Generative AI in M&A Pulse Study: Deloitte, 2026. Human review remains the leading condition organisations attach to high-stakes generative AI use in M&A.
2. 2026 AI Impact Survey: Grant Thornton, 2026. Survey of 950 business leaders (February–March 2026); 78% lack strong confidence they could pass an independent AI governance audit within 90 days; only 5% permit agents to execute high-stakes decisions without human review.
3. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools: Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning and Daniel E. Ho, Journal of Empirical Legal Studies, 2025. Preregistered evaluation of proprietary retrieval-augmented legal research tools across 202 queries; correct-and-grounded answers ranged from 19% to 65%.
4. AI, Liability, and Hallucinations in a Changing Tech and Law Environment: Stanford Legal podcast, Stanford Law School, 2024. Daniel E. Ho on the fact-checking burden long, citation-heavy AI answers create.
5. 2026 MidYear Executive Benchmark Survey: The Verification Gap: Workiva, 2026. Survey of 2,272 finance, risk, sustainability and legal professionals, including 847 C-level executives, fielded across 16 countries in May 2026; 84% confident in AI output accuracy without human review, 26% report an AI-driven error reached external audiences or the board.