What’s the Best Way to Audit AI Decision-Making in Deals?

Updated August 2026 · 6 min read · Deal Room Intelligence Series

Audit the pipeline, not the model

The instinct is to demand transparency into how a conclusion was formed. That is the wrong target, and chasing it produces a false sense of assurance.

The right target is a boundary audit. Treat the model as an opaque component — because it is — and require that everything crossing its boundary is recorded:

Everything in the left column is fully auditable and, taken together, sufficient to reconstruct and defend a decision. Nothing in the right column is available from any vendor, and a vendor claiming otherwise is selling a narrative layer.

The four audits worth running

1. Coverage audit — run this first, every time

Before examining a single finding, reconcile: documents submitted, processed in full, partially processed, failed, and why. A document that failed extraction produces zero findings, and zero findings is indistinguishable from a clean document in almost every report format.

This takes two minutes for an entire data room and it is the highest-value control in the whole process. If coverage is broken, auditing findings is wasted effort — you are auditing a sample you did not choose.

2. Grounding audit — sample and trace

Take a sample of findings and open the cited document. Read the cited sentence against the claim. You are specifically hunting misgrounded citations: a real clause cited for a proposition it does not support. These pass an existence check and fail a reading check, which means they survive exactly the level of review most teams perform under deadline.

Sample deliberately across severity bands, not just the top. Errors concentrate where attention does not.

3. Calibration audit — are the scores meaningful?

Take findings your team overrode and look for patterns. If a category is routinely downgraded, the rubric is mis-tuned for your sector and every score in it is noise dressed as signal. If a category is routinely upgraded, you have a systematic under-detection problem.

This audit is the reason override capture matters. Without it you have no calibration data at all, and the system cannot improve through use.

4. Reproducibility audit — spot-check determinism

Periodically re-run a document processed earlier and compare. Divergence tells you either that the model changed without notice, or that the system is non-deterministic — both of which undermine your ability to reconstruct any prior decision.

What makes any of this possible

All four audits depend on the same underlying property: structured output with source text attached. Narrative reports cannot be audited this way. There is no category to sample by, no threshold to compare against, no score to track overrides on, and no quote to trace — verification means re-reading the document, which is the work the tool was supposed to reduce.

This is the practical reason a scoring stage is worth more than a summarisation stage. Not because a number is wiser than a paragraph, but because a number attached to a quote and measured against a threshold is the only form of output that can be audited at all.

Why it now matters beyond good practice

The governance surface hardened quickly and on a documented timeline:

• NIST AI Risk Management Framework 1.0 (26 January 2023) — voluntary, organised around Govern, Map, Measure, Manage. A reasonable skeleton for your internal audit programme.^[1]
• ISO/IEC 42001 (2023) — a certifiable AI management system standard.
• ABA Formal Opinion 512 (29 July 2024) — lawyers using generative AI must “fully consider their applicable ethical obligations,” including competence and supervisory responsibility.^[2]
• EU AI Act — Article 99 fines to €35,000,000 or 7% of worldwide annual turnover for prohibited practices, applying from 2 August 2025.^[3]

Whether the AI Act reaches your particular use is a classification question for counsel. The direction is not in question: the expectation that you can evidence how a conclusion was reached now sits in instruments with penalties attached, not just in good practice.

The audit finding you should expect

Run these audits honestly and you will find errors. Grounded commercial legal AI has been measured hallucinating between 17% and 33% of the time in an adjacent task, and providers' hallucination-free claims were judged overstated.^[4] A sampling audit that finds nothing is more likely to be a badly designed audit than a perfect system.

So define in advance what you will do with a finding rather than deciding afterwards: what error rate triggers a rubric change, what triggers a vendor conversation, what triggers suspending use in a category. An audit with no pre-committed response is a document-generating exercise.

A workable cadence

1. Every deal: coverage audit before findings are read. Non-negotiable, two minutes.
2. Every deal: grounding trace on everything escalated to committee.
3. Every deal: fixed random sample from the low-severity tail — say twenty findings.
4. Quarterly: calibration audit against accumulated overrides.
5. Quarterly, or after any vendor model change: reproducibility spot-check.
6. Annually: re-run vendor diligence — subprocessors, training terms, retention, versioning.

Bottom line

Stop trying to audit the model; it is not legible and the features that claim otherwise are generated narrative. Audit the boundary instead — inputs, outputs, thresholds, versions, and human decisions — because that is both fully recordable and sufficient to defend a conclusion.

Check coverage before findings, trace grounding on everything you escalate, track your overrides as calibration data, and decide in advance what you will do when the audit finds something. It will.

See an auditable, scored output →