# Approval Is Not Judgment: Decision Records for AI-Assisted Work > Audit trails prove a process ran. They rarely show why anyone agreed with an AI-drafted recommendation. Canonical HTML: https://josejruiz.com/en/insights/ai-decision-records-judgment-trail/ Insights Audit trails prove a process ran. They rarely show why anyone agreed with an AI-drafted recommendation. What a judgment-bearing decision record contains, and where to require one. Published: November 12, 2026 ![A figure in a navy suit stands beside a long, orderly chain of identical stamped document cards that has one blank gap in the middle, holding up a single handwritten card toward the gap, with a warm ochre halo behind.](https://josejruiz.com/insights/ai-decision-records-judgment-trail/ai-decision-records-judgment-trail.webp) An audit trail can be complete and still explain nothing. In AI-assisted work, decision records increasingly capture every signature, timestamp and approval while preserving none of the reasoning that made a decision defensible, because the rationale itself was drafted by a system and accepted without anyone writing down why. That gap tends to stay invisible until someone outside the company asks a simple question about a decision made long ago. What follows looks at why conventional records fail that test, the difference between an audit trail and a judgment trail, what a judgment-bearing decision record actually contains, and how compliance, quality and internal-audit leaders can require one without burying the organization in paperwork. ## Why don’t audit trails explain AI-assisted decisions? An audit trail is proof that process happened. It shows who opened which document, when, for how long, under which version of which procedure. In regulated industries these trails are often excellent, and organizations are right to be proud of them. But an audit trail answers the auditor’s question: did you follow your procedure? It does not answer the harder one that regulators, courts, boards and affected customers eventually ask: when your tool told you not to worry, why did you believe it? Before AI, the gap was partly filled by the work itself. Someone had to draft the evaluation, and the draft carried traces of their reasoning: what they emphasized, what they flagged, what they left open. When a system drafts the evaluation, the document still looks reasoned. It has a conclusion, supporting data and a recommendation. What it no longer contains is any evidence that a person weighed it. The record holds the draft and the acceptance with equal fidelity, and nothing in between, because nothing in between was ever asked for. ## What does a decision with no recorded judgment look like? Consider a composite case from a medical-device manufacturer. A monthly post-market review considers a rising trend in complaints about one product line. An AI module reads the complaint narratives, codes them into categories, weighs them against shipment volumes and drafts a two-page evaluation: the rate is within the historical band, returned samples passed testing, and no single complaint category is trending. Recommendation: continue routine monitoring. The item takes nine minutes. Five reviewers sign. A year later, after a serious incident and a recall, a regulator asks the company to explain the basis for that decision. Not to produce it, which takes an hour. To explain it. Every number in the evaluation turns out to be correct. Every signature is valid. And no one who signed can reconstruct why they agreed. The chair’s honest answer is that nothing in the packet gave him a reason not to. Somewhere in the workflow, a complaints analyst several levels down had noticed that the narratives, read in sequence, sounded like one problem described in different words. She typed a note into the triage field. The next version of the evaluation absorbed the same data and reported, accurately, that all categories remained within band. No one rejected her concern. It simply left no trace in the record of the decision. The minutes read: “Discussion held. Recommendation accepted.” ## What is the difference between an audit trail and a judgment trail? A judgment trail is proof that someone owned the choice. Not that they agreed, which the signature already establishes, but why. It is the record that lets a reader, months or years later, see that a mind was present when the decision was made. The distinction matters most in AI-assisted work, because the system’s output can be accurate in every particular and still be insufficient to decide from. A review board can act diligently, at the level of attention the afternoon allows, with a tool whose numbers are right, and leave behind a record in which ownership exists only on paper. Ownership that leaves no trace is, afterward, indistinguishable from ownership that was never taken. This is where human authorship meets the record. On the Second Map of Work, Reliance is the point where someone begins acting as if the output is trustworthy, and Consequence is where the work lands on a person. A judgment trail records what happened between those two points: what the human brought to the output, what they doubted, and why they relied on it anyway. ## What should a judgment-bearing decision record contain? It can be smaller than most organizations expect. A workable design uses four entries, written by the person accountable for the decision, in their own words, before any signature is applied: **What we considered.** The tool’s conclusion, and whatever was brought to the table beyond it: other data, field observations, a colleague’s concern. **What we are still uncertain about.** The open questions the decision is being made around, including the ones that would change the answer if resolved differently. **Who challenged the conclusion, and what happened to the challenge.** Every challenge has a name attached and can be closed only by a named person giving a reason. Answering doubt becomes a responsibility, not something the system performs on the way past. **Why I agree.** The sentence the decision-maker would be willing to read aloud a year later to a regulator, a court or the person affected. The third entry carries the sharpest edge. It is what would have kept the analyst’s note in the room, with her name on it, until someone answered it. If the answer is “we looked, and the testing does not support it,” that is acceptable. It is a judgment, owned by someone. What is no longer acceptable is a doubt that dissolves without a trace. ## How can leaders require judgment trails without adding bureaucracy? The first instinct of most organizations is to add a template. Resist it. A template invites the pasted phrase, and “consistent with prior data” in a why-I-agree field is just a signature in sentence form. **Apply it where consequences travel.** A judgment trail on everything is a judgment trail on nothing. Reserve the four entries for decisions whose consequences can reach a patient, customer, employee or regulator before anyone gets a second look. Routine records can stay as they are. **Write before signing, not after.** The entries are not documentation of a decision already made. They are where the decision gets made. A chair who has to write down why they agree will sometimes discover, while the meeting is still in the room, that they do not yet know. **Expect the first weeks to be awkward.** People fill new fields the way they fill any field: with filler, or with length mistaken for rigor. Review the entries, give feedback, and narrow the scope before abandoning the practice. **Give the people closest to the data a way into the record.** Analysts, reviewers and front-line staff often see the pattern first. The record should carry their observations under their names, not just their button clicks. **Test it the way an outsider would.** Internal audit can pick a consequential decision from a year ago and ask: could we explain not only what the tool concluded, but why a human agreed? If the answer is no, the record was designed for approval, not judgment. As one character in a recent book puts it, “The rest of the file proves the process ran. These four lines prove a person was in it.” ## Frequently asked questions **What is a judgment trail in AI governance?** A judgment trail is a record showing why a person agreed with an AI-assisted recommendation, not just that they approved it. It captures what was considered beyond the tool’s output, what remained uncertain, who challenged the conclusion and how the challenge was answered, and the decision-maker’s own reasons, written before signing. **How is a judgment trail different from an audit trail?** An audit trail proves the process ran: who touched which document, when, under which procedure. A judgment trail proves someone owned the decision and can explain why. Audit trails answer whether procedure was followed. Judgment trails answer whether the people responsible actually exercised judgment when an AI system’s output shaped the outcome. **Which decisions need a documented rationale when AI is involved?** Decisions whose consequences can reach a patient, customer, employee or regulator before anyone gets a second look. Product safety reviews, credit and claims decisions, hiring and termination calls, and material risk classifications are typical examples. Applying the requirement to every routine decision dilutes it and turns the entries into filler. **How do you document human oversight of AI decisions for regulators?** Record, in the decision-maker’s own words and before sign-off, what was considered beyond the AI output, what remained uncertain, which challenges were raised and how each was resolved by a named person, and why they agreed. Regulators increasingly ask for the basis of a decision, not just evidence that it was approved. * * * Compliance, quality and internal-audit leaders redesigning how AI-assisted decisions are recorded can [start a conversation](https://josejruiz.com/en/contact/). Anker Bioss advises organizations on governing judgment in AI-shaped work. Related reading: [Author, Not Approver: Redesigning the Human-in-the-Loop](https://josejruiz.com/en/insights/human-in-the-loop-author-not-approver/). _Jose J. Ruiz is CEO and Managing Partner of Alder Koten and Chairman of Anker Bioss. He develops these ideas at book length in [AI in the Org Chart: A Leadership Guide to Implementing AI Without Losing Human Judgment, Accountability, and Trust](https://josejruiz.com/en/books/ai-in-the-org-chart/), published by [Elavant Press](https://press.elavant.com) and available on [Amazon](https://geni.us/aiintheorgchart)._ Topics - AI Governance - Decision Records - Compliance - Internal Audit - Judgment --- Site guide for agents: https://josejruiz.com/llms.txt · Sitemap: https://josejruiz.com/sitemap-index.xml