Interrogate the model by treating AI like a junior’s first draft
Part 2 of AI Command Leadership for Finance — a short series on guardrails, judgment, and the operating model CFOs actually need.
When a junior hands you a first draft of a variance narrative or a reconciling pack, you do not file it. You ask questions. What did you pull? What did you assume? Where could this be wrong? Would you put your name on it for the CFO?
Treat AI output the same way.
That is not skepticism for its own sake. It is accountability. The model can assemble a draft in seconds. A person still owns the integrity of the books. Controllers and accounting leads who forget that distinction (who treat a confident draft as a finished workpaper) are the ones who create the risk CFOs are trying to avoid.
Accountability does not move to the model
Across this series we have named four pillars models are still bad at: judgment with context, control design, accountability, and relationships and influence. Part 1 put judgment and control design into a keep / automate / supervise map. This piece sits on accountability and judgment together.
AI does not sign the financials. It does not sit with the auditors. It does not take the call when something is off. A person still does. That means every AI-assisted workpaper still needs a human owner who can answer three questions under pressure: what did we use, what did we decide, and why should leadership trust it.
If you cannot answer those, the draft is not ready - no matter how polished it looks.
Treat the draft like a junior’s first pass
A useful mental model: the model is a fast junior who has never worked a close in your company. It can be thorough. It can also be confidently wrong about timing, policy, related parties, and what “material” means in your business.
Before you accept an AI draft into the pack, run the same interrogation you would with a new hire:
What data did it use? Source systems, periods, filters, exclusions. If you cannot name the inputs, you cannot defend the output.
What did it assume? Accounting policy, cut-off, materiality thresholds, classification rules, prior-period treatment. Assumptions are where drafts go quiet and wrong.
What would make it wrong? A late invoice, a reclass, a one-time contract, a related-party entry, a coding error that looks like margin. Name the failure modes out loud.
Would you sign it? Not “does it sound right.” Would you put your name on it for the CFO and for audit. If the answer is not yes, it stays in supervise - or goes back to keep human.
Those four questions are judgment with context in practice. They also create the audit trail accountability requires.
Questions finance can use this week
You do not need a new platform to start. Pick one process you already mapped into supervise in Part 1 — a variance narrative, an AP exception summary, a first-pass reconciling pack, and require the reviewer to answer the following before the draft moves forward:
Which source systems and periods fed this draft?
What was excluded, and why?
Which assumptions mirror our policy, and which are the model’s guesses?
Where is materiality applied, and by whose threshold?
What would change the conclusion if it appeared tomorrow?
Who is the named human owner of this result?
What evidence will we keep so an auditor can follow the path from input to sign-off?
Write the answers in the workpaper or the review notes. That is not bureaucracy. That is how you earn the right to give the model more autonomy later — the same reliability-first standard from Part 1.
If your landscape is Microsoft, SAP, OpenText, or a mix, the interrogation pattern is the same. The stack is where you log the review, store the evidence, and bind the human owner. The operating model is the decision about what must be asked before anything posts.
What not to do
Accept a fluent draft because it is faster than challenging it. Speed without ownership is how errors reach the pack.
Treat “the model said so” as a control. It is not.
Skip naming a human owner on AI-assisted workpapers. Ownership cannot be anonymous.
Confuse a validation checklist with a one-time training session. Patterns stick when they are in the close calendar.
Expand autonomy before the interrogation habit is boring and reliable.
Validation patterns worth implementing
When teams ask McCloy Data for help here, we are usually not starting with a new model. We are implementing review patterns: what must be asked, who must sign, what evidence must exist before an AI-assisted process graduates from supervise toward more automation.
A concrete deliverable looks like this:
a validation checklist tied to your keep / automate / supervise map
named owners and mandatory review points for AI-drafted workpapers
evidence expectations an auditor can follow
a short pilot design on one or two supervise processes, with residual risk reported back to the CFO
That is accountability made operational — not a demo, not a prompt library.
What comes next
This is Part 2. Part 1 was the map: keep, automate, supervise. Next, in Part 3, we put the four pillars into CFO language — what leadership actually needs from AI in finance, and how close days, errors, PBC, and cost per close become the frame for a roadmap you can run.
If you already have a keep / automate / supervise map and the drafts are landing faster than the review habit, start with interrogation. Ownership of the books still sits with people. The model is only as useful as the questions you refuse to skip.
If you want help building the validation pattern
McCloy Data’s AI & Engineering practice works with finance and accounting teams on validation checklists, review patterns for AI-assisted processes, and pilot designs that keep human accountability clear on the systems you already run. If that pain is live in your shop, we would welcome a brief conversation.