Training Associates and Analysts to Edit AI Output, Not Rebuild It

The instinct you are training against

Junior bankers are taught that quality comes from building. The analyst who rebuilds the model from source learns the business. The associate who writes the CIM from a blank page owns every sentence. That training is correct for a world where the alternative to building is copying someone else's template without understanding it.

A generated draft is not someone else's template. It was built from this client's questionnaire and this client's data room, so it already reflects the business. But the associate's instinct does not know that. The associate opens the draft, sees a paragraph that is phrased differently than they would phrase it, and rebuilds. Two days later the CIM is done, it is roughly as good as the draft was, and the associate reports that the tool did not save time.

The instinct has to be named before it can be replaced. In the first training session, say it out loud: your job on this mandate is to edit, not to write, and if you find yourself with a blank document open, stop.

Three categories: wrong, generic, and different

The core skill is sorting every section of a generated draft into one of three categories, and the training is mostly practice at sorting.

Wrong means the draft states something the data does not support. A revenue figure that does not match the normalized financials. A customer concentration claim the data room contradicts. A product line the company exited. Wrong gets fixed, and the fix gets traced back to source. This is the category junior bankers are already good at, because it is the same skill as checking a model.

Generic means the draft is accurate but could describe any company in the vertical. A market overview that lists industry tailwinds without connecting them to this business. A growth section that says expand geographically and add services. Generic gets rewritten with specifics from the data room and the owner interview, and the rewrite is usually two or three sentences, not two or three pages.

Different means the draft says what the associate would have said, in words the associate would not have chosen. This is the category that eats time. The trained response to different is to leave it alone unless it fails the managing director's standard for client-facing prose. Most of the time it does not, and the associate's preference is not a reason to spend an hour.

An associate who can sort a 30-page CIM draft into these three categories in an hour, and then fix only the first two, has learned the job.

Running the first training session on a real draft

Do not train on slides. Train on the first generated CIM draft from the first live mandate, with the associate who will edit it and the analyst who will support them in the same room, for ninety minutes.

Print it or share the screen. Go section by section. For each one, the associate says which category it falls in and why, and the managing director or senior associate running the session agrees or disagrees. Where they disagree, the discussion is the training. The associate will tend to call sections wrong that are actually different, and the senior banker will usually be able to show that the draft's phrasing is acceptable.

A marked-up draft is the first training artifact. The group should also keep the first version of the list that belongs in the SOP: which sections of a CIM in this vertical came out usable, which came out generic, and which the partner will always rewrite. Then the associate edits the draft for real and times it. The time is the second training artifact.

Training the analyst differently from the associate

The analyst's job in an AI workflow shifts more than the associate's does, and the training should reflect that. The analyst used to build the model, the data book, and the buyer spreadsheet. Now the workflow can produce those from the intake and the database, and the analyst's job becomes verification. That division is consistent with a broader risk-management principle: NIST recommends testing, evaluation, verification, and validation throughout an AI lifecycle, with results documented and measured rather than assumed. [1] The OECD likewise lists transparency, explainability, and accountability among its principles for trustworthy AI. [2]

Train the analyst on two things. First, tracing. When the valuation engine produces a figure, the analyst should be able to click through to the source line in the normalized financials and from there to the owner's upload. Have them do it for ten line items on the first mandate, out loud, so they learn what the trace looks like and what a break in the trace looks like.

Second, the intake check. The readiness view normalizes what the owner entered, and the analyst compares the normalized version to the raw uploads for anything the owner entered inconsistently: a fiscal year that does not match the statements, an add-back with no support. This is the same skill as spreading financials by hand, applied to checking rather than entering.

The analyst should not be trained to edit the CIM. That is the associate's job, and splitting it confuses the review chain. The managing director still owns the final judgment.

The second and third mandates: where the training sticks

One session does not change an instinct. The training sticks on the second mandate, when the associate opens a generated draft in a different vertical and has to sort it without the senior banker in the room. Expect the associate to over-edit again, slightly less than the first time.

The fix is a short check-in after the first draft is sorted: the associate shows the senior banker which sections they marked as wrong or generic, and the senior banker pushes back on any that are actually just different. Fifteen minutes. By the third mandate the check-in may be unnecessary.

Measure it. Hours to edit the CIM on mandate one, two, and three. The number is a local operating measure, not a promised outcome. If it does not fall, either the drafts are getting worse, which the sorting exercise will show, or the associate is still rebuilding, which the hours will show. Repeated measurement is preferable to assuming that a training session worked. [1]

What to do this week

Find the next generated draft that is going to an associate, in any phase. Before they touch it, ask them to spend thirty minutes sorting it into wrong, generic, and different, and to write the three lists down. Then ask them to fix only the first two and time it.

That is the whole training program, run once. Everything else is repetition.

How Acquivo Pro fits

Acquivo Pro produces the drafts that this training is built around: the teaser and CIM generated from the client questionnaire and data room, the deterministic valuation whose every figure traces to its source, and the normalized intake from the readiness phase. Because figures trace to source, the analyst's verification work is a click-through rather than a rebuild, and because the same inputs always produce the same output, a rerun is a check rather than a surprise. The associate edits in place and the managing director reviews on the same login. To see it on a live mandate, request a demo through the Contact page.