Hallucinated EBITDA: The Output Review Rule Every Bank Needs

What a hallucinated number looks like in practice

The failure is rarely a wild number. A model does not usually write that a $12 million revenue company has $40 million of EBITDA. It writes that adjusted EBITDA was $2.4 million when the data room says $2.1 million, because the model saw a $300,000 owner compensation add-back in one file and an already-adjusted figure in another and combined them. Or it writes "EBITDA margins expanded for the third consecutive year" when the margin was flat in year two, because that sentence is the kind of sentence CIMs contain.

Both errors can survive a casual read. The associate who skims the draft sees a plausible margin story and a plausible number. The managing director who reads it on the plane sees the same. The buyer's associate, who builds a model from the data room, may see the discrepancy within an hour and raise it on the first call. The bank's credibility on every other number in the document is then in question.

The worst version is the number that is close enough to be believed through the IOI stage and only caught in confirmatory diligence, after the LOI has been signed on the basis of the wrong figure. A re-trade may follow, and the owner may ask the bank how this happened.

Why fluent output needs stricter review than a junior draft

When an analyst writes a first draft, the errors are visible. The prose is uneven, the formatting is rough, and the reviewer naturally slows down. A generated draft reads like a finished document. The reviewer's eye speeds up, and errors that would have been caught in a clumsy draft can slide past in a polished one.

This is the central reason the review rule has to be explicit. Bankers have reviewed junior work for decades with an implicit standard that depends on the draft signaling its own roughness. Generated output does not send that signal. The standard has to be made explicit and mechanical, or it can quietly loosen.

There is a second reason. A junior analyst who invents a number knows she invented it and will usually flag uncertainty. A model does not know. It produces the number with the same confidence it produces the company's founding year. The reviewer cannot rely on hedging language to find the soft spots, because there may be none.

The rule: every number traces, and a named person traced it

The rule has three parts, and each is necessary.

Every number traces. Each figure in a client-facing document, including revenue, EBITDA, margins, growth rates, customer counts, headcount, and any multiple, is linked to a specific source: a page of the financial statements, a cell in the model, an answer in the questionnaire, or a file in the data room. If a number cannot be traced, it is removed or replaced, not left in because it seems reasonable.

A named person traced it. The reviewer signs the document record. Not "the team reviewed it." A name. The signature means that person opened the sources and checked. Naming the reviewer makes accountability visible in a way that an anonymous process does not. OECD AI principles likewise emphasize human oversight and traceability appropriate to the context. [3]

The review happens before the document leaves the firm, every time. A teaser going to a hundred buyers, a CIM going to twelve, a management presentation going to three. The rule does not scale down for small audiences, because the small audience is the one that may read most carefully.

The rule applies with equal force to narrative claims that imply numbers. "Margins expanded," "revenue doubled," "the top customer represents a modest share": each of these is a numerical claim in words and gets traced the same way.

Building the trace into the workflow instead of adding it afterward

A review rule that lives only in a checklist can be skipped under deadline. The durable version is built into how documents are produced.

The first move is to make the source visible in the draft itself. When a generated CIM shows adjusted EBITDA, the draft should carry a reference to where that figure came from, whether a footnote, a bracketed source tag, or a link back to the data room file. The reviewer's job becomes confirming references rather than hunting for them. Software that produces figures without any trace to the source makes the reviewer's job harder than if the analyst had built the number by hand.

The second move is to separate the valuation from the narrative. Numbers should come from a model with deterministic behavior, where the same inputs always produce the same output, and the narrative should be written around those numbers rather than generating its own. A CIM drafting tool that calculates EBITDA on the fly from whatever it finds in the data room can calculate it differently depending on which file it reads first. A valuation engine that computes once, records the inputs, and feeds the result to the narrative gives the reviewer one place to check.

The third move is a short reconciliation table at the end of the review: each headline number, its source, the reviewer's initials, and the date. The table takes fifteen minutes for a CIM and becomes the record if a question arises later.

Who reviews, and what the reviewer is actually checking

The reviewer should be the person closest to the data, which in most boutique banks is the associate who built the model or managed the data room, not the managing director. The managing director reviews for positioning and judgment. The associate reviews for traceability. Both reviews happen; they check different things.

What the associate checks, in order: every figure in the executive summary, because that is what buyers read first; every figure in the financial summary; every narrative claim that implies a trend or a magnitude; every customer, product, or geographic statistic; and every figure in the appendix tables, because models sometimes regenerate tables with transposed rows.

What the associate does not do is rewrite the prose. Editing for tone is a separate pass. Mixing the two slows the trace and lets errors through, because the reviewer gets absorbed in wording and skips the number.

Something to do this week: take the last CIM the bank sent out, generated or not, and have an analyst trace every number in the executive summary to its source. Time how long it takes and note what could not be traced. That is the baseline, and it gives the bank a concrete control to improve.

How Acquivo Pro fits

Acquivo Pro is built around the trace. The valuation engine is deterministic, so the same inputs always produce the same output, and figures trace to their source. Teasers and CIMs are generated from the client questionnaire and data room rather than from free-form prompts, so the reviewer is confirming a known input path rather than guessing where a number came from. The bank's named reviewer still signs off on every document before it leaves the firm; the platform makes that review faster, not optional. To see it on a live mandate, request a demo through the Contact page.