Why AI Still Can't Make a High-Quality Company Valuation

Every few months, a new AI model arrives with a longer context window, sharper reasoning, and the ability to digest an entire data room in seconds. It's tempting to conclude that company valuation — one of the most judgment- and experience-heavy corners of finance — is finally ready to be automated. Feed the model the financials, ask for a number, and get a defensible answer back.

Reality, however, is a bit more complicated.

AI Bias in Company Valuation

A valuation is rarely requested in a vacuum. A founder preparing to sell their company, a management team pitching investors, or an owner weighing a deal usually already has a number in mind, and large language models are trained to be cooperative rather than skeptical. Given a persuasive story (“our churn is temporary,” “a competitor exited the market, so we'll capture their revenue”), a model will generally take it at face value, because it has no lived experience of the industry and no instinct for spotting self-serving arguments.

Worse, this doesn't stop at the first answer. If the result doesn't match what the requester wanted, they can keep arguing with the model, reframing assumptions and applying pressure, and the model tends to treat each objection as new information to accommodate rather than defend the correct answer against. A human advisor preparing an M&A deal can hold their position under pressure, if only because their name is attached to that number. AI has no equivalent — bias here isn't a one-time flaw, it's a negotiation the requester can keep having until they win.

Data Quality: Why Accurate Numbers Don't Guarantee the Right Answer

The second problem is just as damaging and far less visible: a valuation is only as good as the data behind it, and AI cannot verify that data is complete or honest. It can't “smell” that a customer-concentration risk is missing from the deck, or that a related-party transaction is buried in the accounts. And unlike a human expert, who will add a note such as “I could not verify this point,” a model always answers with the same confidence, regardless of how incomplete the input is.

Past vs. Future: When Financial Statements Don't Show the Full Picture

Even when the numbers are accurate, they only show a company's past, not its current state. Due diligence often reveals that a business is in the middle of major change — new equipment, a relocation, a restructuring — that temporarily distorts financial performance without reflecting the company's real long-term value. A profit decline caused by investment in new production is an entirely different story from a profit decline caused by a shrinking business. The reverse is just as common: an excellent year that turns out to be driven mostly by a one-off asset sale disappears, value and all, once that number is stripped out.

AI is genuinely good at spotting unusual deviations — for example:

• receivables from customers growing faster than revenue;

• unexplained shifts in margins;

• interest expense that looks implausibly low given the level of debt.

What it can't do is take the next step: asking management for an explanation, gathering the missing context, and assessing what the answer actually means for future cash flows.

Valuing SMEs and Searching for Comparable Companies

The gap between “capable in general” and “reliable for this specific case” widens further once you move away from large, well-documented companies. Most of what a model “knows” about how businesses behave comes from data on large public companies. A small or mid-sized enterprise (SME) in a niche market isn't represented the same way, and the model has no way of flagging that mismatch. It will readily assume a growth rate is sustainable, or that a cost base scales the same way it would for a company ten times its size — simply because that pattern dominates its training data.

For the same reason, open-ended AI tasks tend to disappoint in practice. Asked to independently find comparable companies for an M&A transaction, a model will suggest plausible-looking names that don't hold up on closer inspection. Give it a pre-researched shortlist instead and ask it to rank the best candidates, and the results improve markedly. The lesson generalizes: AI is far stronger at making choices within boundaries someone else has set than at open-ended discovery, where it has no reliable way to separate genuine signal from something that merely looks relevant.

Why Company Valuation Is More Than Just a Number

Even if everything above were solved — perfect data, no bias, a clearly bounded task — a deeper problem remains. It's arguably the hardest one to fix, because it has nothing to do with AI's capability at all: most people don't know what they should be asking.

A founder preparing to sell their company who asks “how much is my company worth” gets a far more superficial result than one who asks “what would a strategic buyer pay given our customer concentration, and where is there room to negotiate.” Only someone who already has much of the expertise the tool is supposed to replace knows to ask the second question.

Valuation as Diagnosis, Not Just a Figure

This matters because, in an M&A context, a company valuation is often not just a number but a diagnosis: where the company is strong, where it's exposed, how a prospective buyer is likely to read the numbers, and what similar deals in the market have looked like. That diagnostic layer — together with the ability to judge the counterparty's negotiating position and advise on when to hold firm — is exactly what a client needs the moment a valuation carries real weight in a negotiation or a dispute. If all that's needed is a number for a presentation slide, AI is probably the right tool today. The moment that number has to hold up in deal negotiations, it no longer is.

Where AI Genuinely Helps in Company Valuation

None of this means AI has no place in the company valuation process — its role is simply narrower than the marketing suggests. AI is excellent at the mechanical layer:

• running dozens of scenarios and sensitivity analyses in the time it takes to write a short summary;

• catching inconsistencies in long financial models and documents;

• drafting once the substantive decisions have already been made;

• ranking shortlists a human has already prepared (for example, comparable companies).

Prudentia's Approach: AI as a Tool, Not a Decision-Maker

At Prudentia, we use AI throughout our valuation process for exactly these tasks — faster iteration, broader scenario testing, quicker first drafts. But every valuation we deliver is still built on due diligence our team has done itself, on shortlists we've curated, and on a human advisor who decides which assumptions actually hold up, which are simply what the client wanted to be true, and what that number actually means for the negotiation ahead.

AI makes the process faster. It doesn't yet make it safe to skip.