What Makes an AI-Assisted Valuation Defensible?

 

Carl Losito and Chau Hoang discuss why AI makes human judgment matter more in valuation and what management teams, investors, and counsel should be asking now.

Artificial intelligence (AI) can pull comparable companies, build a discount rate, and summarize a data room in minutes. But checking whether any of that is right takes more time.

Valuation has always been part science, part art. AI is making the science faster but the art more essential.

BRG Director Carl Losito and Managing Director Chau Hoang from the firm’s Transaction & Valuation Opinions practice discuss what they’re asking management teams, where AI-assisted inputs break down, and what makes a valuation defensible when a machine helped build it.

Valuation has always been part art, part science. Does AI change that balance?

Chau: Valuation combines inputs from the company, market, and research data with judgment built from diligence and conversations with management. The risk is that people expect it to be plug-and-chug: data in, number out. I’d argue AI makes valuation more of an art. A model can produce something that looks finished, and someone still needs to explain what it did and whether it reflects reality.

Carl: How do we use AI’s speed while making sure the logic holds against what’s happening in the business? I recently reviewed a letter of intent that said a company was founded in 2000. It was founded in 2020. Because I knew the business, I caught it and told the AI to double-check, and it agreed it was an error. At face value, that date would have turned a six-year-old company into a twenty-six-year-old one.

What happens when the same procedure values a company at $100 million in June and $75 million in September?

Carl: We haven’t seen a swing that dramatic in practice yet, but I expect we will. We often value the same private company at several points in time. Picture a business that looks healthy on paper: revenue is growing, operating profit is improving, and management is executing on its initiatives. Then the model produces a lower value in September because dozens of inputs moved slightly. The AI can list what changed but can’t tell you what drove the change in the real world.

Expert judgment is critical. You must recognize that the result doesn’t align with what’s happening at the company. Sometimes that means rethinking the methodology and incorporating factors that weren’t in the model before.

Chau: A mechanical model reacts to every input, up one quarter and down the next. But human judgment looks at the aggregate and may say that it’s a wait-and-see moment. When value moves that much, the first question is what’s driving it. The answer usually comes from iterative conversations with management and the client. Is the movement really about the inputs, or do we need to understand a broader story? We want that context before making adjustments based on one quarter’s swing in market data.

Inside the black box: How do you check an AI-generated discount rate or comp set?

Chau: Someone can tell AI to pull the comps and produce a discount rate, and on the surface it all looks right. Then you notice the rate is x when you know it should be closer to y, and you have to work backward through the black box to see how it got there. Calibrating backward often takes more effort than building forward.

Every firm is going to need quality-control practices that work for its own engagements. That’s the biggest concern I hear at the managerial level: If people aren’t doing the work, how do they know what to look for? You learn to recognize a mistake by struggling through the work. If AI does the work and the output sounds right, how do you recognize that the comp set was wrong?

Carl: Some parts of a discount rate are set by the industry and valuation date. But a buildup also includes a company-specific risk premium and a judgment about whether this business is riskier or healthier than its peers and why. A human is on the hook for that call.

Comps are even less negotiable. When I challenge an AI tool on a comparable, it usually folds in seconds. We can’t do that. We must have conviction in every public and private comp we use.

What should CFOs and management teams do when providing data?

Carl: Data rooms need better governance. In the past, a person went through the room and indexed everything carefully: the final model, the executed letter of intent, a draft bid that never went anywhere. We only used what we knew to be factual.

If everything is dropped into a room and a bot is sent to organize it, the bot might mislabel a document or not flag that one version is a draft, and another was signed. A person knows to ignore drafts one and two and work from draft three. A model may blend all three. Clients should send final or nearly final data.

Chau: Many clients now use AI to build the models they hand us. If those models aren’t traced back to verified data, errors compound with each iteration and become increasingly hard to unwind.

Who verifies the verifier?

Carl: The accountability structure doesn’t change. On a typical engagement, junior analysts run the modeling, a manager oversees day-to-day execution, and senior leaders like Chau and me own the work and the client relationship. We’re the goalkeepers: nothing goes out without our approval.

The analysts’ job does change. They’ll spend less time scrubbing data themselves and more time making sure it’s clean at the source, which means more communication with whoever on the finance or deal team is providing it.

High-stakes opinions also go to independent committees of senior professionals who weren’t on the deal team. I’ve heard that some banks and sponsors are now adding AI to their deal committees as observers, flagging when one deal’s assumptions don’t match the last one’s. Accountability will run both ways.

Chau: The same test applies to our own tools. When an internal AI tool summarizes market data, I ask where it’s coming from, whether it cites underlying analyst reports, and whether anyone has checked the numbers against the source. The risk is that an analyst reviewing work they’ve never done themselves won’t know what to look for. AI makes everything look polished, which isn’t the same as accuracy.

We may end up with analysts who are highly efficient but miss fundamental learning that makes them better at the job. People coming out of college, who haven’t done the work themselves, are being asked to manage AI and decide whether its output looks right.  You don’t know what you don’t know. Every organization has to learn how to train people to use AI without losing the fundamentals.

What happens when AI-assisted valuations end up in disputes?

Carl: An attorney recently described an opposing party that started with the answer it wanted and asked AI to make the number work. The AI didn’t refuse or flag a problem. It found a way, and the support looked legitimate on the surface. On closer review, the methodology wasn’t standard or supported by recent market transactions. A bot will take that chance because it has no professional reputation. It’s still a gray area how much of a bad number comes from the AI and from human manipulation.

Chau: AI in valuation is relatively new, and disputes take time to surface. When a valuation someone relied on falls apart, there’s often a lag, and litigation itself can take much longer. Valuations performed this year and next that end up in litigation will reveal how courts treat AI-assisted work, but that picture may not come into focus for some time.

What should investors, boards, and counsel ask about any AI-assisted valuation?

Carl: To assess whether a valuation is defensible, ask:

• Can every key input be traced to verified source data?
• Was the data final, or could earlier drafts have mixed in?
• Is there an explanation of what drove each change from the prior period, not just a list of inputs that moved?
• Is the methodology standard and supported by recent market transactions?
• Who with real knowledge of the business reviewed the output and signed off?
• Does the engagement letter say how AI was used and how confidential data was protected?

Chau: I’d add two questions. First, what diligence was performed alongside the AI, and what level of quality control was applied? Second, how much of the valuation did AI actually do? The answers depend on the situation. In a recurring portfolio valuation, AI may handle much of the work. In a transaction worth hundreds of millions of dollars, it might help organize data or build charts but isn’t doing the valuation. The level of scrutiny should match the level of AI use.