Ask ChatGPT to explain how business valuation works and it will give you a genuinely useful answer. It can describe EBITDA multiples, walk you through discounted cash flow mechanics, explain the difference between enterprise value and equity value, and outline the factors that cause one company to trade at a premium to another. The explanation will be coherent, accurate at a general level, and well-organized.

Now ask it to value your business. Something categorically different happens.

The model will produce a response that looks like a valuation. It will use the right language, reference relevant methodologies, and may even generate a range of numbers. But those numbers will not be the result of applying a disclosed methodology to your verified financial data. They'll be a generative output — a plausible continuation of text about valuations — calibrated to nothing specific about your company.

Understanding exactly why this happens, and what a legitimate valuation tool requires instead, is essential for any business owner or operator who wants to use AI intelligently in a sale, capital raise, or strategic review process.

What ChatGPT Actually Does When Asked to Value a Business

Large language models are trained on enormous corpora of text — including financial reports, M&A analyses, investment banking materials, and business journalism. Through this training, they develop sophisticated pattern-matching capabilities around valuation concepts. They can recognize the structure of a valuation argument, reproduce the vocabulary of deal-making, and generate text that has the surface characteristics of a financial analysis.

When you prompt an LLM with your business description and ask for a valuation, the model is doing something like this: it's drawing on patterns from valuations it has seen in training data and constructing a response that fits the pattern of what a valuation of a business like yours tends to look like in text.

The critical distinction: An LLM's valuation output is shaped by the distribution of valuation-related text in its training data — not by applying a defined methodology to your actual numbers. The output looks like a valuation. It is not one.

This matters because business valuation is not primarily a conceptual task. It's a numerical task that requires specific inputs: verified revenue figures, normalized EBITDA, documented growth rates, customer concentration analysis, recurring revenue percentages, and comparable transaction data for your specific sub-sector and size range. An LLM that hasn't been given these inputs — or given them in verified form — cannot produce a reliable valuation, regardless of how confidently its output reads.

What Business Valuation Actually Requires

The mechanics of business valuation for private mid-market companies center on a few key variables:

Verified financial inputs

Revenue and EBITDA are the starting points, but the quality of those numbers matters as much as the numbers themselves. Buyers and their advisors will normalize EBITDA — adding back non-recurring expenses, owner compensation above market rate, and other adjustments — and those add-backs need to be defensible. A valuation built on unexamined self-reported numbers will be renegotiated when the numbers are verified.

Industry-specific transaction multiples

EBITDA multiples — the most common valuation anchor for private mid-market businesses — vary enormously by industry, company size, business model characteristics, and market conditions. A software business with high recurring revenue trades at a fundamentally different multiple than a services business with project-based revenue of the same EBITDA. A business with 40% EBITDA margins in a growing market trades differently than one with the same EBITDA in a declining market. These differences are not small. They can determine whether a company trades at three times EBITDA or eight times EBITDA.

Real multiple data comes from actual closed transactions — comparable company sales, reported in proprietary databases maintained by investment banks and M&A advisory firms. This data is not publicly available in a form that a general LLM can access, verify, and apply to your specific situation.

Company-specific adjustments

Even with the right industry multiple range, individual company characteristics move valuations within and beyond that range. Customer concentration (how much revenue comes from the top one to three customers), management dependency (how much of the business value is tied to a single individual), intellectual property ownership, recurring revenue percentage, contract structure, and employee retention factors all affect where in the range — or outside it — a specific business lands.

The multiple problem: A general LLM, asked for a valuation, will often produce a "typical" or "average" multiple range. For your specific business, this number may be directionally wrong. A software company with 90% recurring revenue in a fragmented market does not trade at the same multiple as an IT services firm with 20% recurring revenue. Using an averaged range as a negotiating anchor can mean leaving significant value on the table — or entering a process with misaligned expectations.

Why General LLMs Can't Hold Relevant Comparable Data

The most fundamental limitation of AI tools for business valuation isn't intellectual — it's informational. The relevant comparable transaction data simply isn't available to general-purpose LLMs in a useful form.

Closed private company transaction data is proprietary. It's collected, curated, and licensed by specialist data providers, and accessed by investment banks and M&A advisors as part of their practice infrastructure. This data is current — market conditions in 2024 are different from 2021 — granular by sub-sector, and specific enough to distinguish between business models within the same broad industry category.

An LLM trained on publicly available text will have encountered discussion of valuation multiples in various contexts — business journalism, analyst reports, some disclosed deal terms — but this represents a fraction of actual transaction data, is not uniformly current, and is not organized in a way the model can reliably query against your specific situation. When a general LLM produces a multiple range, it is generating a number consistent with the valuation language it has seen, not retrieving from a verified comparable transaction database.

General LLM Approach

Pattern-matched estimate

Generates a multiple range based on training data patterns. No verified comparable transactions. No vintage date. No sub-sector specificity. No methodology you can audit or defend.

Deterministic Engine Approach

Calculated from inputs

Applies a disclosed formula to your verified inputs. Uses benchmark ranges from identifiable sources. Produces ranges that reflect your company's specific characteristics. Fully auditable.

What a Deterministic Valuation Engine Does Differently

A deterministic valuation engine applies a set of disclosed rules to a defined set of inputs. The methodology is explicit — you can see which factors are weighted, how adjustments are applied, and how the final range is computed. Change an input; the output changes in a traceable, predictable way.

This isn't just a philosophical preference — it has practical consequences for every use case where you need to rely on a valuation estimate.

When you present a valuation range to a potential buyer, a financial partner, or a board, you need to be able to explain how you arrived at it. "The AI said" is not a defensible answer. A calculation with disclosed inputs and methodology — "this range is based on our normalized EBITDA of X, applied to a multiple range derived from comparable transaction data in our sub-sector, adjusted for our recurring revenue percentage and customer concentration profile" — is.

Deterministic engines also produce ranges rather than false precision. A legitimate valuation output acknowledges the genuine uncertainty in private company valuation. A specific number — "your business is worth $14.7M" — implies a precision that doesn't exist in the valuation of a private company. A methodology-grounded range — "based on these inputs and this methodology, our indicated range is $12M–$17M, with the upper bound dependent on resolving the customer concentration issue" — reflects the actual nature of the exercise.

The Right Role for AI in Valuation Work

This is not an argument against using AI anywhere in valuation or M&A processes. AI tools are genuinely valuable for specific tasks within a valuation process — tasks where synthesis, pattern recognition, and text generation add value without requiring verified numerical precision.

Where AI earns its place in valuation work:

What AI should not provide is the core numerical estimate — the multiple range, the EBITDA figure, the valuation range itself. These require verified inputs, current comparable data, and auditable methodology. That's the domain of deterministic calculation, not generative text.

The practical discipline: Use deterministic scoring engines for the numbers you need to defend. Use AI synthesis for the framing and narrative around those numbers. Label each type clearly. Never allow a generated estimate to substitute for a calculated one in a context where the number will be relied upon.

Preparing for a Real Valuation Process

If you're planning a sale, capital raise, or simply want to understand where your business stands, the starting point is a structured assessment of the factors that drive valuation in your context: the quality and defensibility of your EBITDA, your recurring revenue profile, your customer concentration, the transferability of your management and customer relationships, and your growth trajectory relative to your industry.

This assessment, done rigorously and honestly, tells you two things: your likely valuation range based on current business characteristics, and the specific factors that would most improve that range if addressed before a process. This is actionable information. An AI-generated number that you can't trace to its inputs is not.