Why Your Advisory Firm's Data Is Valuable for AI Training
An HVAC business is for sale. The listing shows $300,000 in annual seller’s discretionary earnings. A buyer who plans to keep their day job asks, “How much of that can I put in my forecast?” My first question would be, “Who runs the shop after the seller leaves?”
For an acquisition advisory firm, the answer might be spread across the listing, a note from a call with the seller, and an analyst’s workbook. Put them together and you can show an AI developer how an underwriter reached a decision. Whether it makes sense to license that work depends on what the firm can share, what the buyer plans to build, and what the firm gets in return.
Two HVAC listings. Both say $300,000.
In these fictional deals, the buyer will hire or retain someone to run the business. The starting figure is seller’s discretionary earnings (SDE). Any pay to the seller for working in the business has already been added back. We’ll hold the other assumptions steady and change one fact.
The seller runs it
The seller schedules crews and manages the technicians. When the seller leaves, a new manager will cost $90,000 a year, including benefits and employer costs.
$300,000 − $90,000 = $210,000Annual earnings after hiring the manager
A manager runs it
The seller is hands-off. The manager will stay, covers those same duties, and their full compensation is already in operating expenses. Deducting it again would count the cost twice.
$300,000 − $0 = $300,000Annual earnings with the manager already paid
What if the seller just says, “I help out”? Ask which duties the seller performs each week, who will take them over, and whether that person’s pay is already in expenses. Leave the adjustment open until those facts are known.
Both figures are before debt service and taxes. A buyer estimating cash available to take home would also need to account for items such as equipment purchases and working capital.
What would the AI developer actually receive?
One training example could fit on a page. Give the model the listing, the relevant seller interview note, and the buyer’s question. Pair those inputs with the answer the analyst signed off on. “Use $210,000. The seller’s pay was added back, the buyer won’t do that work, and the replacement manager is a new expense.” The second case shows when the same deduction would be an error. The incomplete case teaches the model when to ask for evidence.
For training, a developer can use those inputs and expert answers together. For evaluation, reserve fresh deals the model hasn’t trained on, give it only the inputs, and grade its responses against the advisor’s answer key. Give it credit for spotting a missing fact even when there’s no number to put in the forecast yet.
There’s a real example of the evaluation format. Vals AI pairs financial modeling tasks with expert-reviewed Excel workbooks and grading criteria. It lists 51 validation examples as available for license. That gives an advisory firm a concrete comparison for packaging its work, although it doesn’t establish a price for an HVAC case or for training rights. See the Excel Modeling Benchmark methodology.
When would I take the licensing deal?
I’d start with something the firm already uses to teach a new analyst. An original exercise with fictional figures, a working spreadsheet, and a partner-reviewed answer gives a prospective buyer something specific to assess. Then I’d ask three questions before putting a price on a larger collection.
- Which parts can the firm license? A firm might own the workbook it designed while a client supplied the payroll report and a broker supplied the listing. Check the engagement terms, confidentiality agreements, and permissions for each source and the intended AI use. Removing a company name still leaves that permissions question.
- What can the buyer do with the examples? A lender testing an internal tool and a software company training a deal screener for your prospective clients are asking for different rights. Would the second product take over work your clients pay you for? The fee and the scope should reflect that answer.
- What will it cost to deliver a reliable set? Estimate the analyst hours to recover the inputs, check the math, document the judgment, and resolve permissions. If a partner has to reconstruct every decision from memory, include that partner time in the price. If the workbooks and explanations are already paired, a license gives the firm a way to earn revenue from work it has already paid for.
My starting offer would be a small paid pilot with firm-authored cases and a clearly defined use. A standardized adjustment the firm teaches every new hire could be a good first candidate. I’d want a much stronger offer to share a proprietary underwriting method that regularly wins client engagements.
Start with one decision your analysts document
Pick a case and list the source records, the answer, who reviewed it, and who controls each file. Keep the underlying client documents in place while you take inventory. Open the Business Data Checklist.
If you already have a collection you can describe and permission to offer it for the proposed use, explore data licensing for your firm.
