Codifying 25 years of due diligence — without letting the AI touch the arithmetic

Challenge

Valuation and due diligence are where AI is most tempting and most dangerous. A language model will produce a discounted cash flow that reads beautifully and is arithmetically wrong, and it will do so with complete confidence. In a transaction, a dispute, or a court, that is not a quirk. It is a liability.

The underlying problem is that due diligence is genuinely expert work. It is not a checklist; it is knowing which question to ask next given what you have already found, and knowing when you have enough to form a view. That judgement is exactly what does not survive being turned into a form.

So the challenge was to encode the method without handing the method to a model that cannot be audited.

Approach

We separated the two things that AI conflates: judgement and computation.

The valuation engine is deterministic. Discounted cash flow, market multiples, precedent transactions and net asset value are each computed in code, with the inputs, assumptions and workings exposed. Intellectual property valuation runs relief-from-royalty, multi-period excess earnings and cost approaches on the same basis. The language model is never permitted to perform arithmetic — it drafts narrative around figures it did not calculate, which means the narrative can be wrong and the numbers still cannot be.

Workpaper output is cited to the applicable accounting standards, so a reader can trace the treatment rather than take it on trust.

On top of the engine sits a forensic control sheet: thirty-one controls across seven domains, with a court-readiness gate. An engagement either clears the gate or it does not, and if it does not, the report says so rather than quietly proceeding.

The question-first architecture handles the judgement layer. Rather than a fixed checklist, the system works out what to ask next given the evidence so far, updating its view as material arrives. InvestorView publishes a 98% sufficiency target as the standard it works toward. Ding Ventures applies the same engine commercially as advisory for fintech and AI ventures — the method sold as an outcome rather than licensed as a tool.

Outcome

Both properties are live. InvestorView publishes the methodology and its 25-year provenance; Ding Ventures operates the advisory practice built on it.

The practical consequence for a business owner is straightforward. When we value your business — for a sale, a succession plan, a dispute or a shareholder agreement — we are running an engine we built, whose workings we can show you line by line, and whose arithmetic no model has been allowed near.

It also settles a question clients increasingly ask: are you using AI on my numbers? Yes, and here is precisely where it is allowed to operate and where it is not. That boundary is not a policy we wrote. It is architecture.