How an AI Product Proves Value Without Exposing Its Stack (Illustrative Case Study)

1. Case Studies

By Saroj Jha · 7 min read

A fictional AI product demos well and then stalls in the buyer’s security review. An illustrative eight-week plan to disclose what buyers can check, protect what a competitor could copy and prove value on the buyer’s own work.

ProofDesk is a fictional AI product, and this is an illustrative sample engagement: the numbers are modelled projections, not client results. It shows how AAJ would approach a problem many AI founders have, a product that demos well and then stalls in the buyer’s security review or loses to the buyer’s plan to build the same thing in-house with a chatbot.

In my view, when buyers cannot check how an AI product works or how well, the fix is proof, not more adjectives. In this example the team was hiding what buyers needed (the model provider, data use, retention) and showing what competitors could copy (prompts). The plan reverses that: disclose what a buyer can check, keep private what a competitor could copy, and prove value on the buyer’s own work.

Snapshot

The situation

ProofDesk had a product that worked and a sales problem it could not name. Demos went well. Then deals stalled in the buyer’s own security review, or ended with the buyer deciding to build it in-house with a chatbot. The founders answered the first problem by saying less (“proprietary AI”) and the second by saying more (sharing prompt examples in sales calls). Neither worked: the first made security teams suspicious, and the second taught prospects how to copy the product.

Phase 1: Diagnose (weeks 1 and 2)

The aim is to find where deals die and why, from evidence rather than opinion.

• Win-loss review of the last 20 decided deals. Short calls with the buyer, not the rep. The pattern: buyers lost confidence in two places, the security review (they could not tell where their data went) and the comparison with a free chatbot (the product looked like a prompt they could write themselves).
• The security review, read line by line. Five questions came up in almost every review: which model provider, whether customer data trains it, where data is stored and for how long, who checks the output, and how accurate it is. ProofDesk had no written answer to the last three.
• What the team was saying. Sales calls swung between “proprietary AI” and showing actual prompts. Neither answered the buyer’s real question: what does ProofDesk do that a general chatbot leaves to me?

Diagnosis: the product was not the problem. The proof was. Buyers could not check how it worked or how well, and the team was hiding what buyers needed and showing what competitors could copy.

Phase 2: Design (weeks 3 to 5)

The aim is to decide what to disclose, what to keep private, and how to prove value on the buyer’s own work.

1. A disclose-and-keep-private line. Disclose what a buyer can check: the model provider, that customer data does not train it, retention, where a person reviews the output, and the evaluation method and results. Keep private what a competitor could copy: prompts, retrieval rules, the full evaluation set and architecture beyond what a review needs. Offer a penetration test report under a confidentiality agreement. Model providers publish their own data terms, and a buyer will read them: OpenAI, for example, says data sent to its API is not used to train its models unless a customer opts in, and keeps abuse monitoring logs for up to 30 days by default^[1]. A retention answer has to include that second copy.
2. An evaluation buyers can read. 300 real questionnaire questions from consenting customers, answered by ProofDesk and checked by a security reviewer, reported as two numbers: answers accepted unchanged, and answers the reviewer had to correct, plus a limits page for question types where it does worse.
3. A timed side-by-side. One real 150-question questionnaire, done three ways: by hand, with a free chatbot, and with ProofDesk. Report time and corrections for each, and say plainly what the chatbot does well.
4. A pilot with a decision date. One questionnaire the buyer actually owes a customer, a success bar agreed in writing (time saved and share accepted unchanged), and a decision date.

Phase 3: Execute (weeks 6 to 8)

The aim is to put the proof in front of buyers before they ask.

The claims review matters beyond sales. In August 2025 the FTC approved a final order against Workado, LLC, which had advertised the accuracy of an AI content detector; the order bars it from advertising the accuracy or efficacy of such products unless it has competent and reliable evidence that they are as accurate as claimed, and requires it to retain that evidence^[2]. In my view an accuracy number belongs on the page only if a dated test supports it and the test is kept.

AAJ resources used: the Marketing AI Products Playbook (Steps 2 to 6), the Win-Loss Analyzer, the Security Questionnaire Template and Answer Library, the Claims and Proof Checker and the Pilot-to-Contract Planner.

Projected results, two quarters after launch

These are projections, not historical results. They assume the same 40 qualified opportunities a quarter and the same price, so all of the change comes from fewer stalled reviews and more wins. On a real project the targets would be set from the company’s own win-loss data, and the first quarter would be read as an early signal, not a result.

What has to go right

This is AAJ’s arithmetic on the example. At baseline 35 percent of 40 opportunities, 14 deals, stall in security review. Falling to 15 to 20 percent means 6 to 8 deals, so 6 to 8 fewer stalls. A win rate of 24 percent is 2.4 more wins a quarter than 18 percent, so 30 to 40 percent of the unstalled deals have to close. A win rate of 27 percent is 3.6 more wins, so 45 to 60 percent of them have to close. Both are well above the 18 percent baseline win rate, which is plausible only because a deal that clears security review is further along than the average deal. That is an assumption, not a finding. In my view the 24 percent case is the one to plan on, and 27 percent is the best case.

What carries over to any AI product

• Disclose what a buyer can check; keep private what a competitor could copy. Hiding the model provider costs trust; showing prompts costs your advantage.
• Prove it on their work. An evaluation with its method, and a pilot on a real task with a decision date, beat any adjective.
• Concede what the free tool does well. It makes the rest of your case believable.
• Answer the security review before it is asked. A data-use page sent with the proposal turns a stall into a checkbox.

Method and limits

ProofDesk is fictional and every figure about it is illustrative; $18,000 a year and 40 opportunities a quarter are assumptions. The two outside sources are an example of a provider’s published data terms (read on October 22, 2026, and subject to change) and an FTC order about one company’s accuracy claims; neither shows what a questionnaire tool would achieve. This is general information, not legal advice.