In short: AI pilots die from security review, missing owners, demo-grade accuracy and pricing confusion. Design the pilot to pass all four before you sign it.
Key points
- Scope one use case, such as drafting first replies for tier-one billing tickets in English.
- Ask for real data in week one; a sanitised sample proves the demo, not the product.
- Agree a written success test with metric, threshold and method, plus a failure clause.
- Build a dated security pack covering compliance status, data flow, model and vendor disclosure, retention, access and incidents.
- Charge for the pilot; free pilots drift.
Why AI pilots die
Pilots rarely die from the product. They die from everything around it.
Security and data review. Where does the data go, who sees it, is it used for training, how long is it kept? If you cannot answer in writing, the pilot waits for you to write it.
No owner. A sponsor who likes the demo is not an owner. An owner has budget, a deadline and something to lose if it fails.
Demo accuracy versus production accuracy. The demo ran on clean examples. Production has typos, edge cases and missing fields. The gap surprises the buyer and confidence drops.
Procurement. Vendor onboarding, data processing agreement, insurance. A parallel track that nobody started until month three.
Pricing confusion. Seats, usage or outcomes? If finance cannot forecast the bill, finance will not approve it.
Scope one use case and get real data early
"Explore AI for the support team" is not a pilot. "Draft first replies for tier-one billing tickets in English" is.
Narrow scope makes the security review smaller, the success test clearer and the owner easier to find. It also makes the result believable. A win on one ticket type is a reference. A vague win on "support" is a slide.
Ask for real data in week one. A sanitised sample proves the demo, not the product. If real data access takes eight weeks, that is the pilot's real timeline. Better to know now than to discover it in week six.
Write the success test before you start
Agree, in writing, what passing looks like. Three parts: metric, threshold, method.
Example shape: over a fixed volume and a fixed period, what share of outputs are used with minor or no edits, as judged by whom. Pick the number with the buyer, not for them. Then neither side can move it later.
Add a failure clause. If the test fails, the pilot ends and both sides say so. A pilot with no way to fail has no way to win either.
The security pack
Build it before the first enterprise call. Most enterprise security questionnaires ask the same things. Answer them once, in plain language, and keep it dated.
| Section | What to include |
|---|---|
| Compliance status | SOC 2 or ISO status, or a plain timeline if not yet |
| Data flow | One diagram: what enters, where it is stored, where it leaves |
| Model and vendor disclosure | Which models, which providers, whether data is used for training |
| Retention and deletion | How long, where, and how a customer gets it deleted |
| Access control | Who at your company can see customer data, and when |
| Incident process | Who gets called, how fast, what they get told |
If you are not SOC 2 yet, say so and say when. Buyers forgive gaps. They do not forgive surprises in week five.
The model and vendor section matters most for AI. Buyers want to know whose servers see their data and under what terms. Have the sub-processor list ready.
Owner, money and price
Named owner. Get a name, a title and a budget line before you start. If nobody will own it, nobody will buy it.
Paid pilot. Charge something. The amount matters less than the invoice. A paid pilot gets a deadline, a purchase order and an owner, because someone had to approve the spend. Free pilots drift.
Price after, structure before. Fix the pilot fee now. Agree the full-contract price after the success test, but settle the structure up front. Seats, usage or outcomes. Pick one and explain it in one paragraph. Usage-based pricing on an AI product needs a cap or an estimate, or finance will block it on principle.
The checklist
- [ ] One use case, one team, one language, written in a sentence
- [ ] Real data access agreed for week one
- [ ] Success test with metric, threshold and method, signed by both sides
- [ ] Security pack complete and dated, including model and vendor disclosure
- [ ] Named owner with budget, and a pilot fee on an invoice
- [ ] Pricing structure chosen and explained before the pilot starts
Common questions
Why do enterprise AI pilots fail?
They rarely die from the product. They die from security and data review, no real owner, the gap between demo and production accuracy, procurement started late, and pricing finance cannot forecast.
What should an AI vendor security pack include?
Compliance status, a data flow diagram, model and vendor disclosure including whether data is used for training, retention and deletion, access control, and the incident process. Keep it dated and have the sub-processor list ready.
Should AI pilots be free or paid?
Paid. The amount matters less than the invoice: a paid pilot gets a deadline, a purchase order and an owner, because someone had to approve the spend.
What if my startup is not SOC 2 compliant yet?
Say so and say when. Buyers forgive gaps, but they do not forgive surprises in week five.