Two-stage proof

A case says something is worth trying. Proof decides whether to keep it. The Deployment Framework proves everything Build produces in two stages, a sandbox and then a limited live trial, and each item comes out of the trial kept or killed on the evidence.

Most pilots end by drift

Plenty of AI pilots never become anything. That on its own isn't the problem.

Stopping a pilot is often the right call. The waste is in how most of them stop: late, by drift, with nobody having said in advance what success would look like. MIT's researchers found most failed because of brittle workflows, tools that didn't learn, and a poor fit with how people actually work. Those are exactly the things a properly designed trial would have shown early, and cheaply.

Stage 1: a sandbox that can't reach live systems

The first stage happens somewhere safe: the pre-approved sandbox from your guardrails, running on test data or on copies the data map allows, with no connection to anything live. It answers the cheap questions first. Does it do the job at all? What does it get wrong? What does it cost to run?

The important word is can't. In July 2025, an AI coding agent being tried out on a real app deleted the live production database during a declared code freeze, wiping records for more than 1,200 executives and nearly 1,200 companies. It had been told not to make changes. The instruction was there, and so was the access.

Stage 2: a limited live trial

Once the sandbox says it's worth going on, the build is promoted to a limited live trial: a few real users, doing real work, in the systems they already use. Promotion is a decision, not a drift. The builder shows what the sandbox found, the owner agrees, and the trial's limits are set in the tools: who it can reach, which systems, and for how long.

• Limited means a small group, a fixed number of weeks, and an off switch anyone in the trial can use.
• Live means real work and real consequences, inside the trial's fence.
• In the systems they already use means not a demo, and not a separate app people have to remember to open. A tool that only works in a demo hasn't been proved.

What the proof has to show

Three things, written down before the trial starts.

• It works, measured against the value in the case, in the same unit. If the case promised eight more quotes a week, the trial counts quotes.
• It's fair, wherever it helps decide about people. Results are checked by group, and anyone it screens out is reviewed.
• It's safe. Data stays where the data map allows, it stops itself when its output looks wrong, anyone can stop it, and what it changed can be undone.

For anything high-stakes, the evidence comes from someone who didn't build it. Builders aren't dishonest. They're just the people least able to see what they didn't think of.

Kept or killed

Each item leaves the trial with one of two answers. Kept: it gets its full reach and goes to Own, with an owner, a backup, an off switch and a review date. Killed: it stops, its access is removed, and what was learned goes on the friction list. Both are good outcomes, as long as they happen on time.

That depends on writing the kill criteria before the trial begins, alongside the stop-loss from your case. Once people have spent weeks on something, every result starts to look promising.

New offerings run as experiments

Some builds aren't about the bottleneck or a cost at all. They're bets on something new for customers. The framework runs those as experiments, paid for from the experiment budget on your aim page, and borrows the method openly from Lean Startup (Eric Ries, 2011). Each experiment names five things:

• the customer problem it addresses
• its riskiest assumption: the belief that, if it's wrong, sinks the whole idea
• learning milestones: what you'll know, in order, with a chance to stop at each one
• kill criteria
• a customer owner, who answers for what customers experience

Experiments go through the same two stages. Inventing whole new business models is outside the framework: an experiment tests one offering, with a budget and a way to stop.

Plan the proof for your case

The plan below picks up the case you wrote in lesson 3: its stakes, its builder of record and its kill criteria. If the case is high stakes, it asks for someone independent to check the evidence.