Six rules that decide whether an AI-built product survives production. Full text, with checklists you can run today, whether you built the thing yourself or it landed on you when the builder moved on. You can read the guide without giving us your email. The hard part is applying the rules.
AI Delivery is the work of taking AI-generated products to production reality.
AI Delivery checks the generated work and finishes what production requires. It also assigns an owner. At the end there’s a product that works and someone who owns it.
AI made building cheaper. Verifying the result still takes work. The gap between what runs in a demo and what holds on a Monday is where most AI products stall. The six rules cover the work needed to close that gap. Each rule comes from a production problem we have seen.
These are the same rules published on the homepage, word for word. Read each one, then run its checklist against your own product, honestly. The "you" here is whoever holds the risk now: the founder who prompted the thing into being, or the person it landed on when the builder moved on. Start with any box you cannot check.
You asked AI if your product was ready. It said yes. You asked again, differently, forty times. Yes, every time. And somewhere in there, a quiet voice said: in theory.
Take that voice seriously. Not because doubt is always right, but because it points at something untested. Doubt isn’t a verdict. It’s a question that hasn’t been run yet.
So we run it. Every scope starts by turning your doubts into tests: name the fear, define what would prove it wrong, go find out. Doubt in, evidence out. Bring it here. It’s on the sign.
The doubt: “our quote tool gets prices right.” What would prove it wrong: any quote more than 5% off the price sheet. The test: twenty real past orders, quoted again on a test copy. The result: three were off, all of them discounted orders. That’s a fail, and now you know exactly where to look.
Can’t run one of these yourself? That’s normal, and it’s the work we do. We can run it with you. Let’s talk.
The demo is real. It got people leaning in, and that matters. But shipping means surviving what demos never face: a security questionnaire, a traffic spike, a Monday.
AI-built apps land about 70% complete, polished enough to look nearly done. The missing 30% is the unglamorous part: auth, data integrity, scale, cost per request, fallbacks, what happens when the model is wrong. Columbia researchers built more than fifteen applications with AI coding agents and sorted hundreds of failures into nine patterns. The two worst were error handling and business logic.
That’s the part we build. It’s easy to defer from the inside, and it decides everything.
A booking app scored 5 for looks and 2 for works. The new user couldn’t find the cancel button, and a blank phone number field showed an error in code instead of a message. Neither shows up in a demo. Both show up on a Monday.
Can’t run one of these yourself? That’s normal, and it’s the work we do. We can run it with you. Let’s talk.
Somewhere in your company there’s a product AI built that nobody can fully explain. It runs. It matters. And if it breaks at 2am, no one’s name is on it.
Ownership isn’t a feeling. It’s a spec: a named accountable person, documentation your team actually runs on, monitoring, an incident plan, and a decision about who answers when the model or the vendor changes underneath you.
Twenty-five years of scoping comes down to one rule: if you can’t support it, you’ve got no business shipping it. Everything we deliver ships with an owner. In writing. No orphans.
A team found its test alert went to the inbox of a contractor who had left in the spring. The system was fine. The only person it would have woken up was gone.
Can’t run one of these yourself? That’s normal, and it’s the work we do. We can run it with you. Let’s talk.
AI can make a confident case for almost any answer. Confidence is not evidence. A bluff is exactly that: confidence without evidence. So we don’t argue. We test.
The rubric is written and disclosed: architecture under load, security under attack, the workflow with real users in it, costs at real volume. The same dimensions every time, with tests and thresholds matched to the product’s risks, so the verdict comes from evidence, not from a veteran’s hunch.
And because we also build, we say it plainly: the scope is priced to stand alone, no-go is always an acceptable answer, and the verdict is yours to take anywhere. Judgment you can’t afford to hear isn’t judgment.
A row you can run yourself: “My second test account can’t open my first one’s records.” It’s a limited check, not proof of privacy, so the full privacy row stays with a qualified reviewer. A row you probably can’t: “Holds three times our busiest hour. Evidence: the load-test report.” The second one stays unverified until someone qualified runs it.
Can’t run one of these yourself? That’s normal, and it’s the work we do. We can run it with you. Let’s talk.
Working isn’t the same as worth it. Plenty of AI systems run perfectly and produce nothing: the demo impressed, the workflow never changed, the money never noticed.
So we measure the boring way: against the non-AI alternative, in real workflows, with real users, counting the errors and what they cost. Before anything gets built, we agree in numbers what paying off means.
If it doesn’t pay, we say so. That’s the cheapest sentence we’ll ever sell you.
Answering support emails took 6 minutes each by hand. With AI drafting it takes 2 minutes, plus 1.5 minutes to check the draft. That saves 2.5 minutes per email. At 400 emails a month, that’s about 17 hours. If the tool costs more than those 17 hours are worth to you, it doesn’t pay yet, however well it writes.
Can’t run one of these yourself? That’s normal, and it’s the work we do. We can run it with you. Let’s talk.
You know the loop: one more prompt, one more version, almost right, again. The loop feels like progress. It’s motion.
Done is not a feeling. Ask any artist: putting the brush down never feels right. Done is a line drawn in advance: acceptance criteria agreed before the iteration starts, and kill criteria for when to stop entirely.
We draw that line before we begin, and we hold the work to it. Ours included. Done is a gate, not a wall: the product ships, gets owned, and earns its next round. Closure ends the lap, not the road. That’s the job.
Done means: the support bot gets at least 19 of our 20 most common questions right, on each of three separate test runs, by the end of the month, for no more than $2,000. Decided by the operations lead. We stop if it’s below 15 of 20 after two more rounds, or spending passes $3,000, and we keep answering by hand.
Can’t run one of these yourself? That’s normal, and it’s the work we do. We can run it with you. Let’s talk.
AI delivery is the work of taking AI-generated products to production reality: checking what the AI built, finishing what production requires, and assigning an owner, so the result is a product that works and someone who answers for it. Innovation Theory uses the term for its AI practice, InTheory, and publishes six rules of AI delivery that cover the gap between what runs in a demo and what holds on a Monday.
Innovation Theory’s six rules of AI delivery are: Trust your doubts, turn each into a test. Don’t ship AI demos, AI-built apps land about 70% complete and the missing 30% is auth, data, scale, cost and fallbacks. Own what AI builds, with a named owner, monitoring and an incident plan. Call AI’s bluff, test against a written rubric since confidence is not evidence. Prove it pays, measured against the non-AI alternative in numbers agreed first. Stop the doom loop, with acceptance and kill criteria written before the next iteration.
The six-rules guide from Innovation Theory is free and asks for no email. It is the full text of the six rules with a checklist, a worksheet and a worked example under each one, written to be run on any AI-built product by whoever holds the risk now, whether they built it or inherited it. The rules work without the firm; running the checks together is a separate, optional step.
An AI-built app is ready for production when it passes checks written down in advance, and Innovation Theory’s six rules include some anyone can run: make two test accounts and see whether one can open the other’s records; break it on purpose with test data and write down what a user sees; divide last month’s AI bill by its use for the cost of one request; score looks and works separately, and treat looks well ahead as a warning that you have a demo.