Most AI delivery models fail at the seams. Strategy is done by one firm, the build is handed to another, and operations belong to whoever is left holding it — so context is lost twice and accountability is lost entirely. Audit → Build → Operate is one partner across all three, and each phase ends in something that runs rather than something that reads.
That second half is the load-bearing part. Any consultancy can name three phases; the question is what each one produces. A phase that ends in a document has, in the most literal sense, produced nothing that works.
Audit — find the work worth automating
Ends with: a working pilot and a costed build plan.
The purpose of the audit phase is to answer *which workflow, what will it cost, and will it pay* with enough confidence to commit real money. Not to produce an opportunity landscape. We map the workflow, score the ROI, and build a pilot of the thing we're proposing — which is the only way to find out what a build actually involves.
It's deliberately paid and short. Free discovery isn't free; it's priced into whatever comes next, and it manufactures switching cost rather than clarity — we wrote up that mechanism in Bait-and-Bill. Ours is the AI Transformation Sprint: $25,000, two weeks, ending in a deployed pilot, and credited toward the build it scopes.
Skip this phase and you build the workflow someone nominated in a meeting rather than the one costing you money. This is where most of the 95% of pilots that deliver no measurable P&L impact are decided (MIT Project NANDA, 2025) — not in engineering, but in the choice of what to engineer.
Build — ship the production system in your stack
Ends with: a production system, the eval suite, the docs, and the code — yours.
An embedded pod builds inside your systems, not alongside them: two-week sprints, a working demo every sprint, evals before deploy, observability from day one, guardrails and fallbacks. The output is either one or more production agents at a flat fee per agent, or an embedded team when the work is broader than a single system.
Note what's in the deliverable list: the eval suite and the code, in your repository. That's not generosity, it's the difference between buying a capability and renting one — and it's the specific thing whose absence causes Vendor Roulette, where each new vendor restarts from a blank page because the last one left nothing behind.
Skip the discipline inside this phase — evals, traces, guardrails, a rollback path — and you get a demo that impresses a room and can't be trusted with production. Those are the five gates, and they're cheap now and expensive to retrofit.
Operate — stay on the hook
Ends with: a system that keeps working, and keeps improving.
Monitoring, eval regressions, prompt versioning, drift management, vendor-change response, and a quarterly roadmap — increasingly tied to your KPIs rather than to uptime. Managed AI Operations runs $3,000–$20,000 a month depending on how many systems are under management.
This is the phase most models omit and most buyers underfund, because it's the only one that doesn't feel like progress. But AI systems decay rather than crash — models get deprecated, data drifts, someone changes the workflow — so an unoperated build is a depreciating asset. We go through the failure modes in why AI systems break after launch.
Skip this phase and you will rebuild in eighteen months, having spent the interim quietly getting worse outcomes than you think you're getting.
Why the order can't be rearranged
| Phase | Answers | Skip it and |
|---|---|---|
| Audit | Which workflow, and will it pay? | You build the wrong thing well |
| Build | Does it work unattended? | You have a demo, not a system |
| Operate | Does it still work? | It decays and you rebuild |
Each phase produces the input the next one needs. Build without Audit has no defensible target. Operate without Build has nothing to operate. And Audit without a committed Build is the most common shape in the market — the strategy engagement that ends in a deck, which is why so many companies have a beautiful AI roadmap and nothing in production.
Where the growth side fits
One thing sits alongside rather than inside the sequence: Ascent, the growth program. Audit → Build → Operate makes your operation work; Ascent makes buyers able to find you in an answer-engine world. They're independent — plenty of clients run one without the other — but they share the same shape, which is a measured baseline, a build, and a monthly accountability to a number rather than to activity.
What makes this different from three phases on a consultancy's slide
Only one thing, and it's testable: every phase here ends in working software you own. Audit ends in a deployed pilot, not a findings deck. Build ends with the code and evals in your repo, not on a vendor platform. Operate is measured on your KPIs, not on tickets closed.
That's the whole model, and it's why we run it as a software factory rather than an agency — the same sequence, the same quality gates, priced the same way every time, so you learn it once and always know what phase you're in. The full version of the approach is on its own page, including our cadence and what we put first.
