AI deployment runs in three phases. Audit: map one repeated workflow and record a baseline. Pilot: run the improved version on real work with the people who own it, with a test set and a reviewer. Rollout: widen only after the pilot beats the baseline, with named owners, training and a way to report problems. Each phase has an exit test.
What are the three phases of AI deployment?
Each phase answers a different question and ends with a test.
| Phase | Question it answers | What it produces | Exit test |
|---|---|---|---|
| Audit | Which work is worth changing, and how? | A map of one workflow, a baseline, a marked list of steps | You can name the input, rule, output, reviewer and failure mode of every step you plan to change |
| Pilot | Does the new version beat the old one on real work? | A working prompt, Skill, automation or agent, plus results against the baseline | The owner would object if you took it away |
| Rollout | Can the rest of the team run it without the pilot group? | Named owners, training, a runbook, a feedback route | It keeps working a month later with nobody from the pilot watching |
altr calls these Map, Prove and Deploy; see how we altr work.
Phase 1: how do you audit the work and choose the pilot?
Pick one workflow that repeats weekly, has a single owner and produces something checkable. Walk a real recent instance with the person who does it, write down every step, and time a few instances for a baseline.
It ends with a short written choice: the workflow, its owner, the steps to change, the data involved, who reviews the output and the one metric that shows progress. The worksheet is in how to run an AI workflow audit yourself.
Common failure: auditing five workflows at once. You end up with a list and no pilot.
Phase 2: how do you run a pilot that proves something?
Keep it narrow and run it on real work, not a demo. The pilot is the smallest version that could replace the current way: a shared Project and a few prompts, a single automation, or one agent step.
- Build a test set before you start. Ten to twenty real past instances: normal cases, awkward ones, and at least one where the right answer is "a person should handle this".
- Name a reviewer who checks outputs during the pilot and records what they changed.
- Run it on live work for two to four weeks alongside or instead of the old way.
- Compare against the baseline: time per instance, how often output was sent back, errors caught in review.
- Fix the instructions, not the people. Each rejected output is a change to the prompt, the Skill or the rule.
Common failure: judging the pilot on how impressive it looked rather than on the baseline.
Phase 3: how do you roll out without losing control?
Widen only what the pilot proved, to the people who do that work, with the same review rule.
- Three named roles: an operating owner accountable for use, a system owner who makes changes, and a reviewer who samples output.
- Training on the team's own examples. See AI training for employees.
- A one-page runbook: what it does, where it runs, what to check when it fails, who to tell.
- A policy: approved accounts, data rules and review points. See AI governance for small business.
- Metrics leadership can act on: use on the target task, cycle time, rework rate, errors caught, and the owner's verdict.
For Claude, see how to roll out Claude to a team.
Common failure: rolling out company-wide on the strength of one enthusiastic pilot user.
How long does each phase take?
It depends on the workflow, not a calendar. A single-owner process in one system can go from audit to rollout in weeks. A process that crosses teams or needs a custom connection takes longer, mostly in the audit and security review. Be wary of any plan that fixes dates before the work is mapped. After the first workflow holds for a month, start the next audit; the second cycle is faster.
Common questions
What is an AI deployment roadmap?
A plan for moving AI from interest to daily use in three phases: audit one workflow and record a baseline, pilot an improved version on real work, then roll out what the pilot proved with owners, training and review rules.
How many workflows should an AI pilot cover?
One. A pilot is meant to prove that a specific change beats the current way on real work. Several workflows at once split attention, blur the measurement and make it hard to say what worked.
Why do AI projects fail after the pilot?
Usually because the rollout skipped something the pilot had: a named owner, a reviewer, training on real examples or a clear data rule. Scaling a pilot that was never measured against the old way is the other common cause.