How to roll out agentic AI delivery: a 90-day plan

In this lesson you'll learn:

• the five phases of a first rollout, and the trap or signal in each;
• why the first feature is chosen for provability, with the arithmetic that proves it;
• the six objections you will hear, and what answers each.

Sound familiar?#

• The first agentic project was the flagship, and a year later it is still a pilot.
• A demo convinced everyone in week three, and nothing has reached production since.
• The people doing the work today are described as resistant, and nobody has asked them anything.

Each is a sequencing mistake, and none needs a reorganisation to fix.

What does a first rollout have to achieve?#

One feature proven end to end, whose artefacts the second feature can copy. Not a portfolio of pilots, and not tool adoption. The first feature buys the organisation a spec template, a bar derivation, an authority budget, a harness in CI and a rehearsed rollback, and the evidence that the method works here. The second feature should then need less attention than the first.

Roll it out, step by step#

Step 1 · Days 1 to 15: pick the wrong-looking thing#

Choose one feature for provability rather than value: high volume, low damage per mistake, and an existing human process to compare against. Run the three AI-fit questions across the top ten candidates and publish the ones that came back as rules. Take the baseline now: person-days per story, before anything changes. The trap: starting with the flagship, which has the highest bar, the least tolerance for a first attempt and the most spectators. Is it AI work?

Step 2 · Days 15 to 30: write the artefacts nobody wants to write#

The eight-field spec, the acceptance bar per slice, the authority budget and a context file in the repository. Five of the eight spec fields will be undecided, and those five are the value of the exercise. The trap: a demo exists by now and it is persuasive. A demo is the easy 20%; it cannot tell you how often the system is wrong on the cases you did not choose. P1 · Design and Spec

Step 3 · Days 30 to 60: build in slices, prove in CI#

A shippable slice a day, each proven before the next: a walking skeleton with no model in it first, caps moved out of prompts into tool signatures, the golden set at fifty real cases running in CI, review routed by risk band. The signal: something merged most days. If the demo is still the only evidence at day 45, the slices are not slices. P2 · Build and Prove

Step 4 · Days 60 to 90: prove it beside the humans#

Run it in shadow next to the people doing the work, deciding but never acting, for a window agreed in advance and read daily per slice. Keep money actions gated whatever the shadow shows. Rehearse the rollback before the cut-over, then cut over at 5% and widen on live evidence rather than on a date. If it does not match, you learned that for free. Shadow and cut-over

Step 5 · Day 90 onward: report honestly, and let production set the agenda#

Two numbers from the first cycle: the saving and the spend. A surprise bill goes back into the design, an incident produces a control rather than a name, drift is watched weekly, and the maturity check is re-run each quarter. The signal that it took: the second feature needs less of your attention than the first, because the artefacts now exist to copy. The maturity model

Step 6 · Meet the resistance with evidence, not persuasion#

The move that does most of the change management is unglamorous: credit every requirement to the person who raised it, in writing, before consolidating any of them. A voice that felt dropped in week one comes back in week five as a constraint.

When to use it#

• For an organisation's first agentic feature, or its first since a failed pilot.
• When a portfolio of pilots has stalled: pick one, and run it through all five phases.
• When a new business unit adopts the method, with the first unit's artefacts as templates.

Why it matters#

A first rollout is decided in the two places teams most want to rush: the fortnight of artefacts, which turns a demo into something that can be proven, and the shadow run, which is the only way to learn whether the system agrees with the people doing the work. The plan spends its effort there, and the first feature's artefacts are what make the second one cheaper.

Key takeaways#

1. Pick the first feature for provability, and take the baseline before anything changes.
2. Spend the effort where the outcome is decided: the artefacts fortnight, and the shadow run.
3. Meet resistance with evidence, and credit every requirement to the person who raised it.

FAQ#

How do you roll out AI agents in an organisation?#

One feature at a time: choose one that can be proven quickly, write its spec, bars and authority budget, build it in daily slices proven in CI, run it in shadow beside the people doing the work, cut over gradually, and report the saving and the cost together. Then copy its artefacts to the next feature.

How long does it take to get an AI agent into production?#

About ninety days for a first feature run this way, of which roughly a month is shadow and cut-over. Later features are faster, because the spec template, harness and rollback already exist.

Which AI use case should we start with?#

A high-volume, low-damage task with an existing human process to compare against. The most valuable use case usually has the highest bar and needs the most evidence, so it makes a poor first project.

Why do AI pilots fail to reach production?#

Two causes recur: the first feature was chosen for value and cannot clear its bar inside a quarter, or a persuasive demo was treated as evidence and the specification, bars and tests were never written.

Apply it in your role#

Across the enterprise. Roll out one feature and one business unit at a time. The second unit starts from the first unit's artefacts, and that reuse is where the enterprise-wide saving actually comes from.

The ten-minute workflow. A credited requirements register, straight from transcripts:

Here are transcripts of our discovery meetings: <paste>. List every requirement stated, credited to the
person who said it, with the quote. Then group duplicates while keeping every credit, and list conflicts
between people as open questions. Do not merge or drop anything silently.

Sources and credits#