The 90-day AI implementation roadmap I run inside client firms
This is the exact 90-day AI implementation roadmap I run inside client firms, published in full: two weeks of diagnosis, three production systems shipped by Day 90, counsel throughout – and how to run the same plan yourself if you have the engineering capacity in-house.
When a founder asks me what actually happens across the ninety days, I send them this page. There is no abridged version and no gated version; this is the whole method. Keeping it secret has never won me an engagement, because the plan was never the hard part. The hard part is shipping against it, week after week, inside a firm where everyone already has a full-time job. The buyer pays for the execution, not the secret. If you are still weighing whether the role itself makes sense, start with what a Fractional AI Officer does; this piece assumes you are past that and want the plan.
Why ninety days
Ninety days is not a marketing number. It is long enough to ship three production systems and, more importantly, to change the habits around them – the part most AI projects skip, which is why most AI projects end as demos. And it is short enough that accountability survives. A six-month engagement drifts; there is always another quarter to fix things in. A thirty-day sprint produces a prototype nobody trusts with client work. Ninety days is the shortest window in which a firm can watch a system go live, learn to trust it, route real work through it, and then repeat that twice more.
The roadmap at a glance
Three phases. The first two run in sequence; the third runs alongside both.
Stakeholder interviews, a full workflow audit, and an ROI-ranked opportunity matrix, ending in a signed-off ninety-day plan the firm keeps either way.
Three production systems, roughly one every thirty days, the first live by Day 30. Each ships instrumented, documented, and owned by the client team.
A weekly leadership call, async access between calls, and guidance on AI vendors and data policy as those decisions come up.
Days 1–14: diagnosis, not building
Nothing gets built in the first two weeks, on purpose. Most firms carry ten candidate problems and have the budget and attention for three; the expensive mistake is not building badly, it is building the wrong thing well. So the first fortnight is spent interviewing everyone who touches the work, auditing how it actually flows – not how the org chart says it flows – and ranking every opportunity by return: hours consumed, revenue at stake, and how often the founder is the bottleneck in the middle of it.
The phase ends with a signed-off ninety-day plan. The firm keeps that plan whether or not we continue; it is a complete piece of work on its own. This is also why I offer the diagnostic standalone: $5,000 for the two weeks, credited in full toward the engagement if the firm goes ahead.
Days 15–90: build sprints
Then the building starts, and the cadence is deliberate: roughly one production system every thirty days, three by Day 90. The first system is live by Day 30. That deadline matters more than any other in the engagement, because trust in this kind of work is earned by a system quietly doing its job on a Tuesday, not by a roadmap presentation. Once the team has watched the first one hold, the second and third meet far less resistance.
Every system ships with the same six things, without exception:
- Monitoring – every call traced, every error logged, visible on a dashboard.
- An evaluation suite – so a prompt change can never silently degrade quality.
- A founder-controlled kill switch – one toggle, no tickets, no waiting on me.
- A one-page runbook – what it does, who owns it, what to do if.
- Full IP transfer – the code, the prompts, the credentials are the firm’s.
- A trained team – the people who will run it once I step back.
This list is the difference between AI that demos well and AI systems that stay in production. It is also why each system carries a thirty-day stability commitment after handover: the ownership transfer is real, not theoretical.
A 90-day AI plan nobody ships against is a strategy deck with dates on it.”
Days 1–90: embedded counsel
The third phase runs underneath the other two for the full ninety days. A weekly leadership call keeps strategy and delivery in the same head instead of handed between a strategist and a builder. Async access covers the questions that cannot wait for Tuesday – “should we sign with this AI vendor?”, “is this safe with client data?” – and the vendor and data-policy guidance means those decisions stop being made in isolation on a Saturday night.
Running the roadmap yourself
If you have real engineering capacity in-house, you can run this AI implementation roadmap without me, and I would rather you did that than hire the wrong outside help. Three rules carry most of the value. First, pick one workflow – the one consuming the most senior hours – and resist the committee’s urge to start four. Second, ship it to production and let it hold for a few weeks before starting the second; a system in production teaches you more than three in development ever will. Third, instrument everything from day one: monitoring, evaluations, a kill switch, a runbook. The honest caveat is that the constraint is rarely knowledge. It is senior engineering time, and the discipline to keep shipping when client work pulls everyone back.
What Day 90 looks like
By Day 90, most engagements land in the same range: roughly twenty hours a week reclaimed across the senior team, about one FTE of capacity unlocked without a hire, and a 15–25% margin lift from the same revenue base. The quieter change is structural – the founder is no longer in the middle of every decision, because the operational layer absorbs the repetitive work. The first documented engagement, Carol Coelho Náutica, is written up end to end with the numbers the firm reported.
The commercial terms are equally public: $30,000 fixed for the Founders’ Edition – the first three firms, with continuation at $7,000 a month – or $50,000 at Standard rates, paid 50/50 across the engagement. The full breakdown, alternatives included, is in what a Fractional AI Officer costs.
Common questions
How long does AI implementation take for a small firm?
For a firm of five to thirty people, plan on about ninety days: two weeks of diagnosis, then roughly one production system shipped every thirty days, three in total. Shorter timelines usually produce demos rather than production systems; longer ones drift.
What happens in the first two weeks?
Diagnosis, not building: stakeholder interviews with everyone who touches the work, a full workflow audit, and an ROI-ranked opportunity matrix, ending in a signed-off ninety-day plan the firm keeps whether or not the engagement continues.
Can I run the roadmap without outside help?
Yes, if you have engineering capacity in-house. Pick one workflow, ship it to production before starting the second, and instrument everything: monitoring, an evaluation suite, and a kill switch. The method is public; the usual constraint is senior engineering time and the discipline to finish.