The 90-Day AI Roadmap for Companies Starting From Zero

The strategy that never ships

Someone asked you for an AI plan by this quarter — the board, the CEO, or the version of you that set the annual goals. It's this quarter now, and what you have is a slide that reads "Explore AI opportunities" above a green checkmark you don't believe.

This is the normal outcome, and it isn't a discipline problem. "Do an AI strategy" is an instruction with no edges. It doesn't say who, it doesn't say by when, it doesn't say what "done" looks like. So it becomes a committee, then a vendor demo, then a pilot three people know about and no one can point to. Six months in, the honest status is: we looked into it.

The failure isn't a lack of ambition. It's the shape of the goal. A plan that tries to cover everything commits to nothing you can start on Monday. Meanwhile the real leverage — the report your controller rebuilds by hand every month, the quoting logic that lives in one person's head — sits untouched because it was never grand enough to make the deck.

What works instead is boring on purpose: a sequence of narrow, visible wins on a 90-day clock, each with a name attached before it starts. Ninety days is long enough to ship something real and short enough that no one gets to hide inside it.

A plan nobody owns is just a document. Put a name and a ninety-day clock on it, and it becomes a project.

Weeks 1–2: Map the drains, then write one page

Before you automate anything, find out where the week actually goes. Not a process-mapping project — a fast inventory of where senior time and attention leak out. Where does your team rebuild the same thing every month? Where does a decision wait three days for a number someone has to assemble by hand? Where does work stall because only one person knows how?

Run this before you buy a single tool:

1. Pick your five most valuable people by judgment, not title.
2. Ask each, for thirty minutes: what did you do last week that a sharp assistant could have drafted or assembled for you?
3. Write the answer in hours per month, not adjectives.
4. Sort the list. The top three are your candidates. Everything else waits.

Then write the ground rules — one page. What AI is allowed to touch this quarter and what it isn't. What data never leaves the building. What always gets a human signature before it goes out the door. This is not a governance program; it's a single page that makes people safe to move quickly, because the boundaries are written down instead of argued about after something goes wrong.

The output of these two weeks is unglamorous and exactly right: a ranked list of drains measured in hours, and a one-page charter. Nothing has been "implemented." Good. You now know where to point, and you have permission to point there.

Weeks 3–6: Ship one win beside the current process

Pick one item off the top of that list and ship it. One. The best first target is almost always reporting or decision-prep: the monthly package, the board deck's data pull, the weekly ops review — the recurring thing that eats a skilled person's hours and produces a number other people wait on.

Reporting goes first for three reasons. It's visible, so a win gets noticed. It's measurable, so you can prove it. And the right answer already exists — last month's version — so you can check the machine against a known-good output before you trust it. High visibility, low blast radius.

The rule that keeps this safe: run it in parallel, not instead. For a few cycles, the new way and the old way run side by side. The old process stays the source of truth until the new one has matched it enough times that the person who owns the report stops double-checking out of habit. You're not betting the close on it. You're auditing it in the open.

By week six you want one thing you can point to: a report that used to take two days now drafted in an afternoon — checked and shipped by the same person who used to build it from scratch, freed up for the work only they can do. That single, provable result is the story that earns you the next forty-five days.

Weeks 7–12: Capture what one person knows, then rank the rest

Two moves in the back half of the quarter.

First, capture one piece of tribal knowledge. Every operation runs on a handful of people whose judgment isn't written down anywhere — the estimator who knows which jobs go sideways, the service lead who can feel a churning account from the tone of one email. Pick one such domain and get it out of that person's head and into something durable: the rules, the tells, the exceptions. This is slower and less flashy than the reporting win, which is precisely why it belongs here, after you have a visible success to stand on.

Second, rank everything else. By now you have a real list — the drains from weeks 1–2, the extensions your first win suggested, the ideas that surfaced while your team worked. You will not run them all. Score them, and let the score set the order. Give each candidate a 1-to-5 on five dimensions:

• Value — what it's worth if it works.
• Readiness — is the process stable and the data clean enough to build on?
• Risk — what breaks if it's wrong, and how loudly?
• Time-to-value — how fast it shows a result.
• Reusability — does it build capability the next initiative inherits, or is it a one-off?

Add them up, inverting risk so lower risk scores higher. Now the order is defensible. When a CFO asks "why this one first?", the answer is "here's the scoring," which wins that conversation every time against "instinct."

What you hold at day ninety is a ranked portfolio: a short list of funded, owned, running wins, plus a watchlist of bigger bets you're deliberately not starting yet — each tagged with the trigger that would move it up. That is a strategy. It just happens to be built from things you already shipped instead of things you hope to.

The rule that holds all of it together

One discipline makes or breaks this whole arc: no initiative starts without a named owner. Not a committee, not "the ops team" — a person, first and last name, who owns the outcome and whose other work you've agreed to move to make room for it.

The named-owner rule does three quiet, useful things. It caps how much you can run at once, which is the entire point — you cannot name twelve owners for twelve initiatives and mean it, so the constraint forces the sequencing that "do an AI strategy" refused to do. It kills the ideas nobody actually wants, because if you can't find one person willing to put their name on something, that's not a scheduling problem, that's your answer. And it's the honest test for whether a win can scale later: a pilot that works on one person's heroics hasn't proven anything until you ask who owns the process, who owns adoption, and who owns the risk when it's running across the whole company. Better to ask that on day one than after the demo.

Start narrow, start visible

The pull to boil the ocean is strong because it feels like leadership — the platform, the reorg, the "AI-first" mandate. But boiling the ocean is the most comfortable way to avoid shipping anything, because nothing that large can be checked against reality inside a quarter. It buys time. It doesn't buy proof.

Do the opposite. Start on the one drain that costs you the most and embarrasses you the least to talk about. Ship it beside the old way. Name the owner before you begin. In ninety days you'll have traded "we should do something with AI" for three things people can actually see — and that is worth more than any strategy nobody owns.

If you'd rather not assemble the ranked version by hand, the Opportunity Portfolio you produce inside The Operational Intelligence Method scores every candidate on those five dimensions and sequences them into exactly this kind of 90-day roadmap, with an owner named for each line.