Most AI readiness roadmaps are slide decks. Twelve months of workstreams, a maturity model with five tidy stages, a center of excellence nobody staffs. Six months later there's no shipped system, just a nicer deck. If your roadmap can't point at a working thing a real person used this quarter, it isn't a roadmap. It's a stall.
A real roadmap is short, cheap, and ends in a decision. Here's the 90-day version we run: assess where you actually stand, pick one or two opportunities, write the eval before you build, pilot with a real user, measure cost and outcome, then decide scale or stop. Ninety days, one shipped pilot, one honest answer.
Days 1–15: assess the posture, not the ambition
Readiness is four dimensions. Score each one bluntly, because the weakest one caps everything downstream.
- Data access. Can you get the data an agent needs, in under a week, without a favor from someone in another department? If the answer lives in a PDF archive, a legacy ERP nobody has credentials for, or a spreadsheet on one person's laptop, that's not a data problem you solve during the build. Solve it now or pick a different opportunity.
- Security posture. Where does data go when it leaves your walls? Who approved that? If you can't answer what happens when an LLM sees a customer record, you're not ready to send it one. Know your data classification, your vendor terms, and your PII boundaries before line one of code.
- Team. One person needs to own the outcome and have time to validate output. Not a steering committee. Someone who can say "that's right" or "that's wrong" fifty times during a pilot.
- Process maturity. Is the process you want to automate actually stable and documented, or is it tribal knowledge with three undocumented exceptions? You can't automate a process nobody can describe.
You don't need all four at a 5. You need to know your real number on each, because that's what decides what you can build in the next 75 days.
Days 15–30: pick one or two, and write it down
Resist the portfolio. The first roadmap ships one pilot, maybe two — not a program. Pick opportunities that are high volume, high error cost, stable, owned, and measurable. The boring operational ones: the triage queue, the copy-paste tax between two systems, the report someone assembles by hand every month.
Write a one-page scope per pick. What goes in, what comes out, who the user is, what "correct" means, and the single number you'll move. If you can't fill that page, you don't understand the opportunity well enough to build it yet. That's a finding, not a failure — better now than in month three.
Days 30–45: define evaluation before you build
This is where most roadmaps skip a step and pay for it later. Before anyone writes a prompt, define how you'll know the system works. What does a good output look like? What's a failure you cannot ship? Assemble 30 to 50 real examples with known-correct answers and turn them into a test you can run on demand.
If we can't write a defensible eval, we say so and stop here — that's the cheapest money you'll spend. A process too fuzzy to grade is a process too fuzzy to automate, and finding that out on day 40 costs you nothing. Finding it out after you've shipped costs you trust.
Days 45–75: run a scoped pilot with a real user
Now build — narrowly. One workflow, one user, real data, running against the eval you already wrote. Not a demo for the exec team. A person doing their actual job, with the system in the loop, for two or three weeks.
Keep the human in control. The agent drafts, classifies, or pulls; the person approves. You're measuring two things at once: does the output pass the eval, and does the real user actually adopt it or quietly route around it. Demos lie about both. A real user on day twelve tells you the truth — where it breaks, which edge cases you never imagined, whether it saves time or just moves the work.
Days 75–90: measure cost and outcome, then decide
Put two numbers side by side. The outcome — hours saved, error rate down, turnaround faster, whatever you named in the scope. And the cost — model spend per run, plus the engineering to build and maintain it. Not projected. Measured, from the pilot.
Then make the call out loud: scale, iterate, or stop. Scaling a pilot that cleared its eval and got adopted is easy money. Killing one that didn't is discipline — and it's cheap, because you spent 90 days and one pilot, not a year and a program. A roadmap that can't say "stop" isn't managing risk. It's hiding it.
The anti-pattern to name
The failure mode isn't moving too slow. It's the strategy deck with no shipped system — endless assessment, maturity curves, and roadmaps that never touch production. Readiness isn't a document you achieve. It's proven by one working pilot, measured honestly, that earned the right to become two.
That 90-day loop — assess, scope, eval, pilot, decide — is exactly what our AI Readiness Assessment runs for you. We come out the other side with a defensible roadmap, a shipped pilot, and the two numbers that tell you whether to scale. If you've been circling a strategy deck and want a working system instead, that's the assessment to book.