A twelve-month AI enablement roadmap for a company of 50 to 500 people holds four things: one baseline you can re-measure, three to five role cohorts trained on their own work, a monthly cadence that outlives the training, and two gates where stopping is honest. That is the plan. Everything else published under the name belongs to a different company.
Search the phrase and you get year-one plans built around a model gateway, a vector store, an evaluation harness, a center of excellence and a second wave of technical hires, costed in the millions. If you have already bought ChatGPT or Microsoft Copilot seats and the question landed on you as a second job, nine tenths of that document does not apply and the remaining tenth has no date attached.
What a twelve-month roadmap is for
The roadmap closes the distance between a company that has adopted AI and a workforce that can use it. Those are separate events on separate clocks, and US government data measures both.
The Census Bureau’s Business Trends and Outlook Survey found 37% of firms with at least 250 employees, and 32% of firms with 100 to 249 employees, reported using AI in their business operations, against under 20% of firms with four or fewer employees (Grundy, Breaux and Khatiwoda, US Census Bureau, data collected December 14, 2025 to May 3, 2026). The Bureau widened the question in November 2025, so readings either side of that date are not strictly comparable.
Weight the same question by headcount and the picture inverts. A Federal Reserve FEDS Note published April 3, 2026 reports that about 18% of firms had adopted AI by year-end 2025, while 78% of the labor force works at a firm that has adopted it and about 54% works at a firm using large language models. Worker-reported use of generative AI for work sits near 41%.
Read those together. Your company bought in; fewer than half the people report using the tools for work. The purchase took a quarter and the capability takes a year, which is the distinction our main guide to AI enablement draws between a rollout and a program. The roadmap admits that gap and schedules it.
The year at a glance
| Quarter | What happens | Who owns it | Output you can show |
|---|---|---|---|
| Q1, months 1–3 | Baseline, a written use paragraph, first cohort trained | Owner plus vendor | Baseline scores, one working habit in one team |
| Q2, months 4–6 | Remaining cohorts, cadence starts, first stop gate | Owner | Two to four roles trained, re-measured baseline, a decision |
| Q3, months 7–9 | Second-order requests surface, prompt library consolidates | Owner plus cohort leads | Automation candidates ranked by hours |
| Q4, months 10–12 | Proof pack, policy, renewal decision | Owner plus finance | A board-ready result and a year-two choice |
Three details carry the plan. One owner runs the year, not a committee. Training ends in month six while the program runs to month twelve. Two of the four quarters end in a decision, not a deliverable.
Months one to three: baseline, a paragraph of policy, one cohort
Week one is a baseline, not a kickoff. Before anyone is trained, measure what the team can do unaided: give each role a timed task it performs weekly, score the output, and keep the artifacts. Without that file, every number you produce in month eleven is an opinion. License dashboards are not a baseline; seat activity tells you a tab was open.
Week two is one paragraph on acceptable use. Not a policy. A paragraph naming which tools are approved, what information stays out of them, and who to ask when the answer is unclear. A missing policy stalls more programs than a missing budget, and the paragraph removes that dependency. The full document belongs in Q4, once you have watched what people do.
Weeks three to eight are the first cohort. Pick the role with the most repetitive written output and the least political exposure: operations, finance back office, or support. One cohort of 12 to 20 people, trained on their own documents, in sessions under three hours. Hands-on LLM fluency applied to real intake forms, client emails and reports is the format that changes anything by week nine.
Weeks nine to twelve are the habit test. The cohort either reaches for the tool unprompted or does not. Watch, do not intervene. Uneven adoption inside one cohort usually points at a manager’s calendar, not the curriculum.
Q1 contains no tool evaluation, no vendor migration, no dashboard build and no second cohort. One cohort done properly is what the rest of the year compounds on.
Months four to six: the remaining cohorts and the first gate
Q2 trains two to four more roles and then stops training.
Sequence them by evidence from Q1, not the org chart: the role whose Q1 neighbors asked the most questions goes next. Sales and marketing usually volunteer first and should rarely go second, because their output is already fast and the gain is harder to read.
Start the monthly cadence in month four and protect it for the rest of the year. Ninety minutes a month, one per cohort: people bring a task that went badly, someone solves it in the room, the solution goes into a shared library. Course completion is easy to report and tells you little. The monthly session is where sustained use shows up, and it costs almost nothing once it is on the calendar. The cadence, not the training, is what a second year gets built on.
The first gate sits at the end of month four. Re-run the Q1 baseline tasks on the first cohort. Three outcomes, all legitimate:
- Scores moved and use is unprompted. Continue, and add the next cohort.
- Scores moved, use is prompted. The curriculum worked and the workflow did not. Fix a manager’s expectations before training anyone else.
- Scores did not move. Stop. Do not train cohort three to average away the result. Something about scope, roles or access is wrong, and two more cohorts will cost money without telling you which.
A roadmap with no stopping point is a procurement document. Most published ones have none.
Months seven to nine: the part you cannot schedule
Months one to six can be a calendar. Months seven to twelve are a hypothesis, and anyone handing you a dated week-32 deliverable in January is guessing.
What reliably happens is a change in the questions. Teams stop asking how to write a prompt and start asking whether the weekly report can produce itself. That shift is the signal the training worked, and handling it is Q3’s job:
- Collect the asks in one list, one line each, owned by the enablement owner.
- Rank by hours, not enthusiasm. Hours per week times people affected. The top three are usually unglamorous.
- Price nothing yet. Custom automation runs $25,000 to $150,000 in the current market, and Q3 is for sizing the question.
- Consolidate the library. The prompts that survived three monthly sessions are the ones you hand new hires.
Q3 is also where sponsor attention lapses: novelty gone, training over, nothing in a board pack yet. Two defenses work: a one-page monthly note carrying cadence attendance and one named example, and a standing hold you do not cancel on a quiet month.
Months ten to twelve: proof, and the second gate
Q4 produces the three artifacts that decide year two.
A re-measured baseline. The same tasks, the same scoring, every trained role. Change against a recorded starting point is the only capability claim finance accepts, which is the whole argument for measuring in week one.
A written policy. Cheap now, because you have ten months of observed behavior instead of imagined risk. The NIST AI Risk Management Framework organizes this work as govern, map, measure and manage, and a short internal policy mapped to those four headings holds up in a customer security review.
A ranked automation shortlist. Three candidates, each with the weekly hours it would return and the team that would own it.
The second gate sits in month eleven, before the next budget cycle closes. Four options, and three are not “renew as is”: stop and keep the cadence internally, retain a lighter advisory arrangement, convert the budget into the top automation candidate, or renew for the roles you have not covered. Our breakdown of what a second year of enablement looks like shows how those four diverge, and it is better read before you write the month-eleven paper than after.
Month twelve closes with a handover: baseline files, session materials, the prompt library, the cadence calendar, and the policy. A vendor who cannot hand all five over rented you capability instead of selling it, which is the test our guide to end-of-engagement deliverables applies in detail.
What the first twelve months should deliberately exclude
Published roadmaps rarely carry an exclusions list. For a company with one owner and a fixed budget, the refusals are most of the plan.
| Excluded from year one | Why | When it earns a place |
|---|---|---|
| A center of excellence | A standing body needs people to govern. You have one owner | Year two, if headcount supports it |
| Model fine-tuning | Nothing in year one is bottlenecked on model behavior | When a use case proves prompting is the limit |
| An agent platform | Agents need documented processes. Q1 usually proves yours are not | After a process has been redesigned and held a quarter |
| A tool-stack migration | Switching tools mid-program resets the habits you built | At renewal, with usage data in hand |
| Company-wide rollout in Q1 | Training 300 people before one cohort holds a habit buys 300 opinions | Q2, cohort by cohort |
| A custom internal course | Six months of build for content that ages in three | Never, at this size |
Every row is a reasonable year-two item and a fatal year-one one, because each consumes the owner’s attention. That, not budget, is the binding constraint.
What AI enablement costs, quarter by quarter
Market ranges, not list prices, shaped by quarter rather than averaged:
| Quarter | Typical spend shape | What drives it |
|---|---|---|
| Q1 | Largest single block | Baseline design plus a first cohort built from your documents |
| Q2 | Proportional to cohorts | Two to four cohorts, declining per-cohort as material is reused |
| Q3 | Lightest | Cadence facilitation and advisory hours |
| Q4 | Small | Re-measurement, policy drafting, handover |
Corporate LLM fluency workshops run roughly $2,000 to $15,000 per engagement block in the current market, with role count driving price more than headcount. A three-to-five-cohort year at a 150-person company lands in five figures, not six. Our cost breakdown for AI enablement programs sets out the line items; for roadmap purposes the thing to flag with finance in month one is that spend is front-loaded while proof arrives last.
Against the license bill it is usually the smaller number. Teams that compare what their seats cost with what capability costs find the enablement line is a minority of annual AI spend carrying most of the return.
Where twelve-month roadmaps come apart
Four failure modes, in the order we see them:
- No baseline. The most expensive omission in the year, and free to avoid in week one. Without it, month eleven is a sentiment survey.
- A committee instead of an owner. Decision rights split across operations, IT, and HR means renegotiating every cohort. One owner with two hours a week beats three sponsors with authority.
- Training treated as the program. Training is months one to six; the cadence is months four to twelve. Programs that end when the sessions end regress, which is the argument in our case against treating licenses as capability.
- A twelve-month plan presented at uniform confidence. Publish months one to six as a schedule and months seven to twelve as a shape, with the gates marked. Delivery, behavior, and proof run on different clocks, as our realistic program timeline sets out, and a plan that pretends otherwise loses credibility in month seven.
Frequently asked questions
How long should an AI enablement roadmap be?
Twelve months, with the first six scheduled and the last six directional. Eighteen and twenty-four month plans assume a planning function you do not have. A year is long enough to produce a measured result and short enough that one owner can hold it in their head.
What does the first ninety days actually contain?
A baseline measurement in week one, a one-paragraph use statement in week two, one cohort trained on its own documents in weeks three to eight, and an unassisted habit test in weeks nine to twelve. One cohort, not a rollout.
How much should a company of 50 to 500 people budget for year one?
Five figures for a three-to-five-cohort year, front-loaded into the first two quarters. Workshop blocks run roughly $2,000 to $15,000 depending on role count and customization. Budget for automation separately and later; it starts near $25,000.
Who owns the roadmap when there is no Chief AI Officer?
One named person with two hours a week and the authority to put people in a room. Operations and chief-of-staff roles work because they already cross functions. The failure pattern is a steering committee with three co-owners, which produces renegotiation rather than decisions.
What should a twelve-month roadmap leave out?
A center of excellence, model fine-tuning, an agent platform, a tool migration, a custom internal course, and any company-wide rollout before one cohort has held a habit. Each consumes the owner’s attention, which runs out well before the budget does.
Can the timeline be compressed?
The training can. Three cohorts fit inside eight weeks if document access and calendars cooperate. Habit formation cannot be compressed, which is why a six-month roadmap can report delivery and not capability. Compress Q1 and Q2 if you must; do not move the measurement forward.
What if we already started without a roadmap?
Baseline now. A late baseline is weaker, because you cannot separate what training did from what unguided use did, but it beats none. Then treat the current month as month four and run the first gate.
How do we measure progress quarter to quarter?
Q1 on output quality against the baseline tasks. Q2 on unprompted use, cohort by cohort. Q3 on the automation shortlist and the hours attached. Q4 on re-measured scores plus one named process that changed. Seat activity measures none of these.
What happens if the sponsor loses interest in month seven?
Assume they will. A one-page monthly note carrying cadence attendance and one named example, plus a standing hold you never cancel, carries the program through the quarter where novelty has worn off and proof has not landed.
What does month twelve look like if it worked?
Trained roles score measurably higher on the same tasks they were timed on in week one. The monthly cadence runs without the vendor in the room. A policy exists that reflects observed behavior. A ranked shortlist of automation candidates sits with finance, hours attached. That package makes the year-two conversation a choice rather than a renewal.
Key takeaways
- A twelve-month AI enablement roadmap at a 50-to-500-person company contains a baseline, three to five cohorts, a monthly cadence, and two stop gates. The published alternatives assume a platform team and a multi-million budget.
- The gap it closes is measurable: 78% of the US labor force works at a firm that has adopted AI, about 54% at a firm using large language models, and roughly 41% of workers report using generative AI for work.
- Measure in week one. A baseline costs nothing then and cannot be reconstructed in month ten.
- Train in months one to six and run the cadence from month four to month twelve. Programs that end when the sessions end regress.
- Mark gates at month four and month eleven, treat stopping as legitimate at both, and write the exclusions down. Attention, not budget, runs out first.
If you want a second read on which quarter you are in, we offer a free AI-readiness call: thirty minutes, no deck. Book a call and bring your license bill and your last training agenda.
Arthur Wandzel
