Report 14 · AI adoption playbook
AI Upskilling 30/60/90
The week-by-week execution plan for AI adoption in an organization of 50 to 500 people: who owns what, the month-one policy, the month-two training block and prompt library, and the month-three workflow redesign, with pass or fail gates at day 30, 60 and 90.
The short version
Ninety days, three phases, one gate at the end of each month.The strategy case lives in the companion report, The AI-Ready Workforce. This is the calendar. Ninety days in three phases, legitimize in days 1–30, build capability in days 31–60, prove value in days 61–90, sized for an organization of 50 to 500 people with no dedicated AI task force.
The design constraints come straight from the research. Employees are already using AI further ahead than leadership estimates, five hours of role-specific training is where regular usage jumps, and inside the EU a documented literacy program has been a legal duty since February 2025.
Evidence: 1 McKinsey & Company (2025) / 2 Boston Consulting Group (2025) / 3 European Commission (2025)
Employees are this much more likely to run at least 30% of daily work on AI than leaders assume.
1 McKinsey & Company, 2025Training dose where regular AI use jumps, strongest when paired with in-person coaching.
2 Boston Consulting Group, 2025Employees who rank training the number one factor for AI adoption.
1 McKinsey & Company, 2025Checkpoints at day 30, 60 and 90, each with numbers rather than vibes.
4 Omie, 2026What's inside
The question is not whether to adopt AI. It is who owns the rollout, what they ship each month, and how you will know at day 30, 60 and 90 that it is working.
The owner table
Five hats, not five hires.A ninety-day rollout does not need a new department. It needs five hats worn by people you already employ, and it needs the order right. Permission comes first, then skills, then tools. Your people are already ahead of your estimates, so legitimize the usage before you license anything.
One person can wear two hats in a small org, but every hat needs a name against it. An unowned rollout is a rollout that quietly reverts to shadow usage.
Evidence: 1 McKinsey & Company (2025)
| Hat | Usually worn by | Owns |
|---|---|---|
| Program owner | HR or L&D lead | The whole roadmap, the checkpoint readouts, and the Article 4 evidence file. |
| Policy owner | Ops or a legal-adjacent leader | The one-page policy, tool approvals, data red lines, and the incident path. |
| Champion network | About one power user per 10 to 15 staff, self-selected | Office hours, prompt library curation, and the "it worked for me" proof. |
| Exec sponsor | CEO or department head | Visible weekly personal use, budget unblocking, and the amnesty message in their name. |
| Workflow owner (month 3) | Manager of the pilot process | The one redesigned workflow and its before-and-after instrumentation. |
The workflow owner joins in month three; the other four are named in week one.
Evidence: 3 European Commission (2025)
| Gate | When | What it checks (pass or fail) |
|---|---|---|
| Gate 1, legitimize | Day 30 | Policy shipped and findable in under a minute, at least 60% survey response, baseline complete for every role family, champions named. |
| Gate 2, capability | Day 60 | At least 70% of staff past five cumulative training hours, at least 50% using AI weekly on real tasks, three or more prompt-library entries per department. |
| Gate 3, value | Day 90 | The pilot workflow improves on at least one instrument, skill delta positive in every role family, evidence file current. |
Miss a gate and the readout says why. A slipped gate caught at day 30 costs a week; one caught at day 90 costs the program.
Evidence: 2 Boston Consulting Group (2025)
Name the four launch hats this week. The workflow owner can wait until month three, when there is a workflow to own.
Days 1–30, legitimize
Policy, amnesty, and a baseline you can compare against.Month one buys permission, not usage. You ship a policy people can actually find, you invite the shadow users into the light, and you take the skills baseline that day 90 will be measured against.
Expect the McKinsey pattern when the anonymous survey comes back: more usage, and more capable usage, than leadership believed. Inside the EU, the same month starts your Article 4 evidence file.
Evidence: 1 McKinsey & Company (2025) / 3 European Commission (2025)
Week 1, ship the one-page policy
Name three to five approved tools, draw the data red lines (client PII, credentials and unreleased financials never get pasted), mark the human sign-off points, and give people an incident path. One page. Nobody reads the eleven-page version, which makes it worse than none.
Week 2, amnesty and anonymous inventory
Send an exec-signed message: if you have been using unapproved tools, tell us what works, no consequences, we want your head start. Then run an anonymous survey covering which tools, which tasks, hours saved and biggest worry.
Weeks 3–4, baseline skills and pick champions
Run a per-role AI skill baseline (prompting, verification, tool fit) that takes about ten minutes per person and becomes the day-90 comparison. Recruit champions from the survey power users and give them the hat in public. Open the evidence file with the policy version, the survey summary and the baseline dates.
Gate 1 at day 30: policy findable in under a minute, at least 60% survey response, every role family baselined, champions named. Miss two of the four and fix them before training starts.
Days 31–60, capability
Get role by role past the five-hour line.Month two is where behavior changes, and only role-specific practice moves it. Generic webinars do not move usage; task-shaped reps do, and they move furthest when a champion is coaching in person. The one number worth being strict about is five cumulative hours per person.
Evidence: 2 Boston Consulting Group (2025)
Weeks 5–6, daily reps begin, role-specific from day one
Ten minutes a day inside the flow of work. The recruiter drafts and critiques a screening summary; the controller reconciles with AI and hunts its errors. Every rep ends with verification: find what the model got wrong before anything ships.
Weeks 7–8, coaching layer and prompt library
Champions run a weekly 25-minute office hour per department, the in-person coaching that carries the biggest usage lift. The prompt library goes live alongside it, curated by champions and open to everyone.
Evidence: 2 Boston Consulting Group (2025)
What a prompt-library entry needs
Gate 2 at day 60: at least 70% of staff past five cumulative training hours, at least 50% using AI weekly on real tasks, and three or more library entries per department. Guard the five-hour line above all.
Days 61–90, value
One workflow, rebuilt and instrumented.Month three turns capability into a receipt. You pick one workflow, rebuild it around what AI does well and what humans must still own, instrument it before and after, and re-run the baseline so the skill delta is undeniable.
Weeks 9–10, pick and rebuild one workflow
Selection rule: high volume, text-heavy, measurable, owned by a willing manager. Classic candidates are support ticket triage, proposal first drafts, QA summaries and candidate screening notes. Rebuild it explicitly, AI on volume and humans on judgment, and write the new standard operating procedure down. "Use AI for this" is not a process.
Weeks 11–12, instrument, re-baseline and read out
Measure the workflow for two weeks before and two weeks after on cycle time, quality (error or rework rate) and operator satisfaction. Re-run the day-30 skill baseline org-wide. Present the day-90 readout, the usage curve, the skill delta, the workflow before and after, and the next two candidate workflows, then ask for phase-two scope in the same meeting.
Gate 3 at day 90: the workflow improves on at least one of the three instruments, the skill delta is positive in every role family, and the evidence file is current. Pass and you scale to two workflows a quarter; fail and the readout is itself a fundable finding.
Failure modes to pre-empt
The four ways a 90-day plan quietly dies.Most stalled rollouts fail the same four ways, and each shows a clear symptom by day 45. Name them now and you can catch them at a gate instead of at a post-mortem.
Symptom by day 45: licenses everywhere, usage nowhere
When "we deployed Copilot" is the whole strategy, keep the order permission, then skills, then tools, and buy nothing new until Gate 1 passes.
Symptom by day 45: one all-hands webinar, zero behavior change
High attendance and no change means the format is wrong. Run role-shaped daily reps only, and let the webinar budget fund champion hours instead.
Symptom by day 45: execs mandate what they do not model
Teams read the real signal. Make the sponsor visible weekly use a Gate 1 item, in their calendar and in public.
Symptom by day 45: the workflow feels faster and finance shrugs
A pilot with no instruments cannot fund phase two. Choose the instruments and baseline them in weeks 9 and 10, before the redesign, not in week 12.
Every failure mode has a matching gate item. Pre-registering the gates is what turns a miss into a finding instead of a surprise.
Are you ready to start the clock?
Eight questions, about three minutes.Check every statement that is true today.
| Score | Band | What it means |
|---|---|---|
| 0–2 | Not yet | Starting now produces a tool rollout, not a capability. Spend two weeks on owners and policy first. |
| 3–4 | Almost | Close the policy and baseline gaps, then start the clock. |
| 5–6 | Ready | Start Monday. Guard the gates, especially the five-hour line. |
| 7–8 | Running | You are mid-roadmap already. Use the gates as your steering meetings. |
Score yourself honestly. Two weeks spent on owners and policy now is cheaper than a rollout that reverts to shadow usage in month two.
The rollout checklist
What to do this week, this month and this quarter, then the one-page takeaways.The rollout checklist
Takeaways, the one-page version
Start Monday with the five hats and the one-page policy. Everything else in this plan hangs off those two moves.
Sources and method
Every external numeric claim in this report points to one of these 2024 to 2026 sources. Forecasts and self-reported surveys are labelled so they are not mistaken for causal proof.
Superagency in the Workplace
Employees three times more likely to run at least 30% of daily work on AI than leaders estimate; 48% rank training the top factor for AI adoption; about half report minimal or no training.
mckinsey.comAI at Work 2025
Regular AI usage jumps past about five hours of training and is strongest when paired with in-person coaching; more than half of employees turn to unsanctioned tools when none are provided.
bcg.comEU AI Act (Regulation 2024/1689), Article 4 AI literacy duty
AI literacy duty applicable since 2 February 2025 for providers and deployers of AI systems in the EU, of any size; enforcement machinery from August 2026; measures must be role-tailored and documented. Not legal advice.
eur-lex.europa.euOmie rollout playbook and the companion report, The AI-Ready Workforce
The three-phase structure, the five owner hats, the checkpoint gates and the failure modes are Omie’s rollout playbook for organizations of 50 to 500 people. Strategy context lives in the companion report, The AI-Ready Workforce.
tryomie.com