A leadership team prepares for AI by making four decisions before day one (which use cases, which tools you already pay for, what the guardrails are, and who owns the budget), then running a 90-day programme in four stages: Align (days 1–30, leadership sets direction and rules), Enable (days 15–60, cohorts trained on their own work), Prove (days 45–90, three pilots that ship something measurable), Embed (from day 75, champions and playbooks make it stick). Licences are bought in a day; usage is built over a quarter, and it is built by managers doing the work in front of their teams, not by a launch email.
This page is the roadmap Day Seven uses with clients. It assumes a company of roughly 50 to 500 people that has, or is about to buy, Microsoft 365 Copilot, ChatGPT, Claude or Gemini, and a leadership team that would like the money to turn into something visible.
Most stalled programmes stalled before they started. Four decisions, taken in one leadership meeting, prevent most of it.
1. The first three use cases. Not "AI for the sales team". Three specific tasks, each with a name attached, such as "first-draft tender responses from our library of past bids" or "weekly management report from the CRM export". The selection criteria are below.
2. The tools you already license. Most companies discover they are paying for more than they use. Microsoft 365 E3/E5 tenants have Copilot Chat included; Google Workspace Business and Enterprise plans include Gemini; someone in marketing probably has a ChatGPT Team seat. List what exists, what data terms apply to each, and what is missing. Buy nothing new until the list is done.
3. The guardrails. A one-page acceptable-use policy with approved tools, a "never in a prompt" list, and three risk tiers stating what people can try without asking. We cover this in detail in the AI governance framework for UK SMEs. Without it, cautious people do nothing and bold people use personal accounts.
4. The budget owner and the measure. One executive owns the programme and its budget. The same person names the metric the board will see at day 90. If that person is not in the room, the meeting has not happened.
The four stages overlap. Align does not stop when Enable starts, and Embed activities begin before Prove finishes. The table uses week ranges for that reason.
| Stage | Weeks | Leadership actions | Team actions | What to measure |
|---|---|---|---|---|
| Align | 1–4 | Half-day leadership session: agree three use cases, tool list, policy, tiers, owner, metric. Executives complete two hours of hands-on training themselves. Announce the programme with the three use cases named, not a vision statement. | Baseline survey: who uses what, how often, for what. Champions nominated (one per team). Licences assigned to the cohorts that will be trained first, not to everyone. | Baseline: weekly active users per tool, self-reported hours on the three target tasks, confidence score (1–5) |
| Enable | 3–8 | Managers of the first cohorts attend the same workshop as their teams and share one thing they built each week. Remove blockers raised in workshops (access, data, policy ambiguity) within five working days. | Cohort workshops of around 16 people, half-day or full-day, on-site or live online. Each person leaves with one working prompt, workflow or automation for their own job. Champions run a 30-minute weekly clinic. | Workshop completion; number of "things built" per person; weekly active use in the trained cohorts versus untrained |
| Prove | 7–13 | Approve three pilots against the use cases, each with an owner, a before-number and a finish date. Review at week 10 and kill or expand. Consider a Day Seven Sprint (a seven-day hackathon from a real internal problem to a working build) for the hardest use case. | Pilot teams run the use case for real, log time and quality, and write the one-page playbook as they go. | Time per task before and after; error or rework rate; output volume; a quality check on a sample by someone who did not make it |
| Embed | 11–13 and ongoing | Present day-90 results to the board with the metric agreed in week 1. Fund the next three use cases. Put AI use into manager one-to-ones and team objectives. Schedule the quarterly review of policy and use-case register. | Champions own the playbook library and the weekly clinic. Playbooks become part of onboarding. Second-wave cohorts trained by the same method. | Sustained weekly active use at day 90 and day 180; number of live use cases on the register; playbooks in use |
A worked example of the arithmetic leadership should expect. If a 12-person bid team spends an average of 6 hours each per week on first drafts and a pilot brings that to 4 hours, that is 24 hours a week, or roughly 1,100 hours a year at 46 working weeks. Whether that is worth the licence and training cost depends on what those hours are worth and what the team does with them, which is why the before-number and the owner matter more than the tool.
Because a licence rollout is not a programme. We see the same three causes in almost every company that asks this question, whichever tool they bought.
Seats, not usage. The rollout was measured in licences assigned, and licences were assigned to everyone. Nobody was asked to do anything specific with theirs. Microsoft's own Copilot Dashboard reports active usage rather than seats for a reason: the two numbers diverge within weeks. Fix: assign licences to trained cohorts first, measure weekly active use, and treat an unused seat after 60 days as a reassignment, not a failure.
No manager modelling. The team's manager sent the launch email and never used the tool in a meeting. People copy what their manager does, not what the intranet says. Fix: managers attend the same workshop as their teams and, for the first eight weeks, show one thing they built in each team meeting. Two minutes is enough.
No agreed tasks. "Use Copilot to be more productive" gives nobody a starting point. Fix: three named tasks per team, agreed in Align, taught in Enable. The first time most people use an AI tool successfully is when someone shows them the exact prompt for a task they already do every week.
Three smaller causes that are worth checking. Data access: Copilot in particular is only as useful as the SharePoint and Teams content it can see, and permissions sprawl makes IT nervous enough to restrict it. Ambiguous policy: if people are not sure whether customer data is allowed, they assume not and stop. And training that was a webinar: a recorded demo is not training. Sessions need to be lab-style, on the person's own documents, with a working result at the end.
Score candidates against the criteria below. Three is the number: enough to learn something about different parts of the business, few enough that each has a real owner.
| Criterion | What good looks like | Why it matters |
|---|---|---|
| Frequency | The task happens weekly or more, for many people | Rare tasks do not build habits or produce a measurable number in 90 days |
| Time cost today | You can state the hours per week per person with a straight face | If you cannot baseline it you cannot prove it |
| Data risk | Tier 1 or tier 2 in your policy; no personal data at scale, nothing regulated | The first pilots should not need a DPIA or legal review |
| Text or structured input | The task starts from documents, emails, transcripts, spreadsheets or code | Current tools are strongest here; physical or judgement-heavy tasks are poor first pilots |
| Quality is checkable | A person can tell in minutes whether the output is right | You need a quality measure as well as a time measure |
| A willing owner | Someone in the team wants it, has authority to change how the task is done, and will report the number | Enthusiasm in the team beats enthusiasm in the boardroom |
| Visible to others | Other teams will see the result and want it | The first three use cases are also the internal marketing |
Use cases that pass on most rows in most companies: first drafts of recurring documents from an existing library (bids, reports, job descriptions, policies), meeting transcription to actions and follow-up emails, summarising long inbound material (RFPs, contracts, complaint threads) for triage, and internal knowledge questions answered from the company's own documents. Use cases to leave until wave two or three: anything customer-facing without review, anything touching HR decisions, and anything that needs a new data pipeline before the tool can start.
The programme is the Enable and Prove stages made concrete. What works, and what the gov.uk Skills for AI: what works employer guide broadly supports, is short practical sessions on people's own work, followed by application, rather than long courses or generic e-learning.
Cost depends on group size, format and location; ask for a proposal rather than guessing from a price list.
Champions are the mechanism by which Embed happens without a consultant in the building. The pattern that works:
Champions do not enforce policy. If they become the people who say no, teams stop bringing them use cases.
From week 1, before anyone is trained. The baseline is the single most-skipped step and the reason most day-90 reviews are anecdotes.
Three layers, each with a number on day 1 and day 90:
The method, including a worked example and the traps in self-reported time, is in measuring the return on AI training. If the leadership team cannot agree on the outcome measure in Align, that is useful information: it usually means the use cases are not yet specific enough.
Take four decisions in one meeting: the first three use cases, the tools you already license, the guardrails, and the budget owner and metric. Then have every executive complete two hours of hands-on training so they can model use before the first cohort starts.
Audit existing licences and their data terms, write a one-page policy with risk tiers, baseline the time spent on the target tasks, and nominate champions. Buy nothing new until the audit is done.
Usually three things: licences were assigned to everyone with no agreed tasks, managers did not use the tool visibly, and training was a recorded demo rather than hands-on work on real documents. Assign seats to trained cohorts, agree three tasks per team, and get managers into the same workshop.
Team-level time savings on specific tasks show within the first pilot, typically by week 10. Company-level outcomes take two or three waves. Plan for a quarter to prove and a year to embed.
Three. One is fragile, five are unowned. Pick them with the criteria table above and give each a named owner and a before-number.
Not at first. Assign licences to cohorts as they are trained, measure weekly active use, and expand to teams with an agreed use case. Unused seats after 60 days go back into the pool.
Training is a workshop; take-up is what happens in the eight weeks afterwards. The roadmap treats training as one stage (Enable) inside a 90-day programme with pilots, champions and measurement around it.
Yes, first and separately. A two-hour session is enough for a leadership team to use the tools in their own work and to set guardrails they understand. Managers who do not use the tool cannot model it.
Two waves of cohorts of around 16, grouped by job, with a leadership session first, champions nominated in week 1 and a pilot per team after each workshop. Larger groups run the same pattern with more cohorts; see enterprise AI training.
It depends on group size, format (half-day, full-day, blended, sprint) and whether delivery is on-site or online. Tell us the team size and the three use cases and we will send a proposal.
Fund the next three use cases at the day-90 review, keep the champion clinic and monthly meeting, put AI use into manager one-to-ones, and re-measure at day 180. Momentum is a management routine, not a launch.
If you want the Align session run for your leadership team, with the cohorts, pilots and champions planned from it, see AI training for leadership or
get a training proposalTell us what your team needs. We'll get back to you within one working day.