The short answer.

You roll AI out across a company one team at a time, and inside each team one named piece of work at a time: a workflow with written instructions, one owner and one person who checks every output, running in the assistant the company already pays for. The next workflow starts only when the one before it is in daily use, with no more than two or three in development across the company at once. You know it is working when each workflow runs without anyone asking, its checker accepts the drafts with light edits, and the work measures better than a baseline you wrote down before the first run. A short monthly review asks those questions of every workflow and updates the plan the same day.

The articles behind this guide each settle one part of that answer. Read together, they establish one thing none of them says on its own: rolling out and measuring are the same record. The facts you write down to choose and launch a workflow (the team, the piece of work, the owner, the checker, what it starts from, what “in daily use” means and how the work goes today) are the very facts you later read to judge it. Write them once, at the start, and “is AI working?” has an answer for every workflow, every month.

Rolling out is the third of three steps set out in How organizations plan their use of AI: plan it, develop the workflows, deploy them.

Roll out pieces of work, not licenses.

The unit of an AI rollout is a piece of work, not a seat: one named job a team repeats, such as the month-end close checklist, the brief before a client call or the reply to a new inquiry, written down as a workflow the assistant prepares the same way every time. Seats and a training session teach the assistant in general, and on the next working day each person faces their own work with no instructions for it.

Three things are settled before the first team starts.

  • One approved assistant.The ChatGPT, Claude, Copilot or Gemini account the company already pays for.
  • Two paragraphs of rules.Which assistant is approved and on which account, then what information may go into it and whom to ask when someone is unsure.
  • One rollout owner, by name.Usually the head of operations, the chief of staff or the owner of the business. The rollout owner sets the order of teams and decides when the next team starts.

How to roll out ChatGPT, Claude, Copilot or Gemini across a company sets out the company-wide sequence in full. Because a workflow is the thing you roll out, a workflow is also the thing you measure; a seat count says only that the assistant is installed.

Which team goes first, and what does it build first?

The first team is the one whose week repeats most, whose work starts from records someone can export, where one person already reviews the output, and whose lead says yes to owning the first workflow. Keenness and urgency do not decide it: the keenest team often does work that changes shape every time, and a team fighting a fire has no lead with room to test anything.

Put the same four questions to each team lead, in person.

  1. How much of your team’s week repeats?The pieces of work done the same way every week or month.
  2. Where do the inputs live?Exportable today from a system, scattered across inboxes, or in people’s heads.
  3. Who reviews the output today?One name, or “nobody yet”.
  4. Will you own the first workflow?Yes, not now, or not asked yet. Only a yes lets the team go first.

Which team should go first with AI? turns those questions into a comparison sheet.

The first team then builds three pieces of work in sequence, not one and not ten. Each candidate passes four tests: it happens at least weekly, its steps can be said in order, one named person checks it, and it needs nothing from your business systems beyond a file someone already exports. Among the pieces that pass, choose the surest first, then one that reuses the first one’s records or checker, then one that stretches the team. How to choose the first three things to build with AI has the shortlist sheet.

Notice that the test for a team and the test for a piece of work are the same four questions asked at two levels: does it repeat, are the records there, is there a reviewer, is there an owner. Those answers belong on the plan, and FullyPowered AI Planner is where the plan lives: you set out your teams first, then their roles and the work between them, choose what each team builds first, and put an owner on every workflow.

How a workflow is built and deployed to a whole team.

A workflow is built by writing down how the team’s best person does one piece of work, testing it on last month’s real cases, and deployed by running one current version from one shared place: the owner first, then everyone in the role. Building takes five moves: pick the work from the plan, capture how the best person does it, write it as instructions (inputs, steps, checks, output and the point where a person decides), test it on finished cases where the right answer is known, and hand it to its owner. We have a plan. How do we build the first AI workflow? walks through each move.

Common work can start from a free workflow template in the Workflows Directory; templates are samples and a starting point, and work particular to your company is built from your own people’s know-how. When you want the building done with you, a 3-Day Workflow Sprint or a 2-Week Workflow Sprint takes one piece of work from how your team does it today to a tested workflow your people run.

Deploying follows four habits that keep a workflow the same for everyone: one copy in the company’s shared workspace, one person who edits it, a one-line change note per release, and a retest of the saved cases after every change. How to deploy AI workflows to a whole team, and keep them working covers the routine; for the first month, the rollout checklist gives the owner six steps: introduce the workflow on a real case, have each person run it on their own work in the first week, set who reviews, keep an exception log, name one place for questions, and meet a few weeks in to decide what changes. The first 90 days of an AI adoption plan lays the same path across a quarter.

A one-page workflow record to copy. Fill in one record per workflow before its first run. Each later section of this guide reads from it.

  1. Team and lead.______
  2. The piece of work.One sentence, named for what it produces: ______
  3. What it starts from.The records, exports or documents, and where they come from: ______
  4. What it produces, and who receives it.______
  5. Owner.The one person who edits it and decides when it changes: ______
  6. Checker.The one person who reviews every output and approves anything that sends, pays or changes a record: ______
  7. The standard.What every output must contain, what it must never contain, and the sources it may draw from: ______
  8. What “in daily use” means.Who runs it, how often, and what the checker signs off: ______
  9. The baseline.From recent cases, before the first run: time from start to finished, hands-on effort, how often it comes back, how long it waits. Measured by: ______
  10. What the freed time goes to.Named work: ______
  11. Where the exception and send-back log lives.______
  12. The decision.Expand, adjust or stop, the date, who decided and why: ______
  13. What comes next.The team’s next workflow, or the team that receives this one’s output: ______

How do we get people, skeptics included, to use it?

People use a workflow when it does a named piece of their own work the way their best colleague does it, when their manager runs it first, and when their doubts are answered plainly and in writing. A workflow built from the team’s own method changes the job itself: the report or the brief arrives already prepared in the team’s own format.

People who are wary of AI at work are usually wary of three things, and each has a plain answer.

  • Being judged.Nobody is rated or ranked. A workflow is measured by whether the work is used and useful, never by who used it most.
  • Being watched.You learn how work is done by asking the people who do it. Nothing records screens or keystrokes.
  • Losing their way of working.The workflow is written from how the team already works, and a person still decides on every output.

A team that is openly skeptical is usually answering a vague ask such as “use AI more”. Make the ask specific. Let the team name its own most tedious repeating work, let its most experienced people write what good output looks like as pass-or-fail rules, turn each objection into a test case, and run the workflow on cases the team has already finished so it can compare the drafts with its own answers. The team’s lead owns the workflow and has the final say on going live. How do we get people to actually use AI at work? covers the willing team that has not changed its habits, and How do we bring a skeptical team along with AI? covers the team that does not believe the pitch.

The two articles point to the same move from opposite ends: give the most careful skeptic the job of checking. The rules a skeptic writes become the checker’s list, the cases a skeptic doubts become the test set, and the misses a skeptic catches become fixes in the next version. Adoption and quality are built by the same people on the same page.

Quality stays up because a person decides and a workflow prepares.

Quality stays up when AI prepares the work because the check is designed into the workflow, not left to whoever happens to read the draft. Four parts do it.

  • A written standard.What a finished piece must contain, what it must never contain, which sources it may draw from, and one good example. The person who reviews the work today writes it.
  • A self-check on top of every draft.The workflow lists each required part as present or missing, each fact with its source, and each place it was unsure, so the reviewer reads those places first.
  • One named reviewer.The reviewer checks facts against their sources, completeness against the standard, limits against the never list, and fit against this client and this case. Decisions on what to send, promise or charge stay with the reviewer.
  • A fixed list of send-back reasons.Wrong fact, no source, missing part, over the line, wrong case. The reviewer fixes the draft; the owner fixes the workflow, so the same miss does not come round twice.

How do we keep quality up when AI prepares the work? has a review checklist for each workflow. Anything that sends, pays or changes a record waits for a person to approve it. The send-back log is also your best evidence of whether a workflow is useful.

How do we tell whether it is working and paying back?

You tell whether AI is working by asking three questions of each workflow in order (is it installed, is it used, is it useful) and answering the last one against a baseline written down before the first run, with the time spent checking counted in full. Installed, used or useful? has the sheet for the first two. A workflow that is not used cannot be useful, so do not skip ahead.

For the third question, the baseline records how the work goes today on recent cases: time from start to finished, hands-on effort, how often it comes back with a mistake, and how long it waits between steps. Once the workflow is in daily use, measure the same things the same way, with the reviewer noting checking time. Seat counts, logins and surveys of how people feel measure activity, not the work. How to tell whether AI is paying back sets out the baseline sheet and the three decisions that follow: expand, adjust or stop. A stopped workflow with a written reason is a result.

Time freed pays back only when it goes to named work, decided before the workflow is in daily use: the client call nobody got to, the backlog, the next workflow.

A monthly review keeps all of this current. The owner of the plan runs it and the owner of each workflow attends; nobody else is required. How to run a monthly AI review has the full agenda. The questions fit on one page.

Monthly review questions to copy. Ask them in the same order every month.

  1. Is the workflow used?Who ran it since the last review, who did not, and why: ______
  2. Is it useful?Did the checker accept drafts with light edits, or rewrite them? What does the send-back log show: ______
  3. What changed around it?A procedure, a system’s export, the assistant, the people in the role: ______
  4. What is next for it?Keep, adjust and retest, extend to another role or team, or retire: ______
  5. For the plan: what moves from “next” to “now”?The workflow and its owner: ______
  6. For the plan: what did one team learn that another should take?The lesson, the team that takes it, who carries it: ______
  7. Close.The plan’s owner updates the plan the same day. Date done: ______

How a few pilots become the whole company.

A pilot becomes company practice when someone other than its builder can run it from the written version alone, an owner who is not the builder keeps it current, and the plan names the next team. Good pilots stall for reasons that have nothing to do with the AI: only the enthusiast who built it can run it, nobody keeps it current, and nobody decided who goes second. From a few AI pilots to the whole company gives the fix for each and a one-page case for leadership.

The second workflow continues the first. Follow the first workflow’s output to the team that receives it, because that team’s inputs are already clean and reviewed, or take the same piece of work to another team that repeats it. A candidate waits until one named person will own it and a person can check every output. Keep a “now, next, later” list on the plan, so a team that is not next hears where it stands instead of starting its own experiment. What comes after the first AI workflow? has that sheet.

A company with several offices or sites faces the same question sideways: the same team, doing the same work, in several places. The answer is one workflow per piece of work, one central owner, and a short list of what each site may change (names, opening times, local rules and who reviews), with the best site’s method as the starting version. How do we keep AI use consistent across offices and locations? has the “same everywhere, local to each site” sheet.

Across all three articles, the limit on spreading is attention, not ideas or budget. That is why the pace rule holds everywhere: one new workflow per team, two or three across the company, and the next one starts when the last is in daily use.

What good looks like a year in.

A year into an AI plan, you can answer four things for every team by opening a record, not by asking around: which workflows it runs, who owns each one, when each was last reviewed, and what the team now does with the time.

The test is quick. Pick any team, then any workflow in that team, and check four things: the plan shows it with its owner, the workflow shows its current version and last change, the review notes show it was discussed, and the newest person in the team started with it. What good looks like a year into your AI plan sets out a twelve-line checklist to hold your company against.

Every line of that checklist is a field of the workflow record you filled in before the first run, kept current for a year. A company that wrote the record at the start reaches the year-one picture by keeping it up to date.

One rollout, from the first team to the second.

Illustrative example.

Take a 70-person landscaping company with four teams: crew operations, estimating, accounts and customer service. It already pays for one assistant. The general manager owns the rollout, writes the two paragraphs of rules and asks each lead the four questions. Estimating is the keenest, but every site is different and the owner prices each job, so it waits. Customer service is in its busy season, so it waits, with a date. Accounts builds its invoices every week from the job records crews close out in the scheduling system, the office manager reviews every invoice, and the office manager says yes to owning the first workflow. Accounts goes first.

The office manager’s first three are a draft invoice from each closed job record, with any extra work explained; then a weekly list of finished jobs not yet invoiced, which reuses the same records and the same checker; then a draft reply to a client who questions a bill, which stretches the team. Before the first run, the office manager fills in the workflow record, including the baseline from recent invoices and the work the freed time goes to. A crew lead who doubted the whole idea writes the first rule of the standard: every extra on an invoice traces to a line on the job sheet. The draft invoice starts from a workflow template in the Workflows Directory, adapted to the company’s own invoice and tested on last month’s finished jobs.

The general manager runs the monthly review with the office manager on the first Monday of each month and asks the review questions. When the draft invoice is in daily use, the review moves the weekly list from “next” to “now”. Customer service, which answers clients about their bills, receives the accounts team’s output, so it goes second, with a draft reply to billing questions owned by the customer service lead.

What is different in a company of under 200 people.

A company of under 200 people rolls out AI by conversation, not by program, and that makes it faster. Much of the published advice is written for large enterprises, with steering committees, usage dashboards and a pilot division.

  • The rollout owner is an operator.The head of operations or the owner of the business runs it. IT sets up accounts and access.
  • Every team lead is in reach.The four questions for each team are one conversation with each lead, and the order of teams is settled in a week.
  • One person is often expert, owner and checker.The person who does the work best usually owns it and reviews it too. Plan around that person’s week first, and name a cover for leave.
  • Measurement is a sheet, not a dashboard.The team lead measures the baseline on ten to twenty recent cases and knows why each one went the way it did.
  • One good workflow spreads by word of mouth.When one team runs a workflow every week, the next team hears about it and asks for its own.

That is where 10X the ROI of your existing AI comes from in a smaller company: the seats you already pay for, pointed at named work, one team at a time, and checked every month. Start with your teams in FullyPowered AI Planner, choose what each builds first, and put an owner on every workflow.

Questions people ask.

For short answers across the whole subject of planning AI, Answers takes the questions leaders ask in order.

Should we just give everyone a ChatGPT or Copilot license and run a training session?

Seats and a training session produce a few enthusiasts and many unused seats, because training covers the assistant in general and nothing in the job asks for it afterwards. Roll out named pieces of work instead: one workflow per team, with written instructions, an owner and a person who checks. People keep using an assistant that does their own work the way their best colleague does it.

Which team should go first with AI?

The team whose week repeats most, whose work starts from records someone can export, where one person already reviews the output, and whose lead agrees to own the first workflow. The keenest team and the team with the biggest problem are often the wrong first choice. A team without a lead willing to own the workflow waits, however strong its case.

How many AI workflows should we build at once?

One per team and no more than two or three across the company at a time. Each workflow in development needs its owner to test it on real cases, read every miss and decide when it is ready, and that attention runs out before ideas do. The next workflow starts when the one before it is in daily use.

What do we do about people who don’t trust AI?

Make the ask specific and put the skeptics in charge of the standard. Let the team choose its own most tedious repeating work, write what good output looks like, turn its objections into test cases and compare the workflow with its own finished cases. Tell people in writing that nobody is watched or rated, and that a person decides on every output.

Who checks the work AI prepares?

One named person per workflow, usually the person who reviewed that work before the workflow existed. The reviewer checks the draft against a written standard and its sources and sends it back for a reason from a fixed list. Anything that sends, pays or changes a record waits for that person’s approval.

How do we know if AI is paying back?

Write down how one piece of work goes today before the workflow’s first run: time from start to finished, hands-on effort, how often it comes back and how long it waits. Measure the same things once the workflow is in daily use, counting the time spent checking. Then expand, adjust or stop it on that evidence; seat counts and logins do not answer the question.

What happens in a monthly AI review?

The owner of the AI plan meets the owner of each workflow and asks four questions of every workflow: is it used, is it useful, what changed around it, and what is next. Then two questions of the plan: what moves from “next” to “now”, and what one team learned that another should take. The plan is updated the same day.

Our AI pilot worked. Why hasn’t it spread?

Pilots usually stall because they live in one enthusiast’s head, nobody but the builder keeps them current, and nobody decided which team goes next. Write the pilot down until a colleague can run it alone, give it an owner who is not the builder, and put the next team on the plan. The next team is usually the one that receives the pilot team’s output.