Quick Answer: To start an AI transformation, begin with one team rather than the whole organisation. Pick a team whose manager asked for help, hear from everyone in it, and count its week before anything changes. Then move one task to an agent, agree in advance where the saved hours go, and decide after a month whether to keep, widen or stop.
In a six-month Australian government trial in 2024, staff across 56 agencies were given an AI assistant, and only a third of those taking part used it daily, mostly to summarise and rewrite text. Its evaluation recommended detailed workflow analysis to find more uses. Most advice on how to start an AI transformation follows the same order: access for everyone, then training, then a search for something to use it on. That order hands out the tool before anyone knows which work it is for.
Choosing which task first is a decision of its own. What follows is about the team around that task and its first month: which team goes first, what it counts, where its saved hours go and what it hands on. By the end of that month, one agent should be doing work that an agent does well, while people review outcomes rather than produce them.
Why Most AI Transformations Start Too Wide
A 2024 study of 1,665 employees at one large technology firm found that generative AI did not simply spread from experiments chosen at the top and passed from colleague to colleague. The people pushing hardest for wider use and the people resisting it often did the same job, and the authors called for close support with attitudes and skills before scaling up. It is one firm and one questionnaire.
- Licences arrive before tasks. People get access and are left to find uses; in the Australian trial, most use stayed with the jobs the tool suggests by itself.
- Training comes before anyone knows the work. The same trial found training worked best when tailored to each agency, which is hard before you know how a team spends its week.
- One number stands in for the whole organisation. An adoption rate, like a readiness audit, describes the organisation rather than any one team’s week, and hides the teams where something already works.
- The argument happens everywhere at once. Enthusiasts and sceptics share teams, so a wide launch starts the same debate everywhere. One team can settle it on its own work.
Starting an AI transformation everywhere at once teaches the organisation that it has AI, not what AI should do in any one team’s week.
Starting With Everyone or Starting With One Team
A German survey of 440 firms with 50 or more staff found that in 2025 employees were involved in planning AI at only 32 per cent of them, and in putting it in place at 49 per cent. Where works councils were regularly involved, the authors report smoother implementation and greater acceptance among staff. It is a German setting and a correlation, but the point travels: involving the people affected works better than informing them afterwards.
| Starting with everyone | Starting with one team | |
| Who is asked first | Senior leaders and a programme team | Everyone in one team, in confidence |
| What is handed out | Licences and general training | One agent for one named task |
| What gets counted | Logins and usage | Hours on that task, before and after |
| Who has a say | Whoever sits on the steering group | The team, its manager and the checker |
| What a wrong guess costs | Months, and staff patience | A few weeks in one team |
| What the next team gets | A rollout plan | A worked example from inside |
The wide start still has a job. Licences, security settings and rules about the access granted to agents belong at organisation level, and none of it needs an outside programme first. It is groundwork, not the start of anything people notice.
How to Start an AI Transformation With One Team
The justice ministry’s plan, published in July 2025, sets out a scan, pilot and scale approach: pilots test whether a tool is practical and meets users’ needs, stop where they should, and only proven tools spread. It commits to involving frontline users early, and a year on the department reported that a transcription tool for probation officers had been scaled to more than 12,000 of them after its pilots, a count of use rather than of time saved. That is one department’s practice, not a business rule, but it shows how to start an AI transformation without guessing: small, with the people who do the work, and with a way out.
Pick the Team That Asked
The best first team is rarely the one leadership thinks needs AI most. Look for four things: a manager who asked for help rather than being volunteered, work that repeats weekly, someone able to check what an agent produces, and no reorganisation under way. Teams with plenty of small weekly chores usually qualify. So do teams already using AI on their own, sometimes used privately through personal accounts; count what they use first. Avoid a team picked only because its loudest enthusiast volunteered it.
Count the Week Before Anything Changes
Before an agent touches anything, the team counts its own week: each recurring task, how long it takes, how often it comes round and what goes wrong. A week or two of tallies kept by the team beats any outside estimate, and hours per task is a figure everyone recognises. Include the habits nobody mentions because they have become automatic. When you start an AI transformation this way, the end of the month is a comparison rather than an argument, and the figures carry into the full roadmap later.
Agree Where the Saved Hours Go
Saved time does not go anywhere useful on its own. A Korean study from 2026, based on a representative survey of workers, found that generative AI cut working time by 3.8 per cent, but that the time saved and the extra work produced were almost unrelated: people mostly took the gain as a slightly easier day. It is a preprint, but the lesson is direct. Before the agent goes live, the manager and the team write down where the hours will go: the call-back backlog, the work people say they never reach, or a less rushed week, which is fair if chosen on purpose. That note also answers the board’s question later.
The First 30 Days in One Membership Body
Take a professional membership body of around 280 staff, with teams for membership services, events, standards and finance. Starting an AI transformation there began with membership services: its manager had asked for help, the renewal and record queries arrived every week, and the events team was mid-way through a new booking system.
| Week | What happened | Who was involved | What existed by Friday |
| Week 1 | Everyone interviewed in confidence about a normal week | Every adviser and the team leader | The team’s recurring tasks, in their own words |
| Week 2 | Each recurring task timed, with how often it arrives | Two advisers keeping the tally | A starting figure per task |
| Week 3 | One task chosen: replies to members whose development record is missing an event. Checker named | The team, the checker and IT | A working agent, tested on last month’s cases |
| Week 4 | Agent drafts live replies; checker reviews each; saved hours go to call-backs | The team and the checker | Week four against the starting count |
| Day 30 | Keep, widen to the events team later, or stop | Director of membership and the team | A one-page note for the second team |
Because the week one interviews were held in confidence, advisers raised tasks never mentioned in meetings; those were parked with reasons. The rules in the usage policy were settled before week three, so the agent’s access was agreed, not improvised. The named checker was the adviser who knew which records often went wrong, not the most senior person present.
What the First Team Passes to the Second
The second team does not start an AI transformation from scratch: it borrows the method, not the agent. Its work will differ, so what travels is how the first month ran:
- The questions that drew the work out, and those that got polite answers.
- The counting method, so the two starting figures can be compared.
- The checker’s role, written down: what they look at first and how long it takes.
- The parked tasks, with reasons. Some belong to another team, and some will matter once every other team has been asked.
If the method lives only in one person’s head, losing a long-serving adviser can take it with them. The agent rarely transfers. The way of finding the next one does.
What to Leave Out When You Start an AI Transformation
A month with one team needs less than most plans assume. Each of these can wait for the second or third team:
- A central function. A central team earns its place once several agents are running; before the first, it has nothing to look after.
- An organisation-wide list of use cases. One team’s list is enough, and building it shows how to build the bigger one.
- The platform decision. The platform question is easier once you know what one agent needed in practice.
- A detailed roadmap. Order the later steps once you have one result to order them around.
- Company-wide training. Train the first team on its own task until the pilot itself shows what others need.
- An organisation-level return figure. One team’s before and after comes first, and people believe it.
Two things should not wait. Tell everyone else what the first team is doing, briefly and early, rather than saving it all for the wider announcement. And make sure the agent is one the organisation holds from the first day, not one in somebody’s personal account.
Where GrowthNation Fits
GrowthNation is built for organisations that want to start an AI transformation this way. It interviews everyone in the first team, by voice or text and in confidence, so the task list comes from the people who do the work. It turns the chosen task into an agent served centrally and owned by the organisation, so the agent stays when people move on. And it tracks the hours each agent saves against the team’s starting count, so the second team begins with a real figure to beat.
Start With the Team That Asked
Name one team whose manager asked for help. Ask how long its most repeated task took last week. If nobody knows, counting it is the first week’s work.
Book a walkthrough and GrowthNation will interview that team, then hand you three things:
- The team’s week, counted. Every recurring task, in its own words, with hours attached.
- The first task to move. Chosen with the people who do it, so you start an AI transformation on work they want handed over.
- A working agent. With its checker named and its saved hours already allocated.
Hours are measured against the week you counted, so every saving can be checked.
Book a walkthrough: one team heard, one week counted, one working agent.
Join the GrowthNation waitlist. Organisations join in batches, so a place now means your first team is interviewed in the next round.
Frequently Asked Questions
How do you start an AI transformation?
Begin with one team, not the whole organisation. Choose a team whose manager asked for help, interview everyone in it, count its week, move one task to an agent, agree where the saved hours go, and decide after a month whether to keep, widen or stop.
Should an AI transformation start with one team or the whole organisation?
The work should start with one team, while licences, security and rules are set for the whole organisation. A wide rollout spreads access faster than know-how. In the UK government’s 2026 AI adoption survey, 67 per cent of businesses already using AI still named skills, expertise or understanding of use cases as a barrier.
Which team should go first?
The best first team has a manager who asked for help, weekly repeated work, someone able to check an agent’s output, and no reorganisation under way. It is rarely the team leadership thinks needs AI most.
How long does the first phase of an AI transformation take?
About a month is enough to interview one team, count its week, put one agent to work and make a decision. Access or data problems can stretch that, but the decision should still be made, even if the answer is to wait.
Do you need an AI strategy before you start?
A short one, covering ownership and basic rules for tools, helps. A detailed strategy is better written after the first team’s month, when it can draw on real tasks and real figures.
How do you know the first team’s start worked?
It worked if the hours on the chosen task fell against the starting count, the saved hours went where agreed, the checker still changes some outputs, and the team wants to keep the agent. If you start an AI transformation and none of these can be shown, change the task before widening.