UK Business Hours: 9am - 5pm

AI Transformation15 Sept 2026Read: 17 min read

The enterprise AI playbook, cut down for a 50 to 300 person company

What the enterprise AI playbook looks like scaled down for a 50 to 300 person UK firm: one sponsor, one review slot, two training tiers, no new committee.

John Kelleher

The enterprise AI playbook, cut down for a 50 to 300 person company

John Kelleher is a Claude Certified Architect (Foundations and Professional) and leads SpotDev, a Claude Registered Partner and OpenAI Select Partner.

You have read the enterprise material. It tells you to form a steering committee, appoint champions at every level, run three tiers of training, certify people, and stand up a centre of excellence with rotating staff. You have 120 people, a managing director who also runs sales, and nobody with "transformation" in their job title.

Most of that playbook is right about what has to happen. It is wrong about how much structure a company your size needs to make it happen. This post takes each enterprise mechanism, says what it is for, and gives the version a 50 to 300 person UK firm can staff.

The three-step shape (lay the foundation, run a pilot, scale what works) and the 8 to 12 week pilot timeline come from Anthropic's Enterprise AI Transformation Guide (Oct 2025). The guide is written for organisations with thousands of staff. What follows is our scaled-down version, checked against independent evidence where it exists, with the gaps named where it does not. BCG puts 10% of AI value on algorithms, 20% on technology and data, and 70% on people and processes. That 70% is the subject here. The delivery mechanics are in how to roll out AI department by department and are not repeated.

Why the enterprise version does not fit a company this size

The guide's governance and people layer has six parts: a steering committee of the CEO, functional heads, technology, finance and legal; champions at every level with extra training and leadership access; three training tiers (executives, managers, power users); multi-level certification; a centre of excellence staffed by technical architects, domain experts and data scientists, with 3 to 6 month rotations; and formal wins reporting with monthly executive updates and quarterly benchmarking.

At 5,000 staff, each of those is a fraction of a percent of headcount. At 120, the steering committee is your entire leadership team, meeting about something that does not yet produce revenue. The centre of excellence needs people you do not employ. The certification programme needs someone to design, run and assess it.

There is a second problem. The guide cites no external evidence for the committee, the centre of excellence or the certification structure. They are Anthropic's recommendations, drawn from its own enterprise customers. The independent evidence that exists supports one or two of the six mechanisms, not all of them. The sections below keep what the evidence supports and cut the rest.

What replaces the steering committee

Enterprise version: a standing committee with representation from every function, on its own cadence, owning obstacles, budget and governance.

What the evidence supports: the sponsor, not the committee. Prosci's change-management research found projects with extremely effective sponsors were 79% likely to meet their objectives, against 27% with extremely ineffective sponsors. That is the best-evidenced single lever in the adoption literature, and it is about one person's visible, sustained attention, not a group.

Your version: one named sponsor and one 30-minute slot in a meeting that already exists.

The sponsor is usually the managing director. Not because the MD has time, but because at this size the MD is the only person whose say-so removes obstacles in every department without a negotiation. If the MD delegates sponsorship to the operations director, it works only if everyone knows the MD did so deliberately and will back the decisions.

The 30 minutes go into the existing leadership meeting as a fixed item with three questions: what did we try, what happened, what is blocked. No new meeting, no separate invite list, no deck. When the item routinely takes five minutes because nothing is blocked, that is a good sign, not a reason to drop it.

What you lose is formal cross-functional representation. At 120 people the functional heads are already in the room.

Who owns AI when the MD wears four hats

Enterprise version: a centre of excellence. A dedicated team that sets best practice, supports users, experiments, and takes secondees from the business for 3 to 6 months at a time.

What the evidence says: nothing. We found no independent study of centre-of-excellence effectiveness for AI at any company size, and the guide's own section carries no citation. You are choosing a structure on judgement, so choose the smallest one that does the job.

What the job is: continuity. Strip the enterprise CoE back and the one thing that matters at your size is that decisions are recorded, the next ideas queue somewhere, and the standards live somewhere, so that when a person moves on the programme does not reset to zero. Everything else a CoE does either happens anyway in a small firm or is bought in.

Your version: one named owner, a decision log, and a backlog.

The owner is often the operations lead or whoever already runs the CRM. At the smaller end of the range it may be the sponsor. It is a role, not a job title, and it does not need a technical background. It needs someone who will chase.

On time: in our experience a rollout quarter takes the owner roughly half a day a week (scoping the workflow, sitting in on the build, chasing testers, running the review item). Between quarters, or once a workflow is settled, it drops to an hour or two. That is our observation across engagements, not a study finding, and it moves with how much you do in-house. If the owner is spending two days a week, the scope is wrong, not the owner.

On what happens when the owner leaves: this is the question the enterprise CoE answers with headcount, and the one that most often goes wrong at your size. The answer here is paper, not people. Three documents, kept current, stored where the leadership team can find them:

A decision log. One line per decision: what, when, why, who. "Sales AI drafts follow-ups but does not send them. 14 Apr. Agreed by MD and Head of Sales."
A backlog. The workflows people have suggested, ranked, with a sentence on why each is or is not next.
A register of what is running. Each AI workflow, what it touches, who owns it day to day, when it was last reviewed.

If a replacement can read those three and pick up the review item the following week, you have the continuity the enterprise CoE exists to provide, for an hour a month of upkeep. If they have to reconstruct it from memory and chat history, you did not have a programme. You had a person.

Champions at this size

Enterprise version: champions at every level, selected deliberately, given extra training, direct leadership access and public recognition.

What the evidence says: the mechanism (a trusted peer normalising a new way of working) is well established in adoption research from adjacent fields. The AI-specific percentages quoted for champions programmes come from vendor case studies of single companies, not independent studies. Treat the mechanism as sound and the numbers as anecdote.

Your version: one to three people already using AI well, given two things: protected time and a standing invitation to the sponsor's review item. No programme, no selection process, no badge. The argument for why a champion in each team beats a training course is made in full in AI adoption beyond the chatbox. In a firm of 120 you already know who these people are; the only decision is whether to give them the time.

Two training tiers, not three

Enterprise version: three audiences (executives for strategic context, managers to translate strategy into practice, power users for deep technical training), plus hackathons, plus multi-level certification, plus a recommendation to "consider certification status in promotion decisions".

What the evidence says: the constraint is manager confidence, not staff willingness. The Chartered Management Institute found in Jun 2026 that 68% of managers are still at the stage of experimenting or building pilots, and over 80% agree their own and their team's performance would improve with a better understanding of how to manage AI. The CIPD's Good Work Index 2025 found 16% of employees reporting tasks automated with AI, and 85% of those saying it improved their performance. The bottleneck is not staff. It is managers who do not yet know how to run a team that uses AI.

Your version: two tiers.

People who manage people. A short session, repeated when new managers join, on what changes when their team uses AI: what to check, where the approval lines sit, how to spot output produced without judgement, and what a reasonable expectation looks like. The enterprise model splits this into "executive" and "manager". At your size the executives are the managers.
People who use the tools. Hands-on, on their own work, in the workflow they will run. Not a general course on prompting. The four-level framework in is your sales team AI competent defines what competent means at each stage; AI literacy and Claude training cover the training design.

Cut, and why: the executive tier, because it merges with the manager tier at this size. Hackathons, because they need a critical mass of people with slack in their week, and a monthly show-and-tell (below) does the same job in 30 minutes. Multi-level certification, because someone has to build, run and assess it.

The governance layer: policy, sign-off and the paper trail

Enterprise version: a governance framework covering access controls, usage guidelines, quality standards and compliance protocols, owned by legal and compliance on the steering committee.

Your version: the same four headings in one document, owned by the AI owner and signed off by the sponsor. The content is already in an AI governance framework for UK businesses and the AI policy template. Two things a firm this size gets wrong:

The first is treating the policy as the governance. The policy is the rules. The governance is the decision log, the register and the review item. If the policy says "AI does not send customer emails without approval" and nobody can point to where that approval happens, you have a document, not governance.

The second is the EU AI Act. Article 4, the AI literacy duty, has applied since 2 Feb 2025. It binds a UK firm only where it places an AI system on the EU market, puts one into service in the EU, or where the system's output is used in the EU (Article 2). Having EU customers does not by itself put you in scope; using AI in a way whose output reaches the EU can. If that describes you, the wording quoted on most compliance blogs is out of date. Regulation (EU) 2026/1744, adopted 8 Jul 2026 and now in force, replaced Article 4. The duty now is to "take measures to support the development of AI literacy" of staff and others operating AI on your behalf, and the regulation states that this does not require any provider or deployer to guarantee a specific level of AI literacy in any individual. Its recitals say the old "sufficient level" wording was an undue burden on smaller enterprises, and the Commission is to publish practical compliance examples. For a firm in scope, the two-tier training above with attendance recorded is a reasonable measure. For a firm out of scope, it is good practice and nothing more. Check scope with a lawyer if you sell into the EU.

Wins reporting without new infrastructure

Enterprise version: monthly executive updates, internal channels and newsletters, all-hands segments, quarterly user surveys, benchmarking against industry.

Your version: five minutes in a meeting that already exists, and one question tracked monthly.

The five minutes: at a team or all-hands meeting, whoever tried something with AI that month says what happened, good or bad. Bad matters as much as good. A firm where only successes get airtime stops hearing about failures, and failures are where approval boundaries get tested.

The one question: "Who used AI for something real this month, and what was it?" Asked by the owner, answered by the managers, written into the register. Over two quarters it gives an adoption trend without a dashboard. It replaces the enterprise adoption dashboard (daily active users, feature utilisation, session frequency) deliberately: at your size the question is cheaper, and the answer more informative, because it makes a manager name a workflow rather than report a login count.

Rolling out department by department

The delivery layer beneath all this, which department goes first, one per quarter, what a quarter contains, is in how to roll out AI department by department. The structure above sits over that sequence and does not change from department to department. The same sponsor, owner, review item and register carry across.

The first workflow in that sequence is run as a single pilot; the week-by-week plan, with the baseline, hold-out group and stop criteria, is in how to run an AI pilot in 8 to 12 weeks.

What to expect in the first two quarters

The honest version, from the UK evidence.

Adoption will be broad and shallow at first. The ONS found UK businesses with 10 or more staff reporting AI use rose from around 12% to around 35% between late 2023 and 2026, but only 10% of users describe their use as extensive, and adopters use 1.4 to 1.6 AI technologies on average. The same pattern will appear inside your firm: many people trying something, few using it for real work. The two-tier training and the monthly question exist to close that gap.

Revenue will not move in two quarters. DSIT's 2025 survey of 3,500 UK businesses found 77% of AI adopters reporting no change in revenue and 12% an increase; of that 12%, 16% said AI drove it to a great extent, about 2% of all adopters. Expect time recovered and error rates falling in specific workflows, not a revenue line. If someone promises the revenue line by month six, ask what evidence they are drawing on.

Headcount will not change either. The British Chambers of Commerce found in Mar 2026 that 54% of UK SMEs actively use AI, up from 23% in 2023, and that 95% of AI-using SMEs report no impact on workforce size in the past year, with 86% saying roles are unchanged. Say this to staff early, in writing if you mean it. The productivity case is easier to make when people are not wondering whether they are the saving.

The gains will land in unglamorous places. The Department for Business and Trade trialled 1,000 Microsoft 365 Copilot licences over three months from Oct 2024. Time was saved on drafting, summarising research, transcribing meetings, searching for information and brainstorming; none on image generation, scheduling or producing slide decks. Plan your first workflows around the first list.

Spend will be modest. In the Lloyds Bank Business Barometer of Mar 2026, 66% of UK businesses had invested in AI, and the largest group, 33%, had spent under £25,000. What moves the number is data access, whether the AI may write to systems or only read, integration depth, how much human review each output needs, and seat count. A diagnostic scopes those before you commit.

Two quarters in, a firm that has done the above will have one or two workflows running with clear approval boundaries, a register a newcomer could read, a leadership team that has heard eight to twelve short reports of what did and did not work, and a view of which department is next. That is not transformation. It is the foundation the enterprise guide describes, built at a size you can sustain.

Frequently asked questions

Does a 50 to 300 person company need an AI Centre of Excellence?

No. The enterprise centre of excellence provides continuity, standards and support with dedicated staff. A firm of 50 to 300 gets the same continuity from one named owner, a written decision log, a ranked backlog and a register of what is running. We found no independent study showing a formal CoE improves AI outcomes at any company size, so the structure is a judgement call, and the smaller version is cheaper to keep alive.

Who should own AI adoption in a company this size?

Two roles. A sponsor, usually the managing director, whose visible backing removes obstacles across departments. Prosci's research found projects with extremely effective sponsors were 79% likely to meet objectives against 27% with extremely ineffective ones. And a day-to-day owner, often the operations lead or whoever runs the CRM, spending roughly half a day a week during a rollout quarter and an hour or two between them.

Do we need a steering committee for AI, or is that overkill?

Overkill. The evidence supports the sponsor role inside a steering committee, not the committee itself. At 50 to 300 staff the functional heads already sit in your leadership meeting, so a second forum adds a meeting without adding a voice. Put a fixed 30-minute item into the existing leadership meeting (what we tried, what happened, what is blocked) and let it shrink when nothing is blocked.

What is a realistic AI training programme for a company this size?

Two tiers. A short session for anyone who manages people, covering what to check, where approval lines sit and what good output looks like, because manager confidence is the constraint: the CMI found 68% of managers still at the experimenting stage. Then hands-on training for tool users on their own workflows. Skip hackathons and multi-level certification.

What does an AI adoption timeline look like for a mid-sized UK business?

One department per quarter, with the governance layer set up before the first. Expect the first two quarters to deliver time recovered and fewer errors in specific workflows, not revenue: DSIT found 77% of UK AI adopters reporting no change in revenue. Expect broad, shallow use before deep use, and no change in headcount, which matches what 95% of AI-using UK SMEs report.

The next step

If you want this governance and people layer set up around a real first workflow, the AI Accelerator runs one department per quarter with the sponsor, owner, review item and register built in. If you are not yet sure which workflow, or whether your data and access will support one, start with a diagnostic or the readiness assessment on the AI implementation page. Either gives a scoped answer before you commit to a quarter. The human route is Request a Quote.