A managing director at a 12-person consultancy bought Microsoft 365 Copilot licences for the whole team in January. By April, three people were using it regularly. Eight had logged in once. One had never opened it. The tool was fine. The problem was announcing “everyone uses AI from Monday” without any preparation, which turned a promising capability into another overlooked subscription.
The pattern is familiar across UK services firms right now. Around 70% of UK SMEs use some form of AI, but only 31% report a clear return on investment, according to industry analysis. The gap sits almost entirely in adoption: tools are being deployed faster than people are being prepared to use them.
What does “adoption without a grand rollout” actually mean?
Gradual adoption means picking one or two existing workflows, testing them with a small group of volunteers for four weeks, measuring the result, and only then spreading what works. The rest of the team carries on as normal during that period. Adoption spreads because colleagues see time being recovered, not because a company-wide mandate says it must.
A grand rollout assumes the tool will sell itself once installed. A phased approach assumes nothing: it starts where the evidence is thinnest and earns its way outward. The UK Government AI Playbook and the Government Digital Service both frame this as the only responsible path, stressing iterative pilots over big-bang deployments. For a 15-person services firm, the same logic holds at lighter weight.
The practical marker that separates the two: a phased approach has a kill criterion from day one. If the pilot does not produce a measurable improvement within 30 days, it stops. A rollout rarely has one.
A phased approach also fits how small teams actually work. A firm of 15 does not have a change management function. Running two or three people on a 30-day pilot is a manageable ask on top of client work. Asking fifteen people to change how they work simultaneously, before anyone has seen it work, is the ask that produces the empty licences.
Why does a firm-wide mandate so often produce empty licences?
When a tool is announced to the whole firm without role-specific training, staff absorb it as another obligation rather than something genuinely useful. A 2024 CIPD report on people and AI at work found that 49% of UK employees felt they had insufficient training to use AI tools effectively. Among those who felt well-trained, the share reporting positive workload impacts more than doubled.
Microsoft’s own data on Copilot early access deployments showed 29% faster document drafting and 22% faster email handling, but only in firms that provided role-specific training and clear use cases for each function. The licence alone produced almost nothing.
The CIPD finding points to a predictable sequence: announce the tool, skip the training, watch adoption flatten, conclude “our team just isn’t ready for AI”. The sequence repeats in firm after firm. The UK Government AI Upskilling Fund, which offers up to 50% co-funding for AI skills training in SMEs, was launched in response to exactly this pattern. The bottleneck is skills, not software.
Where do you start if you are not running a rollout?
Start by listing the ten most time-consuming weekly tasks per role, scoring each on volume, repetitiveness, and risk. The best candidates are high-frequency, text-heavy, and low-stakes: drafting first-pass responses to standard client enquiries, summarising meeting notes, or reformatting proposals. Recruit two or three volunteers from those roles and agree a simple success metric before the pilot begins.
Three other pieces need to be in place before the pilot starts. First, pick one AI platform: Microsoft 365 with Copilot or Google Workspace with Gemini, plus one external assistant if needed. Tool sprawl kills adoption faster than a failed pilot. People spread across four different products rarely go deep on any of them, and the CMA has noted the value of keeping vendor dependencies under regular review even when standardisation is the right starting move.
Second, write a one-page usage policy. NCSC guidance on public generative AI recommends defining which data can and cannot enter AI tools, who approves higher-risk uses, and how staff report problems. A single page is enough. The goal is clarity, not bureaucracy.
Third, name one internal AI champion: a respected operator with some protected time each week to test and document workflows. The Government AI Playbook emphasises a named service owner for AI initiatives; your champion plays that role at SME scale. They become the person colleagues ask, which is far more effective than a company-wide training event.
For FCA-regulated firms, there is an additional consideration. AI used in client processes falls under existing expectations on operational resilience and outsourcing, and the FCA has been explicit that responsibility for outcomes stays with the firm regardless of what the AI vendor does. That does not mean avoiding AI. It means keeping a clear decision trail.
When should you pause or redesign a pilot?
Kill criteria matter as much as success metrics, and they need to be agreed before the pilot starts. If a process consistently takes longer with AI assistance, or if staff spend more time correcting output than they saved, stop and redesign. The Government Digital Service’s AI Playbook is explicit that every pilot must carry an option to stop or change course, and this applies equally to a 15-person services firm.
Four patterns signal a pilot should stop rather than be pushed forward. Staff are correcting AI output on almost every use. Client-facing work is going out without human review. The tool is being quietly avoided because no one wants to raise problems. Or the success metric was never defined and nobody can say whether it worked.
Two broader patterns undermine this approach before a pilot even starts. If the founder or manager does not use AI in their own work, staff will treat the whole thing as a passing interest regardless of how well the pilot is structured. And if AI is framed primarily around reducing headcount rather than improving the quality and pace of existing work, the CIPD research is clear that quiet resistance follows.
What governance do you need once adoption is working?
Once two or three workflows are embedded, a minimal governance layer matters more than adding the next tool. Keep a simple usage register: what each AI tool is used for, which data it touches, and who is accountable for outputs. This meets ICO accountability requirements under UK GDPR and positions you sensibly if you serve EU clients under the AI Act, which requires deployers of general-purpose AI to maintain appropriate transparency and risk-management records.
Quarterly review is the right cadence. Check where data is stored and who the sub-processors are, aligning with NCSC guidance on access controls for externally hosted tools. Watch for shadow AI: staff will sometimes adopt tools you have not sanctioned, and the better response is to incorporate safe alternatives into your policy rather than issue a blanket ban that gets ignored anyway.
Plan one to three new pilots each quarter rather than a programme. Each pilot is four to eight weeks, with clear metrics and a kill criterion agreed before it begins. This rolling series of small experiments compounds over time without requiring a major restructure, a consulting engagement, or a firm-wide announcement.
The rule for client-facing work stays constant throughout all of it: AI drafts, humans approve. The ICO is unambiguous that where automated processes contribute to decisions with significant effects on individuals, meaningful human involvement is required. For a professional services firm, that means every client-facing output carries a human sign-off, regardless of how confident the AI output looks.
