UK small businesses are experimenting with AI at pace. Simply Business's 2026 SME Insights Report put adoption at 47%, more than double the year before — yet only 19% felt "very confident" using it day to day. Plenty of tools; thinner judgement about where they earn their keep.
Against that backdrop, a late-September Diginomica interview with NEXT CEO Simon Wolfson is unusually useful. NEXT is not an SME. Its AI spend is rising from £200,000 last year to £1.2 million this year, with at least £3 million planned next. The transferable part is not the budget. It is the insistence that application design, guardrails and cost control matter more than the newest model demo.
What Wolfson is deliberately not doing
Wolfson's line is blunt: the value of AI sits in the applications it supports, and those applications have to be designed by people who run the business — not by people who only have access to the technology.
He also draws a hard line around creative work that sells. Early experiments with AI-generated graphics looked exciting in the studio; on the shop floor, he says, they largely failed to sell. NEXT put more time back into human drawing, colour and craft for the designs customers respond to. Trend spotting from a chat window is treated the same way: models summarise what has already happened; they do not replace buyers walking mills, fairs and galleries.
For a growing UK firm, the lesson is not "never use GenAI for creative." It is measure whether the output converts, and keep humans on the work that still needs taste, brand judgement and customer empathy.
Assistive AI vs agentic AI (in plain terms)
Wolfson separates two layers that many SME roadmaps blur.
Assistive AI helps a person do the job faster — coding copilots are his favourite example ("spell check and predictive text for coders"). NEXT introduced that in 2024 and saw a real effect on cost as a percentage of sales.
Agentic AI does stretches of the work itself. He likens the jump to moving from a calculator to a spreadsheet. NEXT's plan is to sit agents alongside roles in its Agile lifecycle — ideas, specification, coding, deployment, support — so people shift from doing every step to managing agents that do them.
That sounds futuristic until you hear the caveat he repeats: you cannot buy an agent out of the box and tell it to "do my business specification." Someone still has to supply business context, rules, security, training and a design that is not locked to a single LLM vendor. Skip that work and you overpay for tokens while the agent makes expensive mistakes.
The productivity story with the awkward footnote
Diginomica records a vivid example. A share function that traditionally looked like 11 days of coding work took a coding agent about 24 minutes to draft; with human review it was still only about half a day. On that slice, the gain looks enormous.
Then comes the honest total: across the wider project, NEXT only saved about 17% of the time. Building and stitching agents together ate a lot of the win. Wolfson's point is that the big gains arrive when agents can hand work to each other safely — and that takes engineering discipline, not a weekend plug-in.
He still targets roughly 30% productivity improvement by February 2028, and cites mainframe modernisation that once looked like a £50 million, business-stopping programme becoming a modular £10 million effort with AI-assisted coding. Again: enterprise scale. The SME translation is smaller: use agents to unlock projects you previously deferred because they were too slow or too risky to staff, not because a vendor promised magic.
A pragmatic checklist for UK scale-ups
Steal the method, not the headcount.
- Pick one measurable workflow, not "AI for the company." Returns processing, catalogue enrichment, support triage, internal reporting — something with a clear before/after.
- Start assistive before agentic. Copilots and tightly scoped automations teach your team what good prompts, data and review look like. Jumping straight to unsupervised agents is how trust dies.
- Write the rules down. Who may the agent email? Which systems can it write to? What must a human approve? Security is part of the product, not a later bolt-on.
- Design for vendor independence. Prefer patterns you can move between models and tools so you are not trapped when pricing or quality shifts.
- Cost the tokens like labour. Wolfson's warning applies at any size: poorly designed agents can burn more on compute than you used to spend on people for the same output. Cap usage, log runs, kill chatty loops.
- Keep humans on brand and taste. If the output sells product, defines brand voice or makes a legal/compliance call, review it. Measure conversion, not just "words produced."
- Plan the stitch. One clever agent that cannot talk cleanly to CRM, ecommerce, finance or support will plateau. Integration and data quality decide whether the 17% becomes something closer to Wolfson's longer-term target.
Kingfisher's same-week results (AI personalisation up 16%, about £100 million of attributed sales, plus heavier bets on search, apps and agent-enabled shopping) underline a parallel point for retailers: personalisation and navigation pay when catalogue data and fulfilment are already trustworthy. Agents do not fix a muddy product feed.
Why this matters for Clarity Growth
Clarity Growth helps UK SMEs and scale-ups choose where AI and automation belong in real operations — websites, ecommerce, connected systems — without buying hype. NEXT's briefing is a useful north star: assistive tools first, agents only where context and guardrails exist, and ruthless attention to whether the application actually serves the business. That is the same discipline we use when we connect tools, tidy data foundations and design sites and workflows that humans (and, increasingly, agents) can trust.
Sources
- Avoiding the traps and keeping AI adoption on a pragmatic level — what's next for UK retailer Next? — Diginomica (22 September 2026)
- AI use among UK small businesses more than doubles in a year — Simply Business (8 September 2026)
- Home improvement retailer Kingfisher taps AI tech driven personalisation as it ups profit target — Retail Technology Innovation Hub (22 September 2026)
- Artificial intelligence in UK businesses — Office for National Statistics (20 July 2026)