The most consequential AI deployment in your organisation is probably one you have never approved, cannot see, and could not name. A new study of 12,637 real-world AI use cases finds that the technology has spread furthest where management has the least visibility: individuals quietly closing tickets twice as fast, drafting their own performance reviews, and in one case building an agent to do half their job in secret. For UK leaders who have spent the past year commissioning AI strategies and signing enterprise licences, the finding is uncomfortable. The revolution arrived from below, and most of it is happening in the shadows.
The gap between the strategy deck and the keyboard
Every UK boardroom now has an AI position. There are steering committees, governance frameworks, pilot budgets and at least one slide promising transformation. What there is rather less of is evidence that any of this drives what employees actually do when they open a chat window. The third annual AI in the Wild study, run by Marc Zao-Sanders and Sara Biuk, analysed 12,637 use cases drawn from nearly 50,000 records collected between March 2025 and February 2026 across Reddit, Quora, LinkedIn, TikTok, YouTube and published articles. It is, in effect, a year-long eavesdrop on how people describe using AI in their own words, and the picture it paints is grassroots, not top-down.
The headline numbers tell you why this matters at scale. Regular ChatGPT users reached 900 million over the year; Gemini passed 750 million. That is adoption on the order of the smartphone, achieved in roughly three and a half years. Yet the study is blunt about the corporate dividend: in the business world, the researchers see “lots of activity producing marginal rather than game-changing benefits, so far.” The mismatch between that sentence and the average UK transformation budget is the story.
Strategic Reality: AI adoption in the workplace is real, enormous and almost entirely employee-led. The study finds 63 of the top 100 use cases are work-related or apply at both home and work — yet notes these “aren’t top-down, centrally-managed, corporate AI initiatives. People are mostly doing these things on their own.” Your AI programme is competing with what staff already do without it.
For a UK economy with a long-standing productivity problem, this should reframe the question leaders are asking. The bottleneck is not getting AI into people’s hands; it is already there. The bottleneck is converting millions of private, ad hoc, often hidden uses into gains the organisation can see, trust and compound.
The real story: usage is widening, not deepening
The instinct is to read each year’s data as a clean break: emotional uses overtake technical ones, agents replace chatbots. The researchers warn against it. Trends are “shifts in emphasis, rather than stark ruptures,” and the pool of users keeps growing, so more people doing one thing does not mean fewer doing another. What the 2026 data shows is breadth: people are pointing generative AI at an ever-wider range of tasks, from the mundane to the intimate.
| What the data shows | 2026 finding | Why it matters for UK organisations |
|---|---|---|
| Use cases analysed | 12,637 (from ~50,000 records) | An order of magnitude larger than prior years — a credible read on real behaviour |
| Top use case | Therapy/companionship | 11% of the dataset, up from 5% the year before |
| Work-related share | 63 of the top 100 use cases | Most workplace use is individual, not corporate-led |
| New work entries | Autonomous agentic operations (#6), vibe coding (#21) | Early and small-scale, not yet production reality |
| Documented business ROI | Rare; one agency cites a 20–30% conversion lift | Quantified returns are the exception, not the rule |
Two findings deserve a UK leader’s full attention. The first is that the single most common use of this technology is not coding or analysis but emotional support — therapy and companionship, now more than 1,400 entries and 11% of everything observed. The second is that genuine business transformation, the thing the budgets are predicated on, barely registers. The examples the researchers found were “early stage, the benefits unquantified, and the beneficiaries SMEs.” That last word matters more in Britain than almost anywhere, given how much of the economy is small and medium-sized firms.
Critical Context: The clearest documented business return in the entire study is a single marketing agency reporting a 20–30% conversion lift from AI-personalised email campaigns. The researchers call explicitly stated ROI like this “rare in the data we have.” Anyone presenting a transformation business case should be honest that hard, repeatable numbers remain scarce in the wild.
What is really happening at work: three tiers and a lot of hiding
Strip the study down to its workplace findings and a structure appears. AI at work is operating at three levels, and the value thins out sharply as you climb.
At the base sits efficiency: first drafts of requirements, summarising long meetings, explaining cost trends in plain language, preparing stakeholder-ready notes. One contributor captured the mood precisely: “The value isn’t intelligence — it’s time saved and clarity gained.” This tier is widespread, low-risk and genuinely useful. It is also unglamorous, which is exactly why it rarely appears in a transformation narrative.
Above that sits growth: using AI to make sales and marketing measurably more effective. Here the evidence exists but is thinner, and the honest returns are mostly anecdotal. At the top sits transformation, rethinking how the business works. The study found gestures towards it — a founder launching a company “from concept to launch in weeks,” someone reworking a restaurant into a takeaway model — but the initiatives were embryonic and the benefits unmeasured. The pyramid is wide at the bottom and very narrow at the top.
Reality Check: Most UK AI value in 2026 lives in the efficiency tier — modest, individual time savings — not the transformation tier the strategy decks promise. Planning as if the pyramid were inverted, with transformation as the base case, sets budgets and expectations against the grain of what people actually achieve.
What complicates all three tiers is concealment. The study describes “shadow usage” as common, and the quotations are striking. “I’m closing tickets 2x faster, my code reviews have fewer issues, and I got praised in my last performance review. But here’s the thing: nobody knows I’m using AI.” Another contributor described building an agent to do roughly half their role in secret after management rejected the idea, then spending the freed time on a side business. The drivers are rational: governance restrictions, cut-down enterprise tools, reputational risk, and the fear of looking like a cheat or a redundancy candidate. The gentle tailwinds — licences bought, training offered, encouragement from the top — are simply weaker than those headwinds.
The human factor: thinkslop, dependency and the agency question
The study’s sharpest contribution is a warning dressed as a coinage. Following the workplace term “workslop,” the authors name a cognitive cousin: “thinkslop” — the lazy, sloppy thinking that excessive AI use can produce. In at least a quarter of the top use cases, people are asking AI to do part of their thinking for them. The mechanisms are familiar to anyone who has caught themselves doing it: reaching for a prompt before forming an intention, outsourcing the mental work that builds judgement, ceasing to write and therefore ceasing to think, and being flattered by a system optimised to keep you engaged.
That last point is a governance issue in disguise. As one contributor put it, “AI is gaslighting you into thinking you’re a genius so you’ll keep using it.” A workforce being told its mediocre work is excellent will stop refining too early, a quiet quality risk no dashboard captures. The counter-pattern, though, is also in the data: people who use AI as a sparring partner, instructing it to poke holes in their arguments, report sharper thinking. “AI is a mirror — not a genie. Use it as such.” The difference between thinkslop and rigour is not the tool; it is how people are taught to hold it.
The emotional findings raise the stakes further, and they land directly on UK public policy. Hamilton Morrin, an academic clinical fellow in neuropsychiatry at King’s College London, told the researchers that “given long waiting lists and difficulty accessing mental health care and therapy in many countries, it’s perhaps not a surprise that increasing numbers are turning to generative AI for support.” For a country living with NHS mental health waiting lists, the surge in therapy and companionship use is not a curiosity; it is a population partially self-medicating with a tool its makers explicitly say is no substitute for trained professionals. The reported cases of AI-related psychosis and harm make clear how high the stakes run.
| Stakeholder | What they assume is happening | What the study suggests instead |
|---|---|---|
| Senior leaders | The AI strategy drives adoption | Adoption is employee-led; the strategy is largely downstream of it |
| IT and governance | Restrictions contain risk | Restrictions push usage into the shadows, where risk is unmanaged |
| HR and people teams | AI is a productivity topic | It is also a wellbeing and cognitive-skills topic, including emotional reliance |
| Line managers | Visible tooling reflects real usage | Much of the highest-value usage is deliberately hidden from them |
| Employees | Hiding AI use protects them | It forfeits shared learning, support and any institutional credit |
Strategic recommendations: govern the reality, not the slide
The correction for UK organisations is not another transformation programme. It is to manage the adoption that already exists — to pull shadow usage into the light, protect cognitive quality, and set expectations at the height the evidence supports rather than the height the vendor promised.
- End the conditions that create shadow usage. Audit why staff hide AI use — restrictive tooling, vague policy, fear of judgement — and remove those drivers. A clear, permissive, well-bounded policy surfaces usage you can then support and learn from. Punishment guarantees concealment.
- Reset the value narrative around efficiency. Make time-saved and clarity-gained the headline, not transformation. These are the gains people actually report; building the business case on them is both honest and achievable, and it stops good projects being judged against an impossible bar.
- Teach the sparring-partner pattern explicitly. Train people to use AI to challenge their thinking, not replace it — to draft last, not first, and to ask the model to find the holes in their work. This single habit is the practical defence against thinkslop.
- Treat emotional and wellbeing use as an HR matter. Acknowledge that staff are using AI for emotional support, signpost real mental-health provision, and be clear about the limits and risks of chatbots in that role.
SME Advantage: The study’s only real transformation examples came from small firms — a founder reaching launch in weeks, an owner remodelling a business with AI as tutor. Without legacy systems, committee sign-off or governance drag, UK SMEs can act on AI’s potential faster than large incumbents. For smaller organisations, the absence of a sprawling AI programme is an asset, not a gap.
Take Action: Run a no-blame “how are you actually using AI?” survey this quarter. The gap between what your governance assumes and what the responses reveal is the most valuable AI map you will get this year — and it costs an afternoon, not a consulting retainer.
Hidden challenges UK leaders should plan for
Acting on this evidence surfaces four problems that are easy to miss until they bite.
The first is that visibility cuts both ways. Inviting staff to disclose AI use builds an accurate picture, but it also creates a record — of who depends on AI, for what, and how heavily. Handled carelessly, that record becomes a redundancy shortlist, which is precisely the fear driving concealment in the first place. Disclosure has to come with a credible, stated commitment that it will not be used against people, or it will fail.
The second is the deskilling curve. Efficiency gains today can become capability loss tomorrow if junior staff never develop the judgement that AI is quietly supplying. The cost is deferred and invisible, which makes it dangerous: a team can look more productive for two years whilst its bench of genuine expertise thins out underneath.
Warning ⚠️: A workforce that has outsourced its thinking is fragile in ways that do not show up until the tool is removed, the model changes, or a problem arrives that the AI handles badly. Build deliberate “AI-off” practice into skill-critical roles so judgement is exercised, not just borrowed.
The third is dependency on infrastructure no one in the building controls. When staff anthropomorphise AI — naming it, gendering it, grieving a model update — the organisation has quietly outsourced part of its emotional and cognitive load to systems owned by a handful of foreign companies. A pricing change, a model retirement, or an outage then becomes a wellbeing and continuity issue, not just an IT ticket.
The fourth is the measurement trap. Because real returns are mostly anecdotal and the best usage is hidden, the temptation is to manufacture impressive numbers to justify the spend. That produces workslop at the strategic level: polished reporting that reads well and means little. Honest measurement of modest gains beats confident measurement of imaginary ones.
The strategic takeaway
The 2026 picture is not the one most UK AI strategies were written for. Adoption is vast but employee-led; the headline use is emotional, not commercial; workplace value is real but modest and frequently hidden; and genuine transformation remains a promissory note, paid out mostly to small firms nimble enough to cash it. The organisations that do well from here will be the ones that stop fighting this reality and start governing it.
Three things are worth carrying forward:
- Adoption is already won; integration is the actual job. The technology is in everyone’s hands. The unmet challenge is turning private, hidden, individual use into visible, supported, compounding organisational capability.
- Manage the shadow, don’t deny it. Shadow usage is a signal that your tooling and policy are out of step with how people work. Surfacing it safely is more valuable than suppressing it.
- Protect thinking, not just productivity. The lasting risk is not that AI does the work badly, but that it does it well enough that people stop developing the judgement to know when it has not. Cognitive quality is now a leadership responsibility.
For UK leaders, the most useful next step is also the cheapest: find out what your people actually do with AI, without judgement, and plan from that truth rather than from the strategy deck. As the study’s authors put it, how much of our agency survives such a powerful, always-on service is partly up to the AI companies — “but most is still up to us.”
If you want to discuss what real-world AI adoption means for your organisation, get in touch — or explore our other strategic insights on AI in UK business.
Analysis based on “How People Are Really Using AI in 2026” by Marc Zao-Sanders, published in Harvard Business Review on 1 June 2026, drawing on the third edition of the AI in the Wild study conducted with Sara Biuk. This article is original Resultsense analysis; the UK framing, interpretation and recommendations are our own.