Britain’s robotics problem is usually told as a story about nerve. We deliberate whilst Japan and South Korea build. The Capgemini Research Institute’s survey of 1,678 senior executives, fielded across January and February 2026, tells a duller and more useful story: 65% of UK respondents place physical AI among their high priorities for the coming three-to-five-year window, against a global average of 66%. On intent, the UK sits within a rounding error of the world. The gap that matters opens somewhere after the business case is signed.
What the numbers actually say about British caution
The convenient version of this argument is that UK boards are too cautious to commit. The survey does not support it. Physical AI — robotics with enough onboard intelligence to perceive a situation, decide, and act without a scripted path — is a stated priority for roughly two-thirds of large organisations everywhere, and the UK is squarely in that band.
Where the UK genuinely trails is in the top quartile. Japan reports 76%, South Korea 72%, China 70% and the United States 69%. Those four have made robotics an explicit plank of industrial policy, and it shows in how their executives rank it. But below them the field is tight: the UK’s 65% is level with Spain and ahead of Australia, the Netherlands, Germany, France, Singapore, India and Italy.
Critical Context: A 65% priority rating against a 66% global average is not evidence of an ambition deficit. Treat any strategy paper that opens with “the UK lags on robotics adoption” as an assertion requiring proof, not a premise.
The number that should worry a British board is a different one, and it is not a British number at all.
The real story: everyone is stuck in the same place
Capgemini found that 79% of organisations globally are engaging with physical AI in some form — 31% exploring, 20% running pilots or proofs of concept, 27% at limited or wider deployment. Only 4% describe themselves as already operating at scale.
That is the finding worth rearranging a strategy around. Nearly four in five organisations are doing something. One in twenty-five has finished. The distance between a working pilot and a production estate is where the entire market currently sits, and no country has solved it.
| Metric (global, unless stated) | Figure |
|---|---|
| Rank it a top automation priority, next 3–5 years | 66% (UK: 65%; Japan: 76%) |
| Engaging with physical AI in some form | 79% |
| Deploying or scaling solutions | 27% |
| Already operating at scale | 4% |
| Expect to be running at scale by 2031 | 65% |
| Top investment driver: labour shortages | 74% |
| Second driver: rising labour costs | 69% |
| Expect humanoids working alongside staff by 2030 | 45% |
Source: Capgemini Research Institute, Physical AI: Taking human-robot collaboration to the next level, April 2026. Fieldwork covered 1,678 respondents at director level or more senior, working for organisations turning over more than $1bn — or more than $500m in aerospace and defence, and in government and public services — spread across 16 countries and 15 industries.
The pattern will look familiar to anyone who read the ONS finding that UK firms have adopted AI widely but shallowly — use tripled to 35% of businesses whilst the average adopter ran just 1.6 tools. Breadth arrives quickly. Depth does not arrive at all unless someone is accountable for it.
Why do pilots stall between proof and production?
Physical AI does not fail at the demonstration. It fails at the second site.
A pilot runs in a controlled cell with a named engineer watching it, a suspended safety case and a tolerance for downtime that no operating line would accept. Production demands the opposite of all three. Capgemini’s respondents identified the blockage precisely: close to 80% find scaling difficult, and the leading reason given is technological and operational readiness rather than strategic doubt. Current systems do not clear the reliability thresholds that industrial and safety-critical settings require, dexterity remains limited, and data drawn from real physical interaction remains both scarce and costly to gather.
That last constraint is the one most business cases miss. A language model can be trained on text that already exists. A robot that must grip an unfamiliar component learns from interactions someone has to physically stage, instrument and repeat. The data does not accumulate whilst you deliberate — it accumulates only whilst you operate.
Hidden Cost: Every month a physical AI programme spends in pilot is a month of operational data not collected. The competitor who deployed a worse system earlier is training on real interactions whilst you are still refining a business case on simulated ones.
The British variable: capital that arrives in instalments
If the UK has a distinctive weakness here, it is not appetite but tenure. Dr Diane Berry, who leads engineering science at Capgemini and wrote the analysis that prompted this piece, points to a domestic habit of pausing or withdrawing funding when short-term pressures appear — a stop-start rhythm fundamentally at odds with a technology that compounds through iteration.
This is the argument that survives scrutiny, because it explains the observed data better than the caution thesis does. If UK executives rate the priority at 65%, they are not sceptical. They are funding on cycles of one to three years for an asset class whose returns arrive on a longer curve, and then reading the resulting flat return as evidence the technology was overhyped. The mechanism is self-confirming and it has nothing to do with courage.
Strategic Reality: Physical AI is closer to a capability build than a capital purchase. The organisation buys competence in integration, safety engineering and operational ownership — and competence decays faster than the equipment does when funding is interrupted.
Where the first deployments genuinely pay
The sequencing here inverts the pattern most executives learned from software. Agentic AI entered organisations through high-volume, low-stakes digital work: drafting, triage, summarising. Physical AI tends to enter at the opposite end, in environments where the alternative to a robot is a person accepting real risk.
Nuclear decommissioning is the clearest case. Handling material remotely inside a high-radiation area is not a productivity play; it is work where the machine is the only responsible option. The same logic runs through energy, utilities, heavy infrastructure, manufacturing and logistics, in which such systems are already cutting downtime and providing continuous cover against persistent labour shortages — the driver 74% of executives named as their primary reason to invest, ahead of labour costs at 69%.
That gives UK operators an unusual advantage. Britain has a large decommissioning estate, an ageing infrastructure base and acute shortages in skilled manual trades. The conditions that make physical AI valuable first are conditions the UK has in abundance.
| Stakeholder | What changes | What they need before it does |
|---|---|---|
| Operations leadership | Continuous cover in roles that cannot be filled | A safety case that survives the regulator, not just the pilot |
| Safety and risk | Autonomy inside a hazard envelope | Deterministic safety mechanisms independent of the AI stack |
| Finance | Multi-year capability spend, non-linear return | Milestones tied to operational readiness, not deployment counts |
| Engineering and skills | Integration and teleoperation competence in-house | Retention of the people who ran the pilot |
| Workforce | Collaboration, not displacement, in hazardous roles | Honest scoping of which tasks move and which do not |
What should a board actually commit to?
Three postures, matched to where an organisation genuinely is rather than where its strategy deck says it is.
If you have not started. Do not begin with a technology selection. Begin by listing the tasks in your estate that are dangerous, unfilled, or both, and rank them by what continuous operation would be worth. Physical AI has a legible business case in exactly those places and a weak one almost everywhere else. Our guidance for smaller manufacturers on where to start with AI applies with more force here, because the cost of a misplaced first project is measured in capital equipment rather than licence fees.
If you are piloting. Set the exit criteria for the pilot before it runs, and make them operational: uptime under real duty cycles, a safety case the regulator has seen, a named owner in operations rather than innovation. A pilot without a defined finish line becomes a permanent one. Capgemini’s 4%-at-scale figure is what a market full of unfinished pilots looks like from the outside.
If you are deploying. Protect the funding line through at least one downturn, and protect the people more carefully than the budget. The integration engineers and teleoperation specialists who made the first site work are the scarce asset; the survey identified lack of that expertise as a named barrier. Losing them in a cost round resets the programme further than cancelling a purchase order would.
Success Factor: The organisations that reach scale treat the second deployment as the real project and the first as the prototype. The second is where reusability, standardised safety cases and shared tooling either exist or do not.
Four problems that surface after the business case is approved
The safety case cannot be inherited from the model. A system that adapts in real time cannot be certified by demonstrating the behaviour you observed. Safety has to be enforced by deterministic mechanisms sitting outside the AI stack — mechanical limits, independent interlocks, bounded operating envelopes — so that assurance does not depend on the part of the system whose whole purpose is to be non-deterministic. Retrofitting this after a successful pilot is expensive and occasionally impossible.
Public acceptance is a UK-specific exposure. Some 64% of UK executives expect weak societal acceptance to prove a decisive obstacle to humanoid adoption, above the 62% global average and close to the most concerned markets. This is not a communications problem to be solved after deployment. Where physical AI touches public-facing settings — transport, healthcare, retail — the acceptance work runs in parallel with the engineering or the deployment stalls at consultation.
Humanoids are absorbing attention disproportionate to their near-term utility. Only 45% of organisations globally expect humanoids working alongside staff by 2030, and the named barriers are unambiguous: 72% cite technology immaturity, 63% upfront cost, 62% unclear return, 58% training difficulty. Meanwhile the growth over that same three-to-five-year horizon will come from autonomous mobile robots, industrial arms and cobots — established form factors with proven safety cases. A board that debates humanoids whilst its competitors deploy AMRs is losing time to the more photogenic option.
Capability leaks through the supplier boundary. It is entirely possible to run a successful physical AI deployment in which every piece of judgement lives in a systems integrator’s team. That works until the second site, the contract renewal, or the first genuine failure. The governance and institutional capability argument that applies to AI policy applies to industrial deployment too: an organisation that cannot interrogate its own system does not really own it.
⚠️ Warning: A programme that outsources integration, safety engineering and operational ownership simultaneously has purchased a demonstration, not a capability. The distinction becomes visible only when something goes wrong.
The strategic takeaway
Physical AI marks, in Pascal Brier’s phrase, “a shift from systems that describe the world to systems that can act within it”. The commercial consequence is that the usual software playbook — deploy broadly, iterate publicly, tolerate failure — does not transfer. Acting in the world means failures with physical consequences and an assurance burden that has to be engineered in from the beginning.
For UK organisations, three things follow.
The intent gap is largely imaginary, and continuing to frame the problem as British timidity misdirects the remedy. At 65% against a 66% global average, the conviction is present.
The execution gap is real and universal. With 79% engaging and 4% at scale, the differentiator over the next three years is not who starts but who finishes. That is a question about funding tenure, retained expertise and operational ownership — three things a UK board controls directly.
The UK’s structural advantages point at the highest-value entry point. Decommissioning, ageing infrastructure and skilled-trade shortages are precisely the conditions under which physical AI pays first. Britain’s stated caution has, almost accidentally, kept it from over-investing in humanoids whilst the form factor remains immature. That is worth something — but only if the deliberation converts into a funded, multi-year commitment rather than another cycle of well-reasoned delay.
Next steps for a board this quarter:
- Identify the three tasks in your estate where continuous operation is worth most and human presence is hardest to sustain.
- Ask what the exit criteria are for every physical AI pilot currently running. If any lacks operational criteria and a named owner outside innovation, it is not a pilot.
- Establish the funding horizon explicitly, in writing, and state what would justify interrupting it — before the first downturn makes the decision for you.
- Audit which of the required competences — integration, safety engineering, teleoperation, operational ownership — currently exist only in a supplier’s organisation.
Take Action: If your organisation’s physical AI conversation is still principally about humanoids, it is being held two years early. The near-term decisions concern mobile robots, arms and cobots in hazardous or unfillable roles — and whether the funding will still be there in year three.
Source citation and attribution
This analysis draws on “Physical AI in the UK: From cautious exploration to trusted adoption” by Capgemini’s engineering science leader, Dr Diane Berry, published by Innovation News Network on 24 June 2026, and also slated for its 26th quarterly edition.
All statistics cited above are taken from the primary source: the Capgemini Research Institute report Physical AI: Taking human-robot collaboration to the next level (April 2026) and its accompanying release. Fieldwork ran during January and February 2026 and covered 1,678 respondents at director level or more senior, at organisations turning over more than $1bn — or more than $500m in aerospace and defence, and in government and public services — spread across 16 countries and 15 industries.
Every figure quoted here has been read directly from that report and its country-level breakdowns rather than from secondary coverage, and national cuts are labelled as such wherever they appear.
Strategic interpretation, the pilot-to-scale framing and all recommendations are Resultsense analysis. Read more of our coverage at Resultsense Insights, or get in touch to discuss what this means for your organisation.