AI Leaves the Screen: Governing the Convergence of Agentic AI and Robotics
Gartner predicts that 15% of all work decisions will be taken autonomously by AI between now and 2028, rising from 0% in 2024, and that 33% of enterprise applications will include agentic capabilities from the same period. According to Deloitte's 2026 survey of 3,237 leaders in 24 countries, 74% of enterprises expect to deploy agentic AI in the next two years and 58% are already using physical AI in production. 21% have a mature governance model for autonomous agents. The next wave of adoption risk is going to fall on warehouse floors, payment rails and hospital wards for the UK. Governance for recommending AI will not govern acting AI.
Gartner predicts that between now and 2028, 15% of all decisions made every day will be made independently by AI, rising from 0% in 2024, and that 33% of enterprise applications will incorporate agentic features during the same timeframe. According to a study by Deloitte of 3,237 leaders from 24 countries, 74% of enterprises are expected to use agentic AI within two years, while 58% have already implemented physical AI in production. 21% have a mature governance model for autonomous agents. In the UK, the next wave of adoption risk will come on the floor of a warehouse, the payment rail, the ward of a hospital. Governance for recommending AI will not govern acting AI.
AI is already making its way off the screen.
The convergence is not a forecast. In 2024, the International Federation of Robotics (IFR) counted 542,076 industrial robots installed worldwide, bringing the total operational stock to 4.66 million. UK installations dropped 35% to 2,500 units, bringing Britain down to 19th place worldwide, following the super-deduction peak, and a clear indication that it is the technology that is not the limitation, but deployment. The Government's AI Opportunities Action Plan has attracted £14 billion in private sector investment pledges and 75% of FCA-regulated companies are using AI in their production.
There are examples in the UK of execution at scale. Ocado has over 3,000 robots in each fulfilment centre, 65 orders are completed per second with 99.5% accuracy and 50,000 orders are assembled in 5 minutes on a five millimetre clearance grid. The research base is anchored by the National Robotarium at Heriot-Watt. Internationally, Amazon has deployed more than a million warehouse robots in more than 300 centres, Waymo has logged 170 million miles driven in the first quarter as a rider only, with 91% fewer serious injury crashes than human benchmarks, and Intuitive Surgical deployed 431 da Vinci systems in one quarter. AI is already operational at an industrial level. The question is, will the UK's operating model be ready for it?
The governance velocity gap
The desire to adopt is greater than governance capacity and the gap is measurable. In fact, 68% of production agents stop after 10 steps or less without needing human intervention, according to LangChain's State of Agent Engineering. According to EY and the AIUC-1 Consortium, 80% of businesses have identified risky agent behaviours, while 17% monitor agent-to-agent interactions on a regular basis and 38% end-to-end monitor AI traffic. The compounding maths isn't kind: if a company has 20 steps in a workflow, and each one has a 95% chance of success, the end-to-end success rate is just 36%, and that's why 64% of businesses with revenues exceeding £1 billion reported losses exceeding £1 million due to AI failures in 2025. The OECD's AI Incidents Monitor now has over 14,000 entries and the framework now explicitly acknowledges the harm of two or more AI systems interacting.
The UK has a strong approach when it comes to AI that recommends, with its principles-based, sectoral approach. It is less well-suited to AI that cuts across sectoral boundaries. Orchestration agents and autonomous mobile robots and humanoids traverse the ICO on data, the HSE on worker safety, the MHRA where devices are health-adjacent, the FCA and PRA where execution is on financial rails, and Ofcom on infrastructure on a single warehouse floor. The seam is not the responsibility of any one regulator. The trend I see in fast automation and digitisation activities in the national operational networks is always the same: governance is always behind the execution, and it is always an incident before it is a policy. It's no longer the frontier of AI that decides. This is what AI should be allowed to do.
The five pillars of AI governance are:
Decision boundaries. Each autonomous system must have a formal set of decisions it can make, decisions it can act on, and decisions it must escalate. The management-system anchor is contained in
;
codifies human oversight; and
outlines the redress route. The failure of
McDonald's IBM voice-order system
is a boundary failure, not a model failure, as a customer was charged for 18,000 cups of water and ended the call.
Intervention and override points for humans. Overrides are not a backup, they are a built-in part of throughput. After discovering that hybrid performed better at scale than fully-automated,
brought humans back in for emotional conflicts and complex transactions. The
Data (Use and Access) Act 2025
and the
new guidance from the ICO on automated decision-making
require override design as a compliance requirement, not a design choice.
Handling of exceptions and failure recovery. Each successive integration is a failure surface.
need their agents to interact with eight or more data sources, so circuit breakers, retry logic, rollback plans and acceptable human-intervention rates need to be established before deployment, not after.
indicates that supervisors will be looking for automated monitoring, not paper compliance.
Collaborative responsibility for software, operational and physical systems. Clearly, the
OECD's interacting-systems framing
, and the
Bank of England Financial Policy Committee's April 2026
explicit request for liability regimes designed for a single decision-maker per harm to be re-engineered for multi-agent execution, are the strongest indicators of this.
Continuous outcome validation.
Waymo's Safety Impact Data Hub
reports crash rates per million rider-only miles, relative to dynamically adjusted human benchmarks. The standard: post market monitoring, in public, in production. UK deployments need to be as disciplined as
EU AI Act post-market monitoring
and
ISO/IEC 42001 continuous evaluation
.
The next 24 months
The UK's principles-based approach is a good one, but that doesn't apply to robots in warehouses, agents on payment rails, or humanoids in hospitals. The DSIT Blueprint, the FCA Mills Review and the Bank of England's FPC agentic-AI ask all draw the same conclusion. The operating model is the next 24 months of policy work. Before agentic AI scales, it is techUK members, regulators and enterprises who should build it together.
Author
Sources referenced
● HumansAreObsolete.com — Gartner predicts 15% of daily work decisions made autonomously by AI by 2028: Agentic AI to control 33% of enterprise software. Available at: https://humansareobsolete.com/articles/gartner-agentic-ai-15-percent-autonomous-decisions-2028-prediction-december-15-2025
● Deloitte — Agentic AI strategy. Available at: https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/agentic-ai-strategy.html
● Robotnik — Industrial robotics in 2025: trends, figures and global outlook. Available at: https://robotnik.eu/industrial-robotics-in-2025-trends-figures-and-global-outlook/
● Drives & Controls — UK industrial robot installations plummeted by 35% in 2024. Available at: https://drivesncontrols.com/uk-industrial-robot-installations-plummeted-by-35-in-2024/
● GLACIS — UK AI regulation: the pro-innovation approach. Available at: https://www.glacis.io/guide-uk-ai-regulation
● TheRoboWire — Warehouse robots 2026: How Amazon and others automate fulfillment. Available at: https://therobowire.com/warehouse-robots-automation-amazon-2026/
● techUK — techUK's 2026 Tech & Innovation focus week. Available at: https://www.techuk.org/resource/welcome-to-techuk-s-2026-tech-innovation-focus-week.html
● HumansAreObsolete.com — Amazon's million robot army expands as physical AI becomes new workforce layer. Available at: https://humansareobsolete.com/articles/amazon-physical-ai-warehouse-expansion-gxo-digit-robots-december-2025
● Waymo — Waymo Safety Impact Data Hub: Release Notes. Available at: https://storage.googleapis.com/waymo-uploads/files/documents/safety/safety-impact-data/Waymo_Safety_Impact_Data_Hub_Release_Notes_20250916.pdf
● Analytics Insight — Best AI healthcare stocks and ETFs to buy in 2026. Available at: https://www.analyticsinsight.net/stocks/best-ai-healthcare-stocks-and-etfs-to-buy-in-2026
● AgentMarketCap — Gartner's 40% warning: Why most agentic AI projects never reach production. Available at: https://agentmarketcap.ai/blog/2026/04/07/gartner-40-percent-agentic-ai-failure-rate-enterprise-readiness
● GLOBAÏA — Already happening: The AI incident record. Available at: https://globaia.org/ai-risks/incidents/
● Beri — Agentic AI ROI: 12 cases show 171% returns, $83M saved. Available at: https://www.beri.net/article/agentic-ai-roi-171-percent-returns-enterprise-case-studies
