Only 7% of civil servants are traditional policymakers, and 55% work in operational delivery, Cabinet Secretary Dame Antonia Romeo told a London School of Economics (LSE) audience on 10 November, giving the yearly lecture that goes with her post. So when she says AI will rewrite every domain the service works in, “from policymaking to direct service delivery”, her own figures put most civil servants in operational delivery, which she placed in prisons, courts, tax offices, job centres and at the borders, rather than in policy units. Our news report on the lecture covers what she announced. This analysis asks a narrower question: what does an AI-enabled state demand of Whitehall, and of the companies that build and run its technology? The lecture answers the first half more fully than the second.
Strategic Insight: Romeo lists AI first among the five imperatives she says make up “the burning platform for change”. When she turns to AI, she also names the cost: the move to an AI-enabled state “could expand our domestic cyber vulnerability”. In our view, for suppliers that sentence matters more than any announcement in the lecture.
Why does the Cabinet Secretary put AI first?
Romeo’s case is that the old model, in which the state runs siloed services for the public, has stopped being enough. Citizens now arrive with their own technology. “The surge of automated, high-volume citizen demand is hitting operational desks today”, she said, and she expects demand to keep rising from what she called “an AI-literate society with low barriers to engaging with government”. In our reading, AI is reaching the counter before departments have rebuilt what sits behind it.
The real story is the workflow, not the tool
For anyone selling technology to government, one warning stands out. Automating single tasks without rethinking the whole workflow, “or indeed the whole operating model”, Romeo said, “will not fully capture the underlying gains from technology”. The examples she gave, from machine learning at HMRC to AI analysis of planning applications at the Ministry of Housing, Communities and Local Government (MHCLG), are already in our news report. She offered them as proof the technology is in use now, then argued that gains task by task are not enough.
Her proposed vehicle for the change is small “Frontier Teams” inside departments, armed with “an explicit licence to operate differently”. She wants them able to hire specialist talent quickly and to redesign processes that have outlived their purpose. That is a structural answer, not a procurement one, and a first sign that the lecture says more about Whitehall’s half of the bargain than about its suppliers’.
| Measure | Figure, as stated in the lecture | Strategic implication |
|---|---|---|
| Traditional policymakers | 7% of civil servants | AI in policy drafting directly affects a small share of staff |
| Operational delivery | 55% of civil servants | The frontline is where Romeo says “the weight of the Service” now sits |
| Responses to the Future Civil Service Portal | 8,000 | Staff told her they want “to use AI and technology to replace bureaucracy” |
| Size of the service | More than half a million people | Romeo calls it “bigger than any single private sector footprint in this country” |
Strategic Reality: Romeo’s stated vision is a civil service that is “a global leader in safe, AI-enabled state transformation”. The qualifier matters. In the same lecture she said deploying AI ethically and transparently, with guardrails, must be the “overriding principle”.
Who answers for a decision an AI tool helped shape?
The most useful passage for this question comes early in the lecture, in Romeo’s history of the service. She cited the Court of Appeal ruling of 1943 that set out the Carltona principle: “there is no distinction between the decisions of officials and Ministers”. As she put it, “Civil servants have no separate legal personality from the Minister.” She said it is still the core principle the service works under.
Later she set a test for AI: “we must remain capable of explaining how and why decisions are made to retain public trust”. And the civil service review will ask “What AI and devolution by default mean for Civil Service boundaries, and accountability?”
In our reading, the principle as she describes it has two parties, officials and ministers. A model built and hosted by a supplier is neither. When an AI tool shapes a casework decision, responsibility still runs to the minister, but the explanation Romeo wants may depend on information only the supplier holds. That is why the review’s accountability question matters to firms as much as to departments.
Critical Context: Explaining the role algorithmic tools play in decisions is already a publishing duty for much of government. The Algorithmic Transparency Recording Standard (ATRS) is “mandatory for all government departments”, and for arm’s-length bodies that run frontline or public services or deal directly with the public, subject to a scope and exemptions policy from December 2024.
What the lecture leaves to the review
Romeo was clear that her current reforms work within today’s structures, and that the bigger questions belong to the civil service review that began in July and is due to report before next summer. It will ask what AI and devolution mean for the service’s boundaries and for accountability, whether today’s departments still suit work that cuts across them, and how capability should be built nationally, regionally and locally. Coordinating a government-wide AI strategy falls to the OPMC, the new office that brings No 10 and Cabinet Office teams together. Our earlier analysis of permanent secretaries’ priorities for an AI operating model sets out the four priorities senior officials gave for government AI.
Where does the cyber warning land?
Romeo named the risks as economic disruption, misinformation and “loss of agentic control”, plus the cyber exposure an AI-enabled state could bring. The government has been working on that exposure from the other end.
A week before the lecture, the National Cyber Security Centre (NCSC) published its State of the Cyber Nation report. It judges it “highly likely” that hostile actors’ use of AI will make their operations larger and more efficient within two years. In remarks the NCSC released for the launch, its chief executive, Dr Richard Horne, argued that the problem is old weaknesses, not new attacks: “vulnerabilities that have been tolerated for far too long can now be discovered and exploited at a scale that was once unimaginable”. The era of “tech on the cheap”, in his phrase, is over. Our news report on the NCSC warning has the detail.
The government’s legislative response is the Cyber Security and Resilience (Network and Information Systems) Bill. The government’s summary says it would bring managed service providers of medium or large size into the regime, with the Information Commission as their regulator, and would let regulators name the most important firms supplying essential and digital services as critical suppliers. The case for the second power is a 2024 cyber attack on an NHS pathology supplier, which the summary says left more than 11,000 appointments and procedures postponed and contributed to a patient’s death.
A separate factsheet, updated on 3 November, deals with AI agents. Ministers plan to consult on measures for incidents where operators are “unable to adequately monitor or control AI agents”, including requirements to preserve evidence of what agents did. Its hypothetical example has an operator keeping records of agents’ actions and their “chain of thought” logs. Ministers would also gain powers to have government experts review such incidents. The factsheet is explicit that these measures are “not intended to establish a comprehensive framework for AI governance”. Our report on the agentic clause has more on the clause.
⚠️ Warning: Romeo’s “loss of agentic control” is the risk the government’s factsheet of the week before calls “agentic loss of control”. In the lecture it sits among advanced AI’s general risks, one sentence before her warning about the state’s own cyber exposure. On the government’s summary, though, the Bill’s duties reach essential services, data centres, designated critical suppliers and managed service providers. In our reading, the exposure of the state’s own AI systems is a separate question.
| Group | What changes for them | What is still unknown |
|---|---|---|
| Civil servants | Romeo says jobs will change and some tasks will be automated, but that human creativity, interaction and decision making stay | Which roles, and how the impact on jobs will be managed in practice |
| Departments and arm’s-length bodies using algorithmic tools | Already required to publish ATRS records, now with the Cabinet Secretary’s explainability test on top | How a whole redesigned workflow, rather than a single tool, will be recorded |
| Managed service providers, medium-sized and up | Would be brought into the regime, regulated by the Information Commission, and required to warn customers an incident has probably affected | When those provisions commence, which follows Royal Assent in phases |
| Firms supplying essential and digital services | Regulators would gain a power to designate the most important of them as critical suppliers | Which firms are designated |
| Operators running AI agents in essential services | A consultation on loss-of-control duties, including evidence preservation | What counts as “loss of control”; the factsheet says industry and oversight bodies will help settle it |
What this means for suppliers to government
The lecture never uses the words supplier or procurement. Read beside the government’s own rules, though, its demands still reach the firms behind Whitehall’s technology. Romeo wants whole workflows redesigned, decisions that can be explained, and an AI transition that does not widen cyber exposure. A point tool sold into a department faces all three tests: it will sit inside a redesigned process, the department using it may need to publish an ATRS record, and a supplier that is a medium or large managed service provider would come into the cyber regime in its own right.
The open challenges
Frontier models with no sourcing story
Romeo wants civil servants equipped with “the most relevant frontier models for their work”, and she wants Frontier Teams to hire specialists quickly. The lecture says nothing about how those models will be procured, or how they will be assured. In our view, if the OPMC’s AI strategy does not answer that, departments will answer it one contract at a time.
Accountability with a third party in the loop
The Carltona principle, as Romeo describes it, ties official decisions to ministers. Her own test is that the state must be able to explain how decisions are made. In our reading, where the reasoning sits inside a supplier’s model, those two commitments meet a contract, and the review’s question about accountability is where that gap will either be closed or left to individual departments.
Exposure moving faster than the rules
The NCSC’s horizon is two years. The Bill has not yet reached Royal Assent: the Lords gave it a third reading on 2 November, and no date has yet been set for MPs to consider the Lords’ amendments. Its measures then take effect in phases, and the agentic measures need a consultation and secondary legislation that Parliament must approve first. In our view, the cyber exposure Romeo warns of could arrive before the duties meant to contain it.
Reality Check: Civil servants told Romeo through her portal that they want new tools to sweep away bureaucracy. Wanting the tools is not the same as running them safely. The NCSC expects new technologies to expose “vulnerabilities, failings and underinvestment at an unprecedented speed and scale”.
What to watch
Four things would show whether Whitehall is building the AI-enabled state Romeo describes, or adding tools at the edges:
- The civil service review’s answer on accountability. It reports before next summer. An answer that says who is responsible when an AI tool shapes a decision, and what the state must be able to obtain from suppliers to explain it, would show the explainability test turning into practice. Silence would leave it to departments and contracts.
- The agentic AI consultation. The factsheet says the measures’ scope, and what counts as “loss of control”, will be settled with industry and oversight bodies. Definitions tied to monitoring and evidence logs would give suppliers a concrete standard to build to.
- ATRS records for the tools Romeo cited. Published records for the HMRC and MHCLG systems she held up as examples would show her transparency principle applied to her own showcase.
- The refreshed Civil Service Code. The government will publish it in the spring. Whether its values say anything about how civil servants use AI will show how far the lecture’s principle reaches into everyday conduct.
Our earlier look at the Tony Blair Institute’s blueprint argued that its AI-enabled state assumes a single centre Britain does not have. Romeo’s lecture shows the head of the civil service treating AI as a question of how the state is organised, not which tools it buys. The harder part is the one she left open: what the state needs from the companies whose models and systems it will depend on. Until the review answers that, the AI-enabled state is a plan for Whitehall, with its suppliers still waiting for their part.
Source: Rewiring the state and the future civil service, Dame Antonia Romeo’s yearly lecture as Cabinet Secretary, given at the London School of Economics (Cabinet Office, 10 November 2026). Also drawn on: the NCSC’s State of the Cyber Nation announcement (3 November 2026); the government’s factsheets on the Cyber Security and Resilience Bill and its agentic AI resilience measures; and the Algorithmic Transparency Recording Standard Hub.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the civil service review says what the state must be able to demand of the suppliers behind its AI. Read more analysis at Insights, or get in touch.