When European public bodies consider purchasing agentic ai for government, they encounter a technology category that barely existed in procurement frameworks two years ago. The vendors are confident; the definitions are elastic; and the distance between a brochure promise and a working system can be considerable.
Agentic AI is not simply generative AI with better marketing. The distinction matters for any public authority writing a tender, and it matters even more for regulators trying to work out which rules apply.
What agentic actually means
The UK government's AI Playbook offers a definition: "Agentic AI refers to AI systems composed of agents that can behave and interact autonomously in order to achieve their objectives." The phrasing is precise. Generative AI produces content from patterns. Agentic systems decide what to do, how to do it, and when. They pursue objectives rather than respond to prompts. An agentic system processing a form does not simply record the data. It checks for missing details, flags potential issues, routes the case to the correct department, and follows up without step-by-step human instruction.
Nadav Mordechai, Product and Strategy Director at Elsewhen, argues that the real shift lies in organisational workflows rather than individual productivity. His report describes the move as going from "adding AI into a system" to "thinking of AI as the system." The change is architectural, not incremental, and it requires procurement teams to evaluate something fundamentally different from the software contracts they are used to.
Singapore's pilots and what they show
Singapore, with roughly a tenth of the UK's population, has published an Agentic AI Primer designed to move beyond buzzwords. One pilot described in that document built a multi-agent system that mimicked data scientists working together: delegating tasks, calling in specialist agents, querying graph databases, and producing contextual insights from customer relationship management data. The primer demonstrates that agentic systems can emulate the way civil servants coordinate, delegate, and problem-solve.
But a pilot in a controlled environment is not the same as a deployed system operating across multiple authorities with varying data quality, legacy infrastructure, and legal obligations. As NewsBriefing has previously reported, the gap between demonstration and deployment in European public services is where most technology projects stumble.
Productivity claims need scrutiny
The Elsewhen report states that "the application of agentic AI in government could lead to productivity improvements of at least 20% to 30%." That figure should be treated with caution. It appears without a cited methodology, sample size, or independent audit. It is a vendor's estimate, not a verified result.
More concrete, though still anecdotal, is the figure from Prime Minister Keir Starmer's speech: a planning system processing a hundred records per day where the previous average was five. Even here, context matters. A demonstration under controlled conditions does not guarantee equivalent performance at scale across different local authorities with inconsistent data.
Regulators start to catch up
The Information Commissioner's Office, the UK's data protection regulator, has published its own Tech Futures report on agentic AI, exploring four scenarios for how the technology might develop over the next two to five years. Its central finding is that the specific design and architecture of agentic systems determine how data protection law applies and how individuals exercise their rights.
Poorly implemented systems, the ICO warns, increase the risk of data protection harms. The choices that matter include what data and tools a system can access, and what governance and control measures surround it. These are precisely the details that vendor presentations tend to gloss over.
The UK government's publication on agentic AI and consumers, while focused on the consumer market, raises questions that public-sector procurers should also ask. What decisions does the system make autonomously? What data does it access? How are errors detected and corrected? Who is accountable when the system causes harm?
The European compliance challenge
For European buyers, the regulatory environment is more complex than in the UK. The EU AI Act classifies AI systems by risk level, and any agentic system processing personal data in public services will face scrutiny under both the AI Act and the General Data Protection Regulation. A system that works in a Singaporean pilot cannot simply be transplanted to a European context without revisiting its compliance architecture from the ground up.
The European Commission's digital strategy work touches on AI in public services, but specific guidance on agentic systems remains thin. As earlier coverage of European digital government has noted, member states are at different stages of digital maturity. The Nordics have invested in digital government infrastructure for years; some southern and eastern members are still building basic digital identity systems.
Why procurement needs new criteria
European public bodies, accustomed to procurement rules that prioritise transparency and equal treatment, may find that evaluating agentic systems requires new criteria. Traditional tenders assess whether a product meets a specification. Agentic systems, by definition, adapt their behaviour to achieve objectives. The specification is not the system; the system's decision-making process is.
The Elsewhen report proposes five principles for identifying where agentic AI can add genuine value in public services, focusing on parts of the system that slow things down, burn staff time, or distract civil servants from higher-value work. The underlying point is sound: automating friction is more useful than automating for novelty. But principles are not the same as procurement criteria, and no shared European framework yet translates them into contract specifications.
More from News Briefing: Cisco firewall management flaw with maximum severity still exploited in Germany · EU Ombudsman probes AI envoy appointment over Siemens conflict concerns.
People mentioned
-
Nadav Mordechai
Organisations
Information Commissioner's Office · Elsewhen · UK Government · European Commission
