Over the last 12 months, the public sector has been swept up in the first wave of AI adoption. Most organisations have introduced some form of generic tool, whether it’s Microsoft Co-pilot, ChatGPT Enterprise, or Gemini.
These tools are genuinely useful. They help staff draft emails, summarise sprawling policy documents and spot patterns in meeting notes. For example and according to the LGA 2025 update, 83% of councils are exploring or using generative AI.
They deliver a clear ‘productivity dividend’, helping overworked teams clear their inboxes faster. They also create the comforting impression that the organisation has ‘started with AI’.
But this creates a dangerous misconception. Rolling out a generic Large Language Model (LLM) is not digital transformation. It is simply the starting point.
The real breakthrough – the shift that will actually reduce backlogs and improve citizen outcomes, comes when organisations stop relying solely on general-purpose tools and begin building AI Agents.
These are not chatbots. They are specialised systems powered by your own sovereign data, the records, workflows, policy rules and statutory processes that define how your services actually operate.
This shift is what will distinguish public bodies that simply use AI from those that deliver faster, safer and fully auditable public services.
Download Feeding the Beast. Managing Data for Public Sector AI
The ‘car v van’ problem
There is a simple way to illustrate the gap between what you have (Copilot) and what you need (Agents).
As Catapult’s Principal Engineer Craig Cook puts it: “Using a generic LLM for public-service automation is like saying: ‘I already have a car, so why would I ever need a van’?
A car is excellent for personal travel. It’s fast, comfortable, and gets you from A to B. But you don’t use it to move furniture. And you certainly don’t judge a Ferrari by its inability to carry a sofa.
Generic LLMs are the ‘cars’ of the AI world. They are great at general tasks, writing text, summarising notes and answering broad questions based on the internet’s knowledge.
But they have no idea how your organisation actually works.
- They don’t understand your local planning regulations (and the specific exceptions to them).
- They don’t know your benefits eligibility criteria or how your appeals process really runs in practice.
- They haven’t seen your historical casework, your policy constraints, your risk thresholds, or your audit requirements.
In other words, they don’t know your service reality. And if they don’t know your reality, they cannot automate anything that matters.
What public-sector AI agents can do (that chatbots can’t)
To move beyond simple productivity, we need AI Agents.
An Agent is a purpose-built system designed to perform a structured, verifiable and repeatable task. Unlike a generic chatbot, it does not rely on the open internet for its logic. It relies on your data.
For example, a well-designed public sector AI agent can:
- Pre-validate casework before it ever reaches a human officer, checking for missing evidence.
- Check eligibility instantly by cross-referencing an application against data already held in legacy systems.
- Triage benefit applications, routing simple approvals to one queue and complex cases to senior staff.
- Accelerate planning workflows by matching new submissions against 20 years of historical precedents.
- Structure decision summaries, pulling together all relevant policy clauses for a human adjudicator to review.
None of this is possible with a generic LLM alone. It requires Sovereign Data – the internal context that only your organisation possesses.
Generic tools give you efficiency but agents give you outcomes.
Why ‘Sovereign Data’ is your superpower
Every council, agency and regulator sits on a massive, untapped asset – their operational memory.
You hold decades of case histories, past planning decisions, benefits rulings and citizen interactions. You have detailed internal policy documents, standard operating procedures (SOPs) and the ‘tacit knowledge’ of how edge cases are handled.
This is your Sovereign Data. It is the asset that shapes every decision you make.
Generic LLMs don’t have access to it. Even if they did, they wouldn’t understand the nuance. Without serious engineering, pasting this data into a public chatbot is a security risk.
AI Agents are designed for exactly this. They are trained (or ‘grounded’) specifically on the rules and records that make your organisation unique. That’s why the real value in public-sector AI won’t come from Silicon Valley’s latest generic model.
It will come from systems built on your domain knowledge.
The AI value ladder. Where are you?
There is a simple maturity model for AI in government. Most bodies are currently stalled at Level 1.
Level 1. Generic LLM tools
- Examples. Copilot, ChatGPT, Gemini.
- Function. Writing, summarising, coding assistance.
- Limitation. No operational understanding and cannot be trusted with decisions.
Level 2. Organisation-specific AI Agents
- Examples. A ‘Planning Application Validator’ or ‘Benefits Triage Bot’.
- Function. Executing specific workflows using internal data.
- Limitation. Supports decisions, reduces backlog, automates drudgery.
Level 3. Custom Sovereign Models
- Examples. A fine-tuned model for clinical coding or legal adjudication support.
- Function. Complex reasoning with full auditability.
- Limitation. Transformational capability.
The shift is not from one vendor to another, it is from generic intelligence to organisational intelligence.
High Stakes. The ‘Data-to-Decision’ chain
We are entering the era of Agentic AI, where systems don’t just chat, they act.
Imagine an AI that checks a planning submission for completeness, applies local policy rules, flags inconsistencies and prepares a recommendation for an officer. The efficiency gains are huge, but so are the risks.
In the public sector, the computer said so is not a legal defence.
If an AI agent influences a decision about a citizen’s life, whether it’s a parking fine, a housing application, or a benefits claim, you must be able to prove exactly how that result was reached.
You need a fully documented Data-to-Decision Chain.
- Input. Exactly what data did the model see?
- Process. What logic or policy rules did it apply?
- Outcome. Why did it reach this specific conclusion?
- Oversight. Who was the human in the loop?
The UK’s Algorithmic Transparency Recording Standard (ATRS) is a good starting point, but operational automation requires robust engineering underneath. Generic LLMs cannot offer this level of granular auditability. AI Agents can, but only if they are built on clean, governed and trustworthy data foundations.
What you need to make the leap
Moving from Co-pilot to real automation is not just a software purchase. It is an engineering challenge. To make the leap, public bodies need:
- Clean, Structured Data. Agents need reliable inputs. They cannot learn from buried PDFs, untagged SharePoint files, or fragmented legacy databases.
- Defined Decision Models. Your actual decision pathways (not just the theoretical ones) must be mapped and made machine-readable.
- Governance Guardrails. You need lineage, versioning, and strict escalation rules for when the AI is unsure.
- Human-in-the-Loop Design. The AI should never be the final arbiter on high-stakes cases.
How Catapult helps you build
At Catapult, we don’t believe in ‘AI hype’. We believe in engineering.
We work with councils, regulators and government departments to turn their internal sovereign data into structured, operational knowledge. We work together with you to:
- Unlock data from legacy systems and SharePoint silos.
- Clean and structure that data so an Agent can read it.
- Design transparent, auditable decision pathways.
- Build proof-of-concept Agents in 4–6 weeks, not months or years.
If you are ready to move beyond the productivity illusion of generic tools and start building AI that actually runs your services, we can help.
For a practical, detailed explanation of how to build AI Agents, govern your data and move up the value ladder, download our exclusive public sector guide.
