How Should Councils Govern AI in Social Care?

The short answer. Councils should govern AI in social care through five layers that sit above the practitioner: a written AI policy, a procurement test, a workforce literacy programme, a practitioner-led use-case register, and a published assurance routine. Each layer assumes the next one exists. Together they turn AI from a risk surface into an accountable tool that improves outcomes without displacing professional judgement.

The question is being asked everywhere this month. At the AI Summit. In the Local Government Chronicle. In Social Work England's guidance. In every commissioning meeting where someone has been asked to sign off on a tool nobody at the table has been trained to evaluate. The answers, where they exist at all, tend to be either too abstract to act on or too narrow to scale. Here is what good council-level AI governance actually looks like in social care, and where to start this week.

What governance means at the council level (and what it does not)

Practitioner-level discipline is one thing. A clear prompt, a Voice-Evidence-Reasoning-Accountability-Human check, a refusal to paste identifiable detail into a public chatbot. That work matters. It is not, however, governance.

Governance is the institutional layer above the practitioner. It is the answer the council gives to four questions, in writing, that hold up under audit, public scrutiny, and a regulator's challenge:

1. What AI is allowed in our social care work, and what is not?
2. How does a tool get from being a vendor's pitch deck to being in a practitioner's hands?
3. Who is accountable when it goes well, and who is accountable when it goes wrong?
4. How will we know, in twelve months, whether this is still safe?

If a council cannot give crisp answers to those four questions, it does not have AI governance. It has AI exposure.

The five layers a council needs

1. A written AI policy practitioners can actually read

The policy is the document a senior practitioner could read in fifteen minutes and use to make a decision the same afternoon. It names the scope (which teams, which tools, which categories of decision), the prohibitions (identifiable personal data into unapproved tools, AI-generated text used in records without human authorship, decisions affecting service users delegated to a system without a named human decision-maker), the accountability model (which named role signs off on use), the council's public-facing position, and the review cycle. The test of a good policy is not legal sign-off. It is whether a Newly Qualified Social Worker can use it without translation.

2. A procurement test that filters before a tool gets near the workforce

An AI procurement test is a structured set of questions a council applies before a vendor's tool can be bought, piloted, or rolled out. It is not a checklist of features. It covers data flows, training data, model behaviour, professional accountability, accessibility, environmental footprint, exit and portability, and supplier governance. Most importantly, it is failed by silence: if a vendor cannot answer a question, that is the answer. Pilot anyway and the council inherits the unanswered question as its own risk.

3. A workforce literacy programme that is more than compliance training

Most "AI training" in social care right now is a forty-minute video and a multiple-choice quiz. That is not literacy. Workforce literacy means practitioners can describe what an AI tool is, what it is not, how it fails, what their accountability looks like when they use it, and how to challenge a poor implementation. It is also, importantly, training for the leaders deciding whether to buy the tool in the first place. A council where only frontline practitioners are AI-literate has its risk inverted.

4. A practitioner-led use-case register

You cannot govern what you do not know is being used. A use-case register is a living document, owned by a named practitioner lead, that records every AI tool currently in use across the council's social care workforce, including the unofficial ones. The first audit of any use-case register surprises every council that does it. There is always more AI in the building than the policy assumes.

5. A published assurance routine

Assurance is the public, repeating answer to the question "is this still safe?". It is not a one-off impact assessment filed before launch and then never opened again. At a local authority, assurance means an annual or biannual statement, published openly, covering what AI is in use, how it has been audited, what has gone wrong and how it was handled, what the workforce literacy data shows, and what the council is changing as a result. Assurance keeps the other four layers honest.

What good looks like (and what we see going wrong)

The pattern is consistent across the councils I have worked with.

Where this fits

Governance is what makes pace possible. The councils with the strongest governance are also the ones moving fastest on the AI work that actually improves outcomes, because they are not having to pause every three months to clean up an unmanaged tool. Without governance, pace is what creates the next service-user harm.

Council-level governance sits on top of national policy, it does not replace it. Social Work England's commissioned research, the Ada Lovelace Institute's findings from seventeen councils, and the Department of Health and Social Care's emerging position are the floor. Council-level governance is the layer that actually keeps people safe in your services, on your watch.

Where to start this week

If you lead an adults or children's services directorate, three concrete actions take less than a fortnight to start.

• Commission a use-case register before you commission anything else. Ask a senior practitioner to map every AI tool currently being used by your social care workforce, official and unofficial. You will not be able to write a credible policy without this baseline. Most councils discover at least one tool nobody in leadership knew about.
• Put a practitioner on the procurement test. If you have an AI tool in active procurement, pause it long enough to add a named social worker to the evaluation panel. They do not need to be technical. They need to be the person who will live with the consequences of the buy.
• Publish, even imperfectly. The first public assurance statement you write will not be perfect. The act of writing it for a public audience is what surfaces what governance you actually have, and what you do not. Better an imperfect statement that someone has to defend than a perfect document nobody outside the council ever reads.

None of this requires a bigger digital team. It requires the right people in the room before the contract is signed, not afterwards.

---

---