The unassigned
workstream.
AI, evidence, and the councils being abolished.

Research note · Enterprise AI Lab · Figures as at August 2026 · No registration

A complaint in 2029

In 2029, a resident complains to the Local Government and Social Care Ombudsman about a decision taken in 2027.

The officer who handled it has moved on. That is normal. The council that employed them no longer exists. That is new.

The successor authority picks up the complaint, because it picks up the predecessor’s liabilities. To answer it, someone has to establish what informed the decision. If an AI tool drafted the summary the officer relied on, the questions are specific: which model, which version, which source documents did it retrieve, what did the officer actually review, and where is that record now?

None of that is hypothetical. The Local Government (Structural Changes) (Transfer of Functions, Property, Rights and Liabilities) Regulations 2008 transfer liabilities from predecessor to successor bodies. The LGA’s guidance on transitional governance is explicit that new councils should set out how they will deal with complaints about the actions of abolished councils — with the relevant new council having appropriate access to the records so it can deal with the complaint. And the Ombudsman has already been critical of new councils that failed to put those arrangements in place, causing confusion and drift.

That records requirement is the whole of it. For most systems being consolidated, “appropriate access to the records” is a migration task with a known answer. For AI-assisted work, it is not yet clear what the record even is.

For most authorities entering reorganisation, there is currently no process that would produce those answers. Not because anyone has been careless — because the systems being consolidated were never asked to keep that evidence, and nobody has been given the job of asking.

The guidance has an AI-shaped hole

Local Digital publishes the sector’s most useful guidance on digital priorities across reorganisation. Six phases, from pre-decision collaboration through shadow authority, the final ninety days before vesting, the first hundred days after, and on into years two and three. It is practical, well-made and widely used.

At the time of writing, it does not address AI at any phase. Not AI tools, not algorithmic transparency, not AI audit records.

This is not a criticism of Local Digital. Reorganisation guidance was built around the systems that dominate a council’s estate — case management, finance, HR, identity, networks. AI arrived in local government’s operational layer very recently and very unevenly, and the guidance reflects a landscape that was accurate when it was written.

But the consequence is concrete. A programme director working diligently through a comprehensive checklist will not be prompted to ask the AI questions, because the checklist does not contain them.

What the sector’s own numbers say

The LGA’s State of the sector: Artificial intelligence research found that around 85 per cent of responding councils are using AI or exploring it. The same research found that only around 18 per cent had a supplier policy or contract clauses requiring suppliers to declare AI use in service delivery.

Read those together. Roughly four in five authorities have no contractual mechanism to establish which of their suppliers are using AI inside services delivered to residents.

Now apply reorganisation. A successor authority inherits the supplier estate of several predecessors. Somewhere in that estate, AI is being used in ways nobody documented, under contracts that never required disclosure. The successor becomes accountable for outputs it cannot trace, produced by tools it cannot inventory, under contracts it did not negotiate. The inventory problem comes before the novation problem — and most programmes have not started the inventory.

The LGA, Socitm and Solace have told Parliament that the AI assurance ecosystem remains underdeveloped, making thorough due diligence difficult for councils. They are right. Reorganisation is where that general difficulty becomes a dated liability.

The four questions

Every reorganisation programme should be able to answer these before vesting day. Most currently cannot.

• Who is accountable for an AI-assisted output produced by an abolished council? The successor authority — that much is settled. The open question is whether it will have what it needs to defend or concede the decision, or will answer a 2029 complaint about a 2027 output with no provenance record at all.
• Where do the predecessors’ assessments and audit records go? DPIAs, equality assessments, model evaluations, human-review logs. If they live inside a supplier’s platform, do they survive contract exit? If they live in a departmental SharePoint, will anyone find them in 2029?
• Which AI supplier contracts novate, and what does exiting the rest cost? The 18 per cent figure means most authorities cannot yet answer the prior question — which contracts involve AI at all.
• How do several predecessors’ AI policies reconcile into one? Three councils will have three retention periods, three positions on whether a human must review a draft before it reaches a resident, and three views on supplier training. On vesting day there is one authority and one policy. Someone decides which — and reconciles the records created under the other two.

Norfolk is the sharpest illustration. Three new unitaries are being formed with parishes distributed between them. Records do not merely merge there; they must be disaggregated by geography. An AI-assisted consultation analysis covering parishes that end up in two different councils belongs to both authorities and to neither.

Self-assessment · Runs entirely in your browser · Nothing is recorded

Five tests. Run them this week.

Pick one AI-assisted workflow you already run and answer honestly. These are deliberately observable, timed and unglamorous — a principle you cannot test is a principle you cannot enforce. If your tools pass, you have less of a problem than this note suggests, and you should ignore the rest of it.

Nothing is recorded. This runs entirely in your browser, we do not see your answers, and there is no form to fill in at the end.

In our experience very little on the market passes test 3, and test 4 is usually answerable in principle rather than in practice. That is not a failure of any individual authority. It is a gap in what the market has been asked to build.

Test 4 is also about to matter more. The Data (Use and Access) Act 2025 replaced Article 22 of the UK GDPR with new Articles 22A–22D. Where a significant decision is based solely on automated processing, the controller must provide safeguards that give the person information about the decision, let them make representations, let them obtain human intervention, and let them contest it. Meeting that duty means producing a record. Test 4 is that record.

Most councils’ current AI use is assistance rather than automated decision-making, so the new safeguards do not bite yet. That is precisely why now is the time to build the evidence layer — before the workflows move into territory where it is legally required rather than merely wise.

A handover schedule for AI workflows

Here is what we think a predecessor authority should hand its successor. Take it, adapt it, put it in your programme plan. For each AI-assisted workload:

• 1. Inventory — what it does, which service, which named owner, which supplier, which model and version.
• 2. Contract position — is AI use declared in the contract; does it novate; what is the exit cost, notice period and data-return obligation.
• 3. Policy state — retention rule, access rule, human-review rule, training-permission position, as actually configured rather than as written in a policy document.
• 4. Assessment record — DPIA, equality impact assessment, any ATRS-style transparency record, with dates and reviewer.
• 5. Evaluation evidence — what was tested, against what task set, with what result, on what date, by whom.
• 6. Audit records — outputs that affected residents, with provenance, retained for at least the complaint and limitation period the successor will be answering against.
• 7. Known failures — incidents, near-misses, prompt-injection attempts, upheld complaints, and any output later found to be wrong.
• 8. Continuity statement — what breaks, and what the service does instead, if the model, supplier or host becomes unavailable.

Why this is genuinely hard

Item 6 of that schedule, and test 3 above, are difficult for a reason worth stating plainly: there is no agreed format for any of it.

There is no vendor-neutral way to represent an AI workflow’s configuration, its evaluation state and its audit provenance such that a different system can read them. This is why no authority can currently move a workflow — not because their supplier is obstructive, but because the interchange format does not exist to be obstructive about.

Building that format is not a procurement exercise. It is a research problem, and it is the one we think matters most: can a workflow be exported and reproduced on a different model and a different vendor, within a stated and measurable tolerance? Nobody has demonstrated it. We also think it has to be built openly and governed by councils rather than by a supplier — including us. A standard one company owns is not a standard.

What we are doing, and what we want

EAIC is convening the UK Civic AI Commons: a twelve-week discovery with six to eight authorities to build the missing layer — common controls, portable evidence, and a tested handover pattern for AI workflows between predecessor and successor authorities.

We are not building AI tools. MHCLG’s Local AI programme is doing that, with 22 councils and around a thousand officers involved, and duplicating it would waste everyone’s time. We are building what sits underneath, so that an authority can run their tools, a commercial supplier’s tools and its own work under one governance model — and move between them.

Design partners are asked for one named sponsor, one bounded workflow they already run, and about two hours a fortnight. They are not asked for resident data, money, a procurement commitment or exclusivity. We would particularly like to hear from authorities inside a reorganisation programme — and from anyone who thinks the four questions have better answers than we have found.

What we are not claiming

We have not built this. There is no product, no customer and no working portability code today. This note describes a gap and a plan, not a track record.

We may be wrong in specific ways, and it is worth naming them. Authorities may reasonably decide that inherited AI workflows are a small problem next to consolidating case management and payroll — and at some authorities they will be right. The interchange format may prove intractable at acceptable cost. Suppliers may build adequate export capability without anyone forcing them to.

If our discovery shows any of that, we will publish it, including the parts that make our own proposition look unnecessary. We would rather be usefully wrong in public than quietly right in a sales deck.

EAIC builds Citadel, a governance and oversight platform, and Sentinel, our assurance and red-teaming work, including published research on prompt-injection attacks against systems handling untrusted content. Citadel enters this on equal terms with any alternative, and an independent evaluation partner — not us — holds the assurance role, with freedom to publish findings critical of our own work. Nobody should mark their own homework.

Sources. Local Digital, Priorities at each phase of reorganisation. LGA, State of the sector: Artificial intelligence — 2025 update. LGA, Transitional governance arrangements during local government reorganisation. LGA, Socitm and Solace, written evidence to Parliament. The Local Government (Structural Changes) (Transfer of Functions, Property, Rights and Liabilities) Regulations 2008. Local Government and Social Care Ombudsman, guidance on jurisdiction. House of Commons Library, Local government reorganisation 2026 (CBP-10494). Data (Use and Access) Act 2025, s.80. MHCLG Digital, Introducing Local AI, March 2026. Figures and dates as at August 2026 — reorganisation timetables differ by area and should be checked for your own.