Minute / Local Transcribe

Minute is an AI meeting scribe built by the Incubator for AI, inside the UK government's Department for Science, Innovation and Technology. One centrally hosted copy transcribes meetings and drafts standardised summaries for every council in the pilot. Caseworkers review each summary before they submit it to the case record.

Who built it

The Incubator for AI is part of the Government Digital Service. The government built Minute for itself rather than buying it. The summaries it drafts are standardised, and councils can customise them.

The pilot

In February 2025 the tool was shown to more than 150 local-government officers. From 52 councils that applied, 25 were selected for a six-week alpha pilot, an early test version. The Local Government Association and the London Office of Technology and Innovation supported it. The selected councils are called the cohort.

Each council could have up to 25 users. Councils got the tool free of charge, on one shared central copy, without a formal contract.

22 councils stayed active. Three withdrew, citing capacity, digital fatigue, and conflicts with their cloud policies. More than 400 users tested it across adult social care, children's services, planning, housing, and democratic services.

Meetings included safeguarding reviews, supervisions between staff and their supervisors, and legal case reviews. The pilot found more than 40 social-care uses in which workers must complete long forms, reports, or assessments. On the shared copy, transcribing a meeting cost about 50 pence. Councils rated the tool 8.5 out of 10 for recommending it.

How it spread

The code was made open source in October 2025. The Ministry of Justice adapted it into a probation tool.

By June 2026 the Ministry of Housing, Communities and Local Government had adapted it into Local Transcribe. It enlisted around 500 housing and homelessness workers to pilot it, the news site PublicTechnology reports. Local Transcribe transcribes a conversation and drafts a summary in a standard format, which the worker then reviews and submits.

Why one central tool

The ministry's director of AI, Tom Smith, explained why the government built one central tool. Councils each adopting their own transcription tools meant duplicate spending, and duplicated assessment and assurance.

He said that matters given the risk of hallucination, meaning invented content, in critical interactions between officers and vulnerable groups. He also wants to avoid being locked in to one vendor.

The shared assurance

Assurance here means the checks and sign-offs a council completes before it uses a new tool. For Minute it was pooled across the cohort. The Local Government Association and the London Office of Technology and Innovation facilitated a central data-processing template. A government department reviewed it as the designated data processor, the body that handles personal data on the councils' behalf.

Councils also shared impact-assessment, equality-assessment, and consent templates, and a council readiness toolkit. Bi-weekly calls ran across the cohort, and most councils rated them useful.

Each council's own internal assurance took very different times. 14 percent of councils finished it in under two weeks, and 41 percent took more than a month.

Standard risk controls had not been completed on the alpha tool. They include penetration testing, where testers try to break into a system, and security certification.

What the sources do not show

No evaluation of how accurately Minute transcribes has been published. There is no error-rate audit, and no data on how often reviewers change a summary.

The time savings cited for it are self-reported by users or claimed by government, not independently measured. Some users reported halving their note-taking. One council estimated up to 90 percent less time on recaps. Government early testing claimed about one hour saved for each one-hour meeting.

No decommissioning, lawsuit, or scandal has been reported.

Independent research

The Ada Lovelace Institute studied AI transcription in social work. It interviewed 39 social workers across 17 local authorities in England and Scotland.

It found that local authorities evaluate these tools for efficiency, not for their impact on people who draw on care, meaning people who receive social care. Risks such as bias and hallucination are not being fully assessed. Social workers' views on how reliable the tools are, and how much human oversight they need, vary widely.

The research covers such tools in general, not Minute.

What to watch

Watch the gap between what the assurance governs and what no one has measured. The assurance covers data protection and process. No one has published whether the summaries are right. A well-documented assurance process is a claim about process, not about the summaries it produces.

Watch the shared copy too. One systematic mistake, such as a recurring mis-transcription, would not stay in one council. It would be written into many councils' records at once.

Pooling makes assurance cheaper, and it also concentrates it. Twenty-two councils share one assessment. So an accuracy question no one asks there goes unasked in every council.

The government-built design removes one exposure. There is no commercial vendor and no data-processing agreement with an outside processor, so there is no route for data to a vendor. That leaves accuracy, not data sharing, as the central question.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network, where a mistake made by one part can be passed on to the other. A pathway counts as closed once it passes on only a few mistakes. It need not stop them all.

Self-correcting means the network's mistakes are corrected rather than building on one another. A tipping point is the edge between the two. A cascading failure is one where mistakes build on one another across the network.

Side-effects and lingering effects are Lab settings, both on by default. With side-effects on, governance carries its own costs. With lingering effects on, damage outlasts its cause.

This case has a budget of 11 units, and each tool costs the same at every target level. Explore (No Targets) sets no targets.

Under Service Targets Only, one tool costing 2 units is enough to meet the targets. Escalate checks, Mark AI-written records, Assign a challenger, or Peer sharing rules each keeps mistakes from building on one another. Each also leaves the scribe helping the work.

Service and Safety Targets and All Governance Targets also ask you to close every failure pathway, among other targets. There the targets can be met. Every way to meet them uses at least three tools and costs at least 7 units.

Each way includes Escalate checks and Mark AI-written records, with Gate record entries or Store less data. Escalate checks closes the summaries given to caseworkers for review. Mark AI-written records closes the recordings to the scribe and the case history read at the next assessment. Closing the recordings pathway also lets the Privacy gauge refill, which both levels ask for. Gate record entries or Store less data closes the summaries written into both kinds of record.

Counting the stronger settings, 27 different sets of tools meet the targets under Service and Safety Targets, and 19 under All Governance Targets. None of the cheapest adds either accuracy check. Assign a challenger and Peer sharing rules add the standing accuracy check, and Check with a second model adds the second read. Some ways to meet the targets include them, but none needs them.

More is not better here. Every tool at its stronger setting costs 38 units, far over the budget. It closes every failure pathway, but it leaves the scribe hurting the work.