Government AI earns its place when document work can be checked

The UK’s Humphrey and Extract programmes show why narrow administrative tasks need source records, reviewers and service measures.

A public office has many documents that must become usable information before anyone can make a decision. Consultation responses need organising. Historic planning maps need digitising. Meeting records need checking. The United Kingdom’s 2025 announcements about Humphrey and Extract place AI in this part of government work. They deserve attention because an administrative bottleneck can be consequential even when it lacks the drama of an automated decision.

The [January Humphrey announcement](https://www.gov.uk/government/news/shake-up-of-tech-and-ai-usage-across-nhs-and-other-public-services-to-deliver-plan-for-change) described a consultation-analysis tool called Consult. The [June explanation of Extract](https://mhclgdigital.blog.gov.uk/2025/06/12/extract-using-ai-to-unlock-historic-planning-data/) described turning historic planning documents into structured data. These are official descriptions of intended applications. They do not establish that every council has adopted the tools or that a national housing target has been achieved. The change worth examining is the movement from general chatbot experimentation to a task tied to an existing public workflow.

Consider an original illustrative consultation with 800 submissions about a bus-route change. An AI assistant groups responses by concerns such as accessibility, frequency and interchange. That grouping could help analysts discover themes sooner. It must still preserve the submission identifiers behind each theme. A summary saying residents support the proposal is insufficient if most submitted objections are buried inside a broad category labelled convenience.

A reviewer could inspect every proposed theme, a random sample of routine items and all entries that mention disability, safety or a missed connection. This is a proposed review design, not a description of the government’s implementation. The important output is a traceable index that analysts can correct. A tool that provides a persuasive conclusion without recoverable underlying evidence saves typing while creating a new accountability problem.

Planning extraction has a different failure pattern. A line on an old map may represent a boundary, a road or a scanning mark. The original document, its date and its relationship to newer records matter. An illustrative council pilot could ask the system to propose polygons and document references, then have planning staff compare them with authoritative records before accepting the data. An uncertain boundary should become an explicit review item.

The [government playbook](https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government/artificial-intelligence-playbook-for-the-uk-government-html) gives civil servants a broad framework for safe selection and deployment. The operational implication is to decide who checks which output before the tool enters routine service. A document assistant should have a narrower permission set than the system that grants an application. A model’s ability to generate an answer does not supply authority to change a public record.

Benefits should be measured at the service boundary. Count the time to produce an accepted dataset, the number of corrections, the age of the backlog and the staff effort needed to resolve exceptions. In the consultation example, also measure how often minority concerns survive the grouping. An impressive reduction in initial processing time may disappear if checking takes longer or if poor records must be rebuilt.

A useful pilot includes difficult materials. Test faint scans, tables split across pages, handwritten additions and documents containing contradictory dates. Include a deliberately irrelevant instruction inside a sample submission to check whether the assistant treats it as content. Keep personal information out of demonstrations and use access controls appropriate to the actual source documents. Public service information can be sensitive even when the model task looks mundane.

There is also a purchasing question. The office needs to know how the proposed data can be exported, how model changes are recorded and what happens when the supplier becomes unavailable. A clean handover includes accepted records and unresolved exceptions. An output locked inside a proprietary interface may make the next stage of work harder, especially for a small council with limited technical capacity.

These tools are promising where a reviewer can compare an output with its source and correct it before public consequences follow. Their strongest case is a visible reduction in verified administrative work. The sensible next question for any office is precise: which document task, which evidence, which reviewer and which service measure? A public announcement opens that inquiry; it does not close it.