Case pattern: a public organisation that started with document search
A composite pattern showing how a public body began its AI work with internal document search, why that was the right first project, and what it enabled.
The short answer
This is a composite pattern rather than a named organisation. The shape: case workers and advisers in a public body spend hours each week looking for things, current policy on an unusual situation, the regulation that applies, a precedent decision, the internal guidance that superseded last year’s, a form’s correct version. The material exists, spread across shared drives, a document management system, an intranet, email attachments and people’s memories, and finding it depends on knowing who to ask. The first AI project was search that understands a question asked in ordinary language, retrieves the relevant passages across all those locations, and answers with citations to the source documents, respecting each user’s permissions. It touched no citizen’s decision, required no new legal basis beyond ordinary data protection review, and returned hours immediately. It was the right first project not only for the gain but because organising the documents, deciding what was current and setting the access rules built the foundation every later project needed.
Why this project and not another
| Candidate first project | Risk | Data readiness | Why not first |
|---|---|---|---|
| Internal document search | Low | Documents exist; organising them is the work | Chosen |
| Citizen chatbot | Medium | Needs current published information and a hand-over route | Later; depends on the same document work |
| Request routing and classification | Low to medium | Needs consistent categories | Second project; benefited from the first |
| Eligibility pre-checks | High | Needs clean case data and legal review | Much later, if at all |
| Automated decisions | Highest | Rarely a lawful basis | Excluded by policy |
How it was built
- Inventoried the document sources and, for each, decided whether it was in scope and who owned it.
- Decided what was current: superseded documents were archived or marked, which was a substantial piece of work and valuable in itself.
- Set access rules so retrieval respects each user’s permissions.
- Indexed the in-scope documents, with the source, date and owner attached to each passage.
- Built the interface to answer with citations and links, never without.
- Ran a data protection review before launch and documented it.
- Piloted with one team, measuring time spent searching before and after.
- Published it in the algorithm register with a plain description.
- Reviewed monthly: questions asked, answers with no good source, documents that turned out to be missing or outdated.
What it enabled
The question log became a map of where guidance was unclear or missing, which the policy team used. Documents that nobody could find turned out not to exist, and were written. The access rules, once set, applied to later projects. Categories that emerged from the questions informed the request-routing project that followed. And staff who had been sceptical about AI had used something genuinely helpful that never pretended to decide anything, which made the next conversation easier.
What this means for you
For a public organisation, internal document search is an ideal first AI project: low risk, immediate measurable gain, no citizen decision touched, and groundwork that every later project needs. Decide scope and currency before indexing, respect permissions, require citations on every answer, review the questions monthly for gaps in guidance, and publish it in your register. The hours come back and the foundation is built.
Frequently asked questions
Why start a public organisation with document search?
Because it is internal, touches no citizen's rights, is legally straightforward, and addresses a genuine daily frustration. It also forces the organisation to inventory and organise its own documents, decide what is current, and set access rules, all of which are prerequisites for anything more ambitious. The gain is immediate and the groundwork is permanent.
How is it different from the search we already have?
Conventional search matches words in filenames and text. This retrieves passages relevant to a question asked in ordinary language, across formats and locations, and presents an answer with citations to the passages it used. The officer reads the cited source rather than trusting the summary. The difference in practice is between knowing a document exists somewhere and getting the relevant paragraph with a link.
What could go wrong?
Two things. Answers presented without sources, so officers act on a summary that may be wrong or out of date; the rule requiring citations and reading the source prevents it. And indexing documents that some staff should not see, which permission-aware retrieval prevents. Both are design decisions made before launch, not features added after.