← All writing

A useful AI report needs a path back to its sources

Source-linked findings, uncertainty, and the review workflow in a property-document analysis project.

A project by the Big0 team, adapted from the company case study. Big0’s original case study.

A Big0 project for a UK legal-services firm explored how AI could help review property-conveyancing document bundles. The bundles included title registers, leases, transfer forms, scans, and supporting documents. Reviewers needed to identify missing material, prepare enquiries, and investigate potential issues.

The most interesting design choice was not the generated summary. It was the connection between each finding and the documents behind it.

Make a finding inspectable

The platform classified documents, extracted information, and organised its output into three groups: missing documents, suggested enquiries, and issues requiring review.

Each finding linked back to specific pages in the source bundle. The reviewer could move from a suggestion to its supporting material rather than search through the entire bundle to work out why the system had raised it.

That traceability changes the workflow. A report is useful only if the person reviewing it can check its claims. A source link does not prove a conclusion is correct, but it gives the reviewer a practical place to start.

Leave room for uncertainty

The design included explicit uncertainty when the system could not confidently resolve something. This is preferable to filling a missing detail with a plausible answer.

Confidence indicators still need to be interpreted carefully. They are signals for review, not proof of accuracy. A readable report and a confident score cannot substitute for checking the underlying document.

Evaluate omissions as well as false alarms

The Big0 case study describes comparing the system’s output with the client’s reports on 50 transaction bundles. This gave the team concrete reference material for assessing missing documents, enquiries, and flagged issues.

There were two different failure modes to examine. Missing an issue could leave a gap in the review. Raising too many unsupported issues could bury the reviewer in work that did not need to be done.

The source does not publish a detailed breakdown of those errors, so that evaluation should not be presented as a universal accuracy score or a guarantee of solicitor-equivalent performance.

Keep the professional review in the workflow

The platform presented suggestions for solicitors to review. Its purpose was to organise the first pass through a bundle and make the resulting findings easier to inspect.

For our team, the project illustrated a useful principle for AI-assisted work: the quality of the review path matters as much as the fluency of the generated output.

More from the blog →