An email from a knowledge lawyer
Your product reviews leases and extracts key terms: break dates, rent review mechanisms, alienation restrictions. A knowledge lawyer emails: on a lease review last week, your product said there was no break clause. There was one, in a schedule, drafted unusually. The associate relied on the summary before checking the lease, and the supervising partner caught it. They want to know why it happened and what you are doing about it.
Your team looks. The product does not store which parts of the lease it read for each answer. You can rerun it, but the model version has changed since, and the output is different. You fix the prompt, but have no way to know whether the fix breaks something else.
Why AI extraction errors are hard to handle
Legal documents are drafted in endless variations, and extraction features will sometimes get things wrong. The problem is not only the error rate. It is the lack of evidence and the lack of a way to measure improvement.
- Answers are shown without the source text they were based on.
- The product shows every answer with the same confidence, so users cannot tell which to check.
- Inputs, outputs and model versions are not recorded, so an error cannot be reproduced.
- Fee earner corrections are made in the final document, and your product never learns about them.
- Prompt and model changes are tested on a few examples, not a representative set.
What unmanaged errors cost
Every error a firm finds reduces trust in every answer your product gives, and knowledge lawyers talk to each other across firms. Without evidence, you cannot explain what happened, which makes the firm's risk team nervous. Without an evaluation set, each fix is a guess that may create new errors elsewhere. Firms may restrict the product to low-value tasks, or tell fee earners to verify everything, which removes the time saving that justified buying it.
How we build extraction firms can check
What we build makes every answer traceable, highlights the uncertain ones, and turns corrections into a test suite.
- Source citations for every extracted term: the clause or paragraph it came from, with a link that opens the document at that point.
- Explicit not found answers that say what was searched, so no break clause found in the main body or schedules is distinguishable from a guess.
- Confidence signals based on the model's answer, agreement between checks and whether the source text is unusual, with low-confidence answers flagged for review.
- A record of each extraction: document version, prompt, model and version, sources and output, so any answer can be examined later.
- Correction capture: fee earners mark an answer as wrong and give the right one with its location, in a few clicks.
- An evaluation set built from corrected examples, with permission, plus synthetic and public documents, run against every prompt or model change before release.
| Firm's question | What the product shows |
|---|---|
| Where did this answer come from? | Cited clause, opened in the document |
| Did it look in the schedules? | What was searched for a not found answer |
| Which answers should we check? | Low-confidence flags |
| What exactly happened last week? | Stored record with model version and sources |
| Is your fix any good? | Evaluation results before and after the change |
Confidentiality matters here. Corrections from one firm's documents are only used in evaluation with that firm's agreement, and never shown to another firm.
The next unusual break clause
A lease arrives with a break clause buried in a schedule. The product finds it, cites the schedule paragraph, and flags it as low confidence because the drafting is unusual. The associate clicks the citation, reads the clause, and confirms it. When the product does miss something on another lease, the associate marks the correction, the case is added to the evaluation set, and the next prompt change is tested against it before release. The knowledge lawyer sees citations on every answer and fewer surprises.
Can your product explain its answers?
- Extracted terms are shown without the source text.
- Every answer looks equally certain.
- You cannot reproduce an answer from last month.
- Fee earner corrections never reach your team.
- Prompt or model changes are released without a proper test set.