Two hundred contracts, and one question nobody can answer

Someone asks a question in a management meeting. Of the contracts we currently hold, how many let the counterparty terminate for convenience — and which ones?

Legal knows the answer is in the files. Finding it means opening contracts one at a time, which takes weeks. So the question is deferred, and the decision gets made without the answer.

This is not a shortcoming of the legal team. It is that the organisation's knowledge is stored in a form that cannot be searched.

Two very different jobs

Discussions about AI and legal work usually conflate two things.

The first is finding and extracting from documents that already exist: term, expiry, renewal conditions, liability caps, governing law, confidentiality obligations. Current technology does this well and verifiably, because every answer can point back to the text it came from.

The second is legal judgement — whether a clause is acceptable in this context. That requires discretion, accountability, and an understanding of what is not written in the contract.

A well-designed system does the first thoroughly and leaves the second to a person.

What works in practice

A searchable contract register. Once key terms are extracted as structured data, the question above is answered in seconds — and questions nobody used to ask, because the answer was known to be expensive, start being asked.

Date alerts. Many contracts renew automatically unless notice is given by a deadline. Organisations without alerting regularly find themselves committed for another year by accident.

Comparison against a template. When a counterparty sends a draft, the system marks what differs from the company's standard, so a lawyer reads the differences rather than the whole document looking for them.

Pre-signature checks. Automated checks catching basic errors — party names that differ between the first page and the signature block, or dates that contradict each other.

Limits worth stating plainly

Language models produce text that sounds correct, and in legal work, text that sounds correct and is wrong is the most dangerous output there is.

A citation to a provision that does not exist, or a summary that shifts a condition slightly, are errors that are hard to catch precisely because they read reasonably.

The workable defence is to require every answer to cite the source text it came from. If the system cannot show where an answer came from, the answer should not be used.

Contract data is also among the most sensitive an organisation holds. Deciding where these documents are processed, who can reach them, and where they are stored is a decision that has to happen before a tool is chosen.

Start with one question

Take the question executives ask often and legal answers slowly, and make it answerable. That is the whole scope.

Registering every contract in the organisation at once is a project large enough that it often does not finish. Answering one question takes weeks, and it demonstrates the value to whoever approves the next budget.

What GIPSIC does here

We build document management and enterprise search, including work with data under strict access control. On one project, tracing a document went from hours to under five minutes.

What we raise in the first meeting is where documents will be processed, because for this class of material that answer determines the entire architecture.

If you would like to talk about your organisation's documents, get in touch.