Insight NDA

Insight NDA works with agents and negotiators who handle a high volume of NDAs and commercial agreements. Before each discussion, an agent needs the key terms in their head: duration, scope, jurisdiction, exclusivity, limitations on liability. Reading a 20-page NDA in the minutes before a call is not practical, and pushing the review down to junior staff produced summaries that varied with whoever wrote them. The team needed structured, reliable summaries of legal documents without a lawyer at the reviewing end.

Volume was the pressure point. A single agent might need to review three or four documents in one morning, and a manual pass took around 20 minutes per document. That reading time was the ceiling on how many engagements the team could run at once.

The platform had to be in production in three weeks. A client engagement was starting in week four, a series of negotiations where the team intended to use the tool live, on real agreements, from the first call. If it was not ready, it would be shelved until the next suitable project came along, with no certainty of when that might be. The engagement was the launch, and the launch was the engagement.

That shaped every decision. There was no phased rollout and no beta period. The build was scoped to the one thing that had to work under time pressure: upload a document, get the terms, trust them enough to walk into a call.

The platform accepts a PDF upload and extracts the full text. That text goes to OpenAI with a prompt engineered around the specific terms that decide an NDA negotiation: effective date, term length, governing law, parties, scope of confidentiality, carve-outs and breach consequences. The model returns them as a structured data object rather than prose, and the platform renders that object as a formatted summary alongside the original document.

Each extracted term is paired with the clause it came from. Agents can annotate a summary, flag sections for follow-up and export a one-page summary PDF before a call. The document stays in the system, linked to the client record, so the analysis outlives the single engagement it was prepared for and is there the next time the same counterparty appears.

Keeping the AI output beside the source text, rather than in place of it, was the central design decision. Agents did not want to trust a summary they could not verify quickly. Seeing the relevant clause next to the extracted term answered that directly and turned the summary into a reading aid rather than a replacement for the document.

The alternative was a general document Q&A interface: upload the PDF, ask questions, receive answers. It is more flexible and faster to build, and it is what most document AI demos look like. The problem is the moment of use. An agent with minutes before a call does not have time to work out which questions to ask, and a question phrased slightly differently can pull a different answer from the same clause.

A fixed extraction format delivers the same seven terms in the same position every time, regardless of how the underlying agreement is drafted. That consistency is what makes the summary skimmable under pressure, and it is what makes two agents' preparation comparable. Flexibility was traded for predictability, and predictability was what the engagement needed.

An agent drags a PDF into the platform and has a structured summary within 90 seconds. They skim the seven terms, click through to any clause that looks unusual, add a note for the client file and export the one-page summary to take into the call. Nobody reads the whole document unless a flagged clause deserves it.

The client record accumulates every document analysed for that engagement, with the annotations attached. Junior staff no longer produce ad hoc summaries, and senior agents no longer re-read documents that were reviewed the week before. Review has become a check rather than a task, and the check takes the time it should.

The platform handled 40 documents across 12 client files in its first month. Review time before a negotiation call dropped from around 20 minutes to under 5, and the extraction has been consistent enough that agents treat the summary as the first thing they open rather than the last thing they check.

That reclaimed time had an operational effect. The team took on two additional concurrent engagements in the month after launch, something that had not been feasible while every document required a manual read. The concept went to production in 3 weeks, and the client-linked document store now gives every future negotiation a head start.