Contract review with agents: what still needs human eyes

A law firm taking 40 NDAs a week has a real productivity problem. Each NDA is small, but it still takes focus. Combined, they eat principal capacity from work that actually requires the lawyer's judgement. Agents solve part of this in 2026, but not by replacing the lawyer. They solve it by building out the prep work that's traditionally been deprioritised because it's tedious.

In seconds, an agent can do six things that used to cost an hour. It classifies the document, identifies parties and signature setup, finds deviations from the firm's standard template, summarises material changes in the clauses that matter, checks against guidelines for liability, term and data processing, and proposes either approval or markup with reasoning.

That sounds dangerous, but it's only dangerous if the flow is built wrong. A useful principle is to separate what the agent reads from what the agent signs off on. Reading a contract is a reliable task. Concluding on a contract is not. Separate the two cleanly in the flow and you have a tool that multiplies capacity without moving risk exposure.

What still needs human review

The first thing that still has to be checked manually is clauses where a human negotiation has changed semantics without changing form. An example: a liability cap that's technically within the firm's tolerance, but combined with an unusual delivery obligation creates effectively unlimited exposure. The agent doesn't see the relationship between these without a partner who's read a hundred similar contracts. That's not a weakness in the technology. It's the nature of the craft.

The second is regulation-specific requirements that move quickly. Privacy, export controls, sector-specific compliance like financial regulation. An agent tuned in January may have outdated references by April. Humans read regulator letters. Models don't, unless someone feeds them.

The third is client strategy. A company heading into a transaction has a different risk appetite on an NDA than one running normal business. The agent doesn't see the transaction context. The partner does. If the flow has no field for "client context," the agent will consistently give recommendations that aren't calibrated to the situation.

Where the agent fits well

Where does the agent fit well? Three clear areas. Standard-template deviation is the first: the agent flags what differs and proposes a standard response with source reference. That's pure scissor work that's been manual for ten years. The second is clause consistency in long contracts. Conflicts between clause 4.3 and Schedule 2 are rarely caught by a human reading sequentially, but they're obvious to a model reading the whole document at once. The third is signal against history. When this counterparty has been in 17 prior agreements, what have they typically tried to negotiate in, and where in the agreement does it sit this time?

An undervalued piece of work is how the flow captures learning. Every time a partner overrides the agent's recommendation, that's data. If the flow records "agent said approve, partner overrode with markup on 7.2" along with the partner's reasoning, you have a training set that actually improves over time. Without that loop, the agent is static. With it, it's learning.

Security and confidentiality

Security and confidentiality aren't trivial. Client material shouldn't go to general model services with unclear data retention. In practice, the agent's reading happens either in a self-hosted environment or with a vendor that has a documented no-train policy, EU residency, and contractual security measures meeting the bar professional bodies set. With those in place, client work can be processed. Without them, the flow has to be limited to non-confidential material.

Billing

What about billing? Many firms worry that agents undermine the hourly rate. In practice we see the opposite. Clients who know the firm uses agents pay happily for the result, not the hours. An NDA reviewed, signature-ready and back to the client the same day on a fixed fee is better service than 90 minutes on hourly billing. It's a business model shift, not a threat.

How firms get started

For a firm starting out, the most robust pilot is a bounded document type and a bounded client portfolio. NDA flow for SMB clients has low conflict risk, high volume, and immediate measurable effect. If the flow holds for six weeks, it can extend to distribution agreements, employment contracts, and service agreements. Each extension is a new checklist, not a new tech stack.

What separates law firms that capture real value in 2026 isn't that they use agents. It's that they've built clean boundaries between what the agent does, what the human owns, and how learning circulates. That's process work. Like good legal practice has always been.