An in-house legal team asked me a difficult question recently.
If ChatGPT or Claude can read a contract, follow instructions and produce a redline in Word, what is a specialist legal AI product doing?
I run one, so I have an obvious bias. But I thought the challenge was right.
If you give ChatGPT good instructions on a straightforward contract, it can produce a decent redline. Also, the models will keep improving, and any feature-by-feature answer is likely to age badly.
But a redline is the output. The harder question is one level higher: what positions should the AI take?
This goes to Legal’s core job: help the business get deals done without accepting risks the company never meant to take.
For teams that have not standardised how they apply the company’s risk tolerance, every lawyer tries to strike that balance individually, one contract to the next. Some will push too hard and slow the deal down. Others will accept more risk than the company intended.
Left to its own devices, AI will not know where that company wants to sit either. So a lawyer still has to fix the output according to their own view of the right balance.
Why contract playbooks exist
That has always been one reason to use contract playbooks. You get your most experienced people to calibrate the risk/reward trade-off once, then help the rest of the team follow it. As AI takes on more of the redlining, the return on doing that increases.
Here’s a chart that summarises what happens across a team without playbooks:
At one end, the team under-lawyers. It closes the contract faster, but the additional legal risk costs more than the speed was worth.
At the other, it over-lawyers. It reduces the legal risk further, but the commercial cost of a protracted negotiation is greater than the benefit.
The playbook’s goal is to place the dot in the middle, as best as you can. You will never place the dot with mathematical precision. The right position will also vary by company, sector and contract type (and occasionally counterparty).
The point is to make the calibration deliberate and reusable (as much as possible), rather than asking every lawyer and every AI session to recreate it from scratch for every contract.
So the playbook is the team’s best current answer. The governance system that underpins it determines who can set positions, who can apply them, when they may depart from them for a given deal or counterparty, and how the answer changes over time.
Who writes the instructions?
Much of a legal team’s useful knowledge sits with the practising lawyers responsible for different types of work or other subject-matter experts (SMEs) across the org, like IT security, finance, insurance, privacy etc.
The lawyer who negotiates the company’s DPAs knows which positions are preferred, where the team normally compromises and which points need to be escalated. The IT team may hold the same knowledge for cyber provisions.
If that knowledge stays in their heads, the AI cannot use it and the rest of the team cannot reuse it.
So the first problem is not model intelligence. It is making the knowledge easy enough to capture and maintain while lawyers or contract managers are doing the work.
Asking a central legal operations or technology team to do 100% of that creates control, but it can also create a queue. Every update requires someone to interview the subject-matter lawyer, translate the advice into instructions and test it.
But if you let everyone maintain their own playbooks you get the opposite problem: several versions of the same policy, with no reliable way to know which one is approved (and you also reintroduce the over/under-lawyering problem).
The better structure is federated. The lawyers closest to each type of work own the substance. The organisation controls who may edit, approve, publish and use it.
How much governance do you need?
The answer depends partly on the size and shape of the team.
In a small legal team, it is quite common for lawyers to manage and use their own playbooks, or simply use a general-purpose AI chat tool. They may not have standardised every position or fallback across the team.
As we saw, that’s a bit of a missed opportunity. The lawyers duplicate work and the company may take different positions depending on who reviews the contract. Some will over-lawyer relative to the team’s ideal balance; others will under-lawyer.
But it is not necessarily a disaster. Each lawyer can apply judgment to the AI’s suggestions and correct them. Not ideal, but it will still save time.
That changes as the team grows. It changes even more when contract managers, paralegals or procurement managers handle contracts under Legal’s guidance.
Imagine a counterparty redlines the company’s own template on a point where the playbook has no fallback.
A senior lawyer may want the AI to suggest several options. They can assess the legal and commercial trade-offs and decide whether one of them is acceptable. If the same issue appears several times, the legal team should be able to decide whether to add a fallback, change the playbook or amend its own template.
On the flip side, the legal team may not want a procurement manager to receive the same AI-generated options for a novel issue and choose a new company position. It may instead want the AI to explain that the point sits outside the approved playbook and escalate it to Legal. Legal can then deal with it, and ideally update the playbook so that on the next run, procurement won’t need to escalate.
In both cases, the AI should proactively propose the playbook update.
The underlying model is equally capable in both cases. The difference is the authority of the person using it.
That is what the governance layer is for. It determines who can use each playbook, choose between approved fallbacks or receive a new option, when an issue must be escalated, and who can turn the answer into guidance for future contracts.
It does not force every user to behave identically. It keeps them working within the same risk/reward calibration while giving a senior lawyer more room to exercise judgment than eg a procurement user.
Where DraftPilot fits
At DraftPilot, lawyers can create and refine playbooks conversationally through AI chat (by uploading historical mark-ups, signed copies etc), share them through team workspaces with different permissions, and apply them to tracked-change contract review in Word. DraftPilot flags when a position needs to be escalated internally and provides data analytics on which contracts and topics carry the most escalations, so it’s easy to prioritise where to add more playbook detail. Throughout the process, the lawyer remains in control of every change.
The fuller loop goes further:
more detailed role-based AI outputs (eg a non legal user can use AI Chat for summaries but not for legal advice, the chat should refer the user to Legal in that case);
explicit escalation workflows including logging all escalation decisions;
analysis of positions taken across matters; and
AI-proposed updates to both the playbooks and the team’s contract templates based on repeated authorised decisions.
That is the direction specialist legal AI needs to move in and what we’re laser focussed on at DraftPilot.
What this changes for in-house Legal
This governance layer provides the “checks and balances” I wrote about in my previous post:
“The company has a legal team to keep deals moving without accepting undue risk. Lowering legal cost is attractive, but not if it weakens the judgment and controls the function was hired to provide. There’d be no point in that.
As a result, full automation of the in-house legal team without checks and balances on quality offers an unattractive trade. The potential saving is a percentage of a relatively small cost line. But the downside is potentially worse performance on the dimension that justifies the team’s entire existence”
But as AI does more of the contract work, those checks cannot depend on a lawyer reviewing every line themselves.
The role of in-house Legal moves up a level. Less time doing every redline, more time deciding what positions the AI should take, who is allowed to use them, what needs to be escalated and when a decision should change the playbook.
Legal still owns the judgment. But instead of recreating it clause by clause on every contract, the team applies more of it once through playbooks, permissions and escalation rules.
Back to the original question: a decent redline is useful, but it is not enough. A specialist system should help Legal govern how AI handles contracts across people and matters, and improve that system over time.
So if the demo ends when the redline appears, it has stopped too early!
Thanks for being here,
Daniel
CEO at DraftPilot
P.S. My day job is running DraftPilot, where we help legal and contracting teams review contracts faster, using legal-approved playbooks to keep the work fast, consistent and under legal’s control.
If this could be useful for your team, just hit reply and we can give you a demo!