AI for In-House Lawyers in a Small Company: Contract Triage and Playbooks

A lone in-house lawyer can use AI to triage incoming contracts by sending every request through one intake form, comparing each contract to a clause playbook that marks positions red, amber or green, and asking the AI to produce a first-pass issues list and redline. You then review the flagged clauses yourself, so your time goes to the contracts and clauses that actually carry risk.

Two cautions shape everything that follows. AI misses things and sometimes invents them, so it triages, it does not advise. And contract text is confidential, so the tool, the plan and the categories you may paste are decisions made before the first upload. None of this is legal advice; check privilege, confidentiality and professional-conduct rules with your professional body or outside counsel.

Follow me on Instagram@sagnikteaches

The problem sole counsel actually has

In a company of 20 to 200 people, one lawyer receives contracts from everyone: sales sending customer paper, procurement forwarding supplier terms, HR asking about a contractor agreement, a director pasting a clause into a chat message with "quick look?". Most are routine. A few matter a great deal. The review itself is rarely the expensive part; the interruptions and the lack of a queue are.

Connect on LinkedInSagnik Bhattacharya

So the routine has three jobs: put requests into one place (intake), decide how much attention each needs (triage), and apply your standard positions consistently (playbook). AI helps with the second and third, and with drafting. It does nothing for the first unless you build the form.

Subscribe on YouTube@sagnikteaches

Step one: an intake form that gives the AI something to work with

Much of triage time goes on asking the same four questions by email. Build a short form (a shared form tool, a ticket template or even a structured email) and refuse to open contracts that arrive without it. Fields that earn their place:

• Counterparty and what they supply or buy, in one line.
• Whose paper: ours, theirs or negotiated from a template.
• Value and term: total or annual value, start date, length, renewal.
• Data involved: none, business contact data, personal data, special categories, source code, customer data.
• Business owner and signer: who is accountable and who has authority.
• Deadline and why: "needed by Friday" is not a reason; "launch on Monday" is.
• Anything already agreed by email that the contract might contradict.

The fields feed the triage rule below. Value and data answer most of the question before you read a line.

Step two: a triage rule that sorts contracts into three lanes

Write the rule down so it is yours, not the tool's. A workable version:

Set the value thresholds with your finance director, for example green under $10,000 a year and amber under $100,000. Those figures are illustrative. The point is that the lane is set by the form fields and written rules, so that a rushed afternoon does not quietly downgrade a risky contract.

Step three: the clause playbook

A playbook records your preferred position, the fallback you can accept and the point at which you escalate, per clause. Without one, AI reviews against generic "market standard", which may be nothing like your risk appetite. With one, you can ask a precise question: does this clause fall within green, amber or red?

Here is a filled-in extract (illustrative positions; yours depend on your business and advice).

Build the first version from the last ten contracts you negotiated. Ask the AI to extract each clause family and the position finally agreed, then you correct it. It will propose "market" positions you have never held; delete those, because a playbook full of positions you do not take is worse than none. Date and version the document, and keep the previous version, so you can say what applied on a given date.

Step four: the triage prompt and what it gives back

Use the AI inside an environment you have cleared for confidential material (see the data section). Paste the playbook table, the intake fields and the contract text, then ask for a structured output. Do not ask "is this contract OK?".

You are assisting sole in-house counsel. Do not give legal advice;
produce a structured first-pass review for the lawyer to check.

Inputs: (1) our clause playbook, (2) the intake form, (3) the contract.

Output:
1. One-paragraph summary: parties, subject, value, term, whose paper.
2. A table of every playbook clause family: where it appears (clause
   number, quote the first 15 words), which colour it falls in
   (green, amber, red, or "not found"), and why in one sentence.
3. Clauses NOT in my playbook that look unusual (list with clause number).
4. Contradictions with the intake form or with other clauses.
5. Definitions used in the liability and indemnity clauses.
6. Suggested redline wording for amber and red items only, using my
   amber-position language where it exists.
Quote exact text for every flag. If you cannot find a clause, say
"not found", do not infer one.

An illustrative extract of what comes back for a supplier agreement: "Liability cap (cl. 14.2, 'The Supplier's total liability shall not exceed...'): amber. Cap is 6 months' fees; our green position is 12. Indemnity (cl. 15.1, 'The Customer shall indemnify the Supplier against all losses...'): red. One-way, against us, uncapped by reference to cl. 14. Auto-renewal (cl. 3.4): amber. 20 days' notice; our green is 30. Not in playbook: cl. 11 gives the supplier a right to change prices on 30 days' notice."

What you fix: check that clause numbers match the document, because assistants sometimes cite the wrong one, and that "not found" means not found. Search the text for the keyword yourself (a 10-second find). A "not found" that is wrong is the most dangerous output, because it looks like a gap in the contract when it is a gap in the reading.

Step five: redline checks before anything is sent

AI-drafted redlines go wrong in quiet ways. Before a redline leaves your desk run these checks, in order.

1. Compare to the original, mechanically. Use your word processor's compare function between the counterparty's version and yours, not the AI's description of what it changed. Assistants occasionally alter text they were not asked to touch, or silently tidy a definition.
2. Read every changed defined term. A tweak to "Losses" or "Confidential Information" can reach every clause that uses it.
3. Check cross-references. If clause 12 is deleted, clauses that cite it need adjusting.
4. Check numbers and dates. Caps, notice periods, amounts and governing-law wording are where an invented "typical" figure appears.
5. Read the tone of comment bubbles. AI explanations written for the counterparty can be needlessly blunt or reveal your internal position.

A realistic mistake and how it showed up: the assistant redrafted a limitation of liability to "mutual" but left the carve-out list saying "the Customer's liability for fees is not limited", so the clause was still lopsided. It surfaced because the compare view showed the carve-out had not changed while the cap had. Reading only the AI's cheerful summary would have missed it.

Data handling and privilege, decided before the first upload

These are the decisions to make once, in writing, and attach to the playbook.

Choose the environment

Use a business or enterprise plan whose own terms say your content is not used to train models by default, and read those terms on the vendor's page rather than relying on a summary. As of November 2026, OpenAI's help centre says ChatGPT Business content is not used to improve its models by default, Anthropic's privacy centre says the same of its commercial products such as Claude Team, and both plans start at two seats. Individual plans on both services may train on chats unless the setting is switched off. Seat prices change, so confirm the current figure on each pricing page. A free consumer account is not the place for anyone's contract. For tool selection see ChatGPT or a legal AI tool and legal AI pricing for small firms.

Check the contract you are about to paste

Some confidentiality clauses forbid disclosure to third-party processors, and some customers prohibit AI use in their contract with you. Add a check to the intake form: "Does the existing agreement with this counterparty restrict third-party tools or AI?". If yes, the contract goes to the red lane by hand. See mentioning AI use in contracts and questions to ask before buying AI that touches client data.

Mind privilege

Whether using a third-party AI tool affects legal professional privilege depends on the tool's terms, who can access the content and the rules where you practise, and the position is still developing. A cautious approach: use only tools with confidentiality commitments, keep your analysis notes and the AI output in your matter file rather than in shared chat history, label them privileged and confidential, and avoid forwarding AI output to business colleagues unless you intend to. Take advice from your professional body or outside counsel for your own situation.

Keep an excluded list

Write down what never goes into AI: litigation material, regulatory investigations, deal documents, personnel disputes and anything marked as restricted by an external party. Related reading: legal tasks a small firm should never hand to AI.

Five worked mini-cases

A bakery chain's supplier agreement. Procurement forwards a flour supplier's terms that let it change prices on 14 days' notice. Intake fields put it in amber. AI flags the price clause as "not in playbook" and the 20-day auto-renewal notice window as amber. You send a two-line redline: price changes need 60 days' notice and a termination right if prices rise over 5 per cent. Elapsed time 25 minutes.

A driving-school group's insurer endorsement. A document that looks like a contract is really an insurance endorsement; the playbook has no clause family for it, so every row says "not found". The sensible outcome is the red lane: you read it, and the AI's only role is a plain-English summary you check.

A before and after on a vague request. Before: an email saying "Can you check this? Needs signing today." After the form: counterparty, annual value $38,000, their paper, includes customer contact data, signer is the managing director, deadline tied to a launch. You triage to amber, reply with the expected turnaround, and ask for the deadline reason in the form next time.

An NDA, green lane. A mutual NDA on your template with the standard term. The AI check is a one-minute comparison against the template, with differences listed. You approve in five minutes and sample-check at month end.

A quick sum on time. Suppose you handle 30 contracts a month: 12 green at 5 minutes, 12 amber at 35 minutes, 6 red at 2 hours. That is 1 hour, 7 hours and 12 hours, 20 hours in all. Without triage the same set at 90 minutes average would take 45 hours. The numbers are illustrative; measure your own for a month before you promise anyone a saving.

A filled-in intake form, and the reply you send back

Seeing the form completed makes the lanes concrete. Here is an illustrative submission from a sales manager at a company that sells equipment servicing.

Counterparty: [customer], a regional chain of 14 sites
What they buy: annual service contract for 40 machines
Whose paper: theirs (master services agreement, 22 pages)
Value and term: $64,000 a year, 3 years, auto-renews
Data involved: site staff names and phone numbers
Business owner / signer: [name], sales director / managing director
Deadline and why: tender award letter says signed by 28th
Already agreed by email: 4-hour response time on breakdowns
Restrictions on AI or third parties in existing agreements: none known

By the rules above this is amber on value ($64,000 is under the $100,000 limit) but has two points that could move it to red: a three-year term that auto-renews, and a service-level promise made by email that the draft may not contain. Your reply, drafted with AI in a minute and edited by you, might read: "Received. Amber review. I will have an issues list to you by Thursday noon. Two points to confirm first: does the 4-hour response time appear anywhere in their draft, and who is responsible for the penalty clause if we miss it? Please do not sign or confirm anything verbally until you hear from me."

That reply does three things a busy lawyer often skips: it commits to a time, it narrows the questions, and it stops the business making promises while you review. All three reduce the chance of an urgent, late-night escalation.

Giving the business a self-serve lane that holds

The green lane only saves time if colleagues use it correctly. Three small pieces of paper do most of the work.

• A one-page "can I sign this?" guide. It says: our template, under the value limit, no personal data, no changes to the text. If any answer is no, send it to legal. Ask the AI to turn your triage rule into a plain-language flowchart, then simplify it by hand.
• A list of approved templates kept in one shared folder, each with an owner and a review date. Out-of-date templates cause more green-lane errors than clever counterparties do.
• A delegated authority table showing who can sign what value. You do not set this alone; the board or finance director does. Your job is to make sure the intake form asks the right question about it.

Test the guide on a colleague who has never read it. If they cannot classify three sample contracts correctly in five minutes, it is too long. Run a short session when you launch it, and re-send it when the playbook changes.

Keeping the playbook current as positions move

A playbook fossilises quickly if you never feed back what you conceded. After each amber or red negotiation, add one line to a change log: clause, what the counterparty asked for, what you agreed, and why. Every quarter, paste the log into the assistant and ask: "Which clause positions in my playbook have I conceded more than three times? Propose a revised amber position for each, based only on what is in the log." Then decide for yourself. Where you keep conceding, your green position is probably unrealistic; where you never concede, you may be able to tighten. It is a measured way of keeping the document honest, and it gives you a record you can show a new colleague or a board.

Measuring whether the routine is working

Track four numbers monthly in a simple sheet: contracts received by lane, median turnaround per lane, number of AI flags you rejected as wrong, and number of issues you found that the AI missed. The last two are the important ones. If you reject more than a third of flags, the playbook wording or the prompt needs tightening. If you find a missed issue in a green contract, move that clause family to a stricter threshold.

Once a quarter, pull three contracts already signed through the green lane and review them in full. It is a quality sample of the sort a larger legal team would run, and it will tell you whether the delegated authority limits are right.

Finally, add a short note to the playbook on when to bring in outside counsel: unusual jurisdiction, anything over a stated value, anything novel in the AI area itself. AI helps the lone lawyer cover more ground; it does not change when a second pair of expert eyes is the right call. If you want help designing the intake and tooling around your own contracts, my AI implementation consultation covers that, and AI contract review for small firms explains what these tools catch and miss.

Questions sole counsel ask about AI and contracts

Can AI replace my review of a contract?

No. It can sort contracts by risk, compare clauses to your playbook and draft first-pass redlines, but you remain responsible for the advice. Treat output as a junior's first pass: useful for speed, unreliable on judgement, and always checked against the actual text.

Will using an AI tool waive privilege?

It can raise the question, depending on the tool's terms, who can access the content and your jurisdiction's rules. Use a business plan with no training on your content and clear confidentiality terms, keep matter notes within approved tools, and check with your professional body or outside counsel.

Which contracts should never go through AI triage?

Anything where disclosure would itself be harmful: pending litigation, regulatory investigations, M&A, personnel disputes and anything under a confidentiality term that forbids third-party processing. Check the clause before you paste, and keep a list of excluded categories in the playbook.

How long does a playbook take to build?

A useful first version covers six to ten clause types and takes a morning if you start from your last ten negotiated contracts. It grows as you add the positions you actually concede. Version it and date it so you can show what applied at the time.

Further reads

• Can Lawyers Use AI Note Takers in Client Meetings? — Whether AI note takers suit confidential meetings.
• How to Summarise Bundles and Transcripts With AI Safely — Summarising long documents safely.
• Should You Sign an NDA Before Sharing Data With an AI Consultant? — NDA points when a third party sees company data.
• AI Clauses for Agency Contracts: Disclosure, Ownership, Approvals — AI clauses worth putting into services contracts.
• AI Literacy Requirements: What Your Staff Need to Know — What the rest of the team needs to know about AI.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Drafting routine letters with AI in a law practice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: claude.com pricing page and Anthropic's privacy centre article on model training for commercial products; OpenAI's ChatGPT pricing page and help-centre article on how data is used to improve models (all checked 4 November 2026); general practice for contract playbooks and clause review. No legal advice is given.