AI Agents for In-House Legal Teams: Contracts and Requests

AI agents for in-house legal teams work best on the work around the law, not the law itself: sorting incoming requests, tracking where each contract stands, and finding the clause a lawyer needs. The agent prepares, routes and retrieves with sources, and the lawyer decides and signs off. Set up this way, the team spends less time on triage and status chasing and keeps full control of every opinion.

Key takeaways

• The strongest first uses are intake triage, contract status tracking and clause lookup, because they are repetitive, checkable and do not require legal judgment.
• Every answer should carry its source: the document, the version, the system and the time it was last updated. An answer without a source is a guess.
• Legal opinions, risk acceptance, negotiation positions and anything sent to a counterparty stay with lawyers, enforced by permissions and approval steps rather than by instructions alone.
• Start read-only, with one request type and one contract stage, and widen access only after lawyers have compared the agent's work with their own.
• Treat contracts and incoming emails as untrusted text. Instructions hidden inside a document must never change what the agent does.

What work around legal can an agent actually take on?

An in-house legal team is a service desk for the rest of the company. Sales wants a confidentiality agreement today. Procurement needs a vendor contract reviewed before a kickoff. A manager asks whether a renewal notice window is still open. Very little of the effort in these requests is legal judgment. Most of it is finding out who is asking, what they need, by when, which template applies, whether an earlier agreement exists and who must approve.

That surrounding work is where an agent helps. Three jobs fit well:

• Intake triage: read each request, check it is complete, classify it and route it to the right person.
• Contract status tracking: answer where a contract is, who is holding it and what happens next.
• Clause lookup: find the relevant wording across executed agreements and the team's own playbook, and show it exactly as written.

This differs from the setup in AI agents for law firms, where client intake and court or filing deadlines are central. In a company, the requesters are colleagues, the memory is the company's own contract base, and the main risk is not a missed client but a wrong answer that someone acts on without a lawyer seeing it.

How does intake triage work for legal requests?

Without triage, requests arrive by email, chat and hallway conversation, each with different missing details. An agent can sit in front of those channels and do what a good legal coordinator does: acknowledge the request, check it against a short list of required fields, ask the requester for what is missing and hand a complete package to the right lawyer.

A useful intake package usually contains:

• The counterparty's full legal name and the business owner of the relationship.
• The type of document: new agreement, amendment, renewal, termination or a question.
• Whose paper it is, ours or theirs, and any draft attached.
• The requested deadline and the reason for it.
• Related existing agreements the agent can find in the repository.
• Whether personal data, regulated activity or unusual terms are involved, as stated by the requester.

Routing rules should be written by the legal team in its own words and kept short. A request for a standard confidentiality agreement on the approved template can go to a fast lane where a lawyer confirms rather than drafts. Anything involving personal data goes to whoever handles privacy. Anything that does not match a rule goes to a named person with a short summary, never to a silent default.

The agent's replies to requesters should be limited to four things: acknowledgement, requests for missing information, status, and self-service guidance that lawyers have written and approved in advance. It should not tell a colleague that a clause is acceptable or that a plan is lawful. When someone asks that kind of question, the correct behavior is to route it and say that a lawyer will respond.

How can an agent track contract status without guessing?

Status questions consume more legal time than most teams admit, because the answer is scattered. The draft sits in a document folder, the approval in a ticket, the signature in an e-signature tool and the last counterparty message in someone's inbox. An agent with read access to those systems can assemble the answer and cite where each part came from.

A workable status model has a small set of stages: requested, in drafting, with counterparty, in internal approval, awaiting signature, signed, and nearing expiry or renewal. The agent maps each contract to one stage using the system of record and reports the time of the last update. If two systems disagree, for example the ticket says approved but the signature tool shows nothing sent, the agent says so instead of picking one.

Here is an illustrative scenario. A sales manager asks in chat whether a customer agreement has been signed. The agent finds the contract in the repository, sees that it was sent for signature and that one of two signers has not yet signed, and replies with who is pending and when the request went out. It offers to send a reminder. Because a reminder goes to an external party, the agent drafts it and places it in the coordinator's queue for approval. The manager gets a precise answer in moments, and nothing leaves the company without a person agreeing to it.

The same logic covers renewals. The agent can list agreements whose notice windows are approaching, based on dates already recorded in the repository, and open a task for the owner. It does not decide whether to renew.

How does clause lookup stay trustworthy?

Clause lookup is retrieval, and retrieval is only as reliable as its discipline. The agent searches executed agreements and the legal playbook, then returns the original wording, not a paraphrase presented as a conclusion. For the technical background, [AWS explains retrieval-augmented generation](https://aws.amazon.com/what-is/retrieval-augmented-generation/) clearly. The business rule is simpler: quote first, summarize second, and never conclude.

A lookup answer a lawyer can rely on includes:

• The exact clause text, quoted, with the section reference.
• The document name, version and whether it is executed or still a draft.
• Effective and expiry dates as recorded.
• Any amendments or side letters linked to the same agreement, flagged as possibly changing the clause.
• A plain statement when nothing was found, instead of the closest-sounding alternative.

The amendment point matters. A clause may read one way in the original and another after a later change. The agent can surface the connected documents, but whether the amendment alters the meaning is a legal reading, so it goes to the lawyer. Many wrong answers in this area come from an agent that fills a gap with something plausible. The practical controls are covered in how to reduce AI hallucinations in business workflows: require citations, allow the answer not found, and test with questions whose answers you already know.

The agent finds and prepares. The lawyer decides.

Where does the line sit between agent and lawyer?

The line should be written down and enforced by the system, not left to good intentions. A simple decision guide has three tiers:

• The agent acts alone: acknowledging requests, asking for missing details, reporting status with sources, quoting clauses with citations, creating internal tasks and reminders for owners.
• The agent prepares and a person approves: any message to a counterparty, a first-pass comparison of a draft against the playbook, a summary of an agreement for a lawyer, and any change to a record in the contract repository.
• Lawyers only: legal opinions, interpretation of ambiguous wording, risk acceptance, negotiation positions, deviations from the playbook, privilege decisions and signatures.

Two controls make the line real. The first is permissions. Give the agent the narrowest access that lets it do its job, ideally read-only at the start, and restrict sensitive matters entirely. The approach in least privilege for AI agents applies directly, and the [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a good reference for the underlying principles. Confidential matters such as disputes, investigations or board topics should sit outside the agent's reach altogether.

The second control is treating text as untrusted. A contract, a pasted email or an attachment can contain sentences written to look like instructions. The agent must read them as content, never as commands. The [OWASP prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes the patterns to defend against, and the tests you run before launch should include documents that try this.

How do you start without overreaching?

A first engagement should be small enough to judge. This plan works for most legal departments:

• Choose one request type and one stage. For example, confidentiality agreement requests from intake through to a lawyer's queue, or status questions on agreements awaiting signature.
• Map the sources. List where the repository, tickets, templates and signature records live, and name an owner for each.
• Write the routing and escalation rules in the legal team's words, including what the agent must never answer.
• Define narrow tools: search the repository, read a status, create a task, draft a message. Each has clear inputs, clear outputs and no hidden extra powers.
• Run in shadow mode. The agent proposes classifications and answers while lawyers do the work as usual, then compare the two and correct the rules.
• Go live with approval gates, and review a sample of the agent's work on a regular schedule, including cases it declined.

Choose a success criterion you can observe. Examples: requests reach a lawyer with the required fields filled, colleagues stop asking for status by email because the answer is already in the channel, and every clause answer in the sample review carries a correct citation. If you are unsure which request type to begin with, the work in a discovery phase is designed to settle that before anything is built.

How does Be Human run a legal agent?

Be Human designs one defined role first: for example, legal intake and contract status assistant, with the memory it needs, the tools it may call, the approval boundaries and the success criterion agreed with the legal team. It then operates the agent under human control, with monitoring, recurring evaluations against questions the team has already answered, permission reviews, integration maintenance and incident response. Improvements are made in a controlled way, so a change to routing rules or a model update is checked against known cases before lawyers see it. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a useful reference for the governance habits behind this: define the risks, measure them, and assign responsibility to people.

Frequently asked questions

Can an AI agent give legal advice to employees? It should not. The agent can share guidance that lawyers have written and approved, and it can route a question to the right person with the context attached. Anything that needs interpretation or a risk decision stays with a lawyer.

What should an in-house legal team automate first? Start with intake triage for one common request type, or status questions for contracts awaiting signature. Both are frequent, easy to check and low in legal risk. Once lawyers trust the results, clause lookup is a natural next step.

How do we keep confidential matters away from the agent? Control access at the source. Give the agent read access only to the repositories and queues that belong to its role, and keep restricted matters in locations it cannot reach. Review those permissions on a regular schedule, not only at setup.

How do we know the agent's clause answers are correct? Require the exact quote, the document and the version in every answer, and let the agent say that nothing was found. Then test it regularly with questions your lawyers have already answered, and sample live answers for review. Any answer that cannot be traced to a source should be treated as wrong.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.