Legal Terms for AI Agents: Who Signs, Who Answers, Who Proves It
Your contracts describe two kinds of actor: people and accounts. An agent is neither. It is software that takes actions, and your terms have to say, in words a lawyer can defend, who it acts for.
What the current language assumes
Most terms of service define "you" as the entity holding the account. Access is granted to a user, an employee, a contractor. Liability attaches to the account holder. Audit rights are written around user identifiers. Each of those assumptions depends on an identifier that resolves to a person.
Agents break that resolution. When an agent signs in with a shared API key or a service account, the identifier on the request belongs to nobody in particular. The contract says "you". The logs say service-account-prod-3.
Four clauses that stop working
- Identity. A clause that binds obligations to a named user has nothing to bind to when the caller is software.
- Authority. Terms rarely state who approved a machine's access, on what evidence, and for how long.
- Evidence. Audit rights are only as good as the records behind them. A log line that names a shared key is not evidence of who acted.
- Exit. Termination language assumes you can end one person's access without ending anyone else's.
Authority needs a name, not a role
"Authorized personnel" is not an answer to the question a regulator, an auditor or an opposing counsel will ask. The question is which person approved this agent, what they were shown at the time, and how that approval can be withdrawn.
Write the clause so that every agent has a person who answers for it, and that person is confirmed at enrollment rather than asserted later. Delegation belongs in the same clause: an agent may hand work to another agent, but the second agent cannot hold more authority than the first.
Evidence and the standard of proof
A useful terms clause states what a record must show. Not that a call happened, but which agent made it, which person answers for it, what the exact operation and arguments were, and whether the request was permitted at the moment it ran. A timestamp and an IP address do not meet that standard.
Say plainly what the record is not. A model's own account of its reasoning is not evidence. Neither is a prompt log. The record that survives cross-examination is the one produced at the boundary, before the tool ran.
The exit clause
Revocation is a legal promise, not a feature. Your terms should state how quickly access ends, what remains valid until it expires, and what happens to work in flight. They should also cover departures: when the person who approved an agent leaves, someone else at the same accountable organization approves that agent again within a stated grace period, or the agent stops.
Pair that with a way back. Suspension followed by reinstatement after the host is checked again, with reduced access first, belongs in the contract as a described process, not as an improvisation.
Where the boundary sits
Be explicit about scope. Oathera does not verify people. It relies on your identity provider and your proofing sources, and it does not replace the identity provider. It does not judge model output, inspect prompts or rewrite them. Those remain your obligations, and your terms should keep them there.
How Oathera Helps
Oathera records your agent's DNA at enrollment: what it is, where it runs and who answers for it. An Ed25519 anchor key is created and kept on the agent's host, and the agent's identity is the key digest of that key. The host is observed and the host digest recorded, then compared at every token request. A named, verified sponsor is checked against your proofing sources.
Before the agent acts, a principal signs in at your identity provider with OIDC and PKCE, sees the requested and granted clearance, and approves. The enrollment record is written, and every action traces to an agent and the person who answers for it.
At runtime, short-lived identity tokens name a session key held in memory, and each call carries a single-use operation proof signed over the operation, its arguments and the token. Oathera Agent Enforcement checks the token and proof and asks the Oathera Identity Service for a signed, single-use boundary decision. No valid decision, no call. Services receive a gateway assertion in place of the agent's token.
For the exit clause, revocation refuses the next token request and the next boundary decision, and issued tokens stay valid until they expire. Deprovisioning a person stops every agent they approved. Handover lets another principal of the same sponsor approve each agent again. Agent revoked, suspended and host-changed events publish over SSF and CAEP.
For evidence, a hash-chained audit log replaces credential-like values with a correlation ID, carries signed seals, verifies the chain, and exports exactly once to a SIEM or write-once store. Clearance and risk tier are recorded with the decision, and rollout stages let you watch first, enforce for a few agents, then enforce everywhere.
Rewrite one clause first
Start with the clause that defines who a caller is. Become a design partner at https://oathera.ai/contact, or watch the simulator at https://sim.oathera.org.