Fear 1
It emails a client something wrong.
Client dispatch is a gated capability. The AI can draft the update, but it cannot send it. A lawyer approves first, and the approval is recorded.
For law firms
Use AI in your practice without losing control of it.
Your team drafts client updates, runs intake, extracts deadlines, and reconciles billing all day. Nothing reaches a client, a calendar, or a trust ledger without a lawyer's approval.
And every action leaves a signed record that proves a lawyer stayed in control. Not proof that you used AI. Proof that you supervised it.
The three things you are afraid of.
Most firms are not afraid AI gives a wrong answer in a chat window. They are afraid of what happens when it acts on its own.
Fear 2
It misses a deadline.
Deadline extraction runs on every matter document, and a calendar write is gated behind your approval. You get the date and the proof it was surfaced, not a silent miss.
Fear 3
Client confidences leak.
AgentGuard runs on your own model key and never sees the content of a matter. It governs what the agent is allowed to do, not what it reads. Your data stays in your runtime.
Proof of control, not proof of AI
The receipt is built to defend you, not expose you.
A chat history shows that you used AI. A supervision receipt shows that a lawyer stayed responsible: who reviewed the work, what capability the agent was granted, whether a human approved before anything left the firm, and which policy was in force. It is signed, tamper-evident, and verifiable by anyone you hand it to, with the content itself never stored.
It maps to the duties the ABA's 2024 guidance puts on a lawyer using generative AI:
CompetenceWhich model and policy version did the work.
ConfidentialityContent never leaves your runtime. Hash only.
SupervisionReviewer identity and approval, on the record.
CommunicationNothing sent to a client without sign-off.
Reasonable feesExact cost of each action, itemized.
What the agent can and cannot do.
Every teammate starts read-only. Anything that touches a client, a calendar, or a trust ledger has to earn a higher capability, and earning it is recorded.
| Capability | What it allows | What it cannot do |
|---|---|---|
| READ_ONLY | Classify, summarize, extract deadlines, reconcile billing, prepare a packet. | Send a client email. Touch the trust ledger. |
| TRANSACT | Prepare a dispatch-ready draft that still waits for a lawyer's approval. | Move trust funds. Change a policy. |
| ADMIN | Trust-account actions, retention changes, capability grants. | Dispatch to a client without review. |
The human stays the lawyer of record. AgentGuard is an audit and supervision layer. It proves who did what and keeps the agent inside its lane. It does not give legal advice, does not practice law, and does not guarantee an outcome. Review the operating pack with your own counsel before production use.
Let your firm automate the work it has been afraid to.
Pick law, and your AI team is built and working in about a minute. Free to start, no card, and you can kill any agent in one click.