Intake
Lead follow-up and case-fit prep
Agents prepare the next step after the Managed Firm Brain answers: summarize facts, flag fit issues, draft follow-up, and queue the handoff for staff review.
Managed AI Agents
Agents are jobs inside one brain — not a catalog you babysit.
Intake follow-up, records, drafts — those jobs run inside one brain on the systems you already have. Not a pile of agents for sale.
Stay in ChatGPT or Claude; boxed legal-AI apps cannot match those desktop apps. Same team can run the website, SEO, ads, portal, and onboarding. People keep the phones.
Updated October 31, 2026 — added what the MCP spec does and does not require for approvals, with sources
What agents run first
Start where the work is repetitive, measurable, and safe to supervise.
The best managed agent is not the flashiest one. It is the one that saves time every week and can be checked before it acts.
Records
Medical records and document follow-up
Track requests, surface stale items, draft follow-ups, and keep the matter timeline visible without asking staff to rebuild the same list every week.
Client communication
Draft updates without auto-sending
Managed agents read the file and draft status updates or next-step messages. Humans approve before anything leaves the firm.
Operations
Dormancy, dashboards, and task cleanup
Agents find stuck work, explain why it is stuck, and recommend the next move so managers can run the firm from facts instead of memory.
Managed-service model
Managed agents need an operator loop, not a launch-and-leave install.
FirmOps runs agents like an operating system: start narrow, monitor the work, and tune the workflow before expanding.
Design
Pick one recurring workflow, define the safe read paths, and decide which actions must stay human-approved.
Deploy
Connect the agent to approved context, run it read-first, and compare its recommendations against staff expectations.
Monitor
Review outputs, missed edge cases, escalation patterns, and whether the workflow is saving manager time.
Tune
Update prompts, policies, integrations, and approval thresholds as the firm changes.
Capability proof
MCP and integrations are how the agent reads the real firm.
MCP is not the client-acquisition headline. It is the capability proof: the integration layer that lets supervised agents read approved firm context across Clio, Dropbox, Gmail, documents, and reporting tools before recommending action. For the deeper company-brain pattern, read the Claude Desktop MCP guide in the FirmOps operator notes.
- Read-first by default: answer from firm context before writing anywhere
- Approval-gated writes: client sends, Clio updates, and task changes stay supervised
- MCP and integration builds as capability proof: connect Claude to Clio, Dropbox, Gmail, and reporting tools
- Operator-led rollout: decide what to automate first, what to leave human, and how to measure lift
Protocol vs. operator
MCP recommends a human in the loop. It does not enforce one.
The MCP tools specification describes tools as model-controlled and leaves the approval experience to whoever builds the client. Its safety language is mostly SHOULD, not MUST. A managed agent turns each recommendation into a rule the firm can check.
| What the MCP spec says | How FirmOps runs it |
|---|---|
| There "SHOULD always be a human in the loop with the ability to deny tool invocations."[1] | Every client send, filing, and system write waits for a named person to approve, reject, or retry it. |
| Clients "SHOULD" prompt for user confirmation on sensitive operations and show tool inputs to the user before calling the server.[1] | Drafts land in the requesting staff member's email drafts or an approval queue with the full proposed change visible. |
| Clients "MUST consider tool annotations to be untrusted unless they come from trusted servers."[1] | Read-first is enforced by the credentials the agent holds, not by a tool calling itself read-only. |
| Clients "SHOULD" log tool usage for audit purposes.[1] | Every read, draft, and approved write is logged so a manager can see what the agent touched and why. |
For the systems agents connect to, see law firm AI integrations. For the written rules behind the approval gate, see the law firm AI policy guide and the read-first definition.
See it live
Watch a managed AI agent working in the wild.
jonny.bot is a public-safe, chronological log of what a FirmOps-managed AI agent actually does day by day. Same pattern this page describes: read-first, human-approved, never sends on its own.
Legal document automation
A practical first managed-agent use case.
Legal document automation is one of the cleanest ways to prove the model: read the file, prepare the draft or packet, and route the result for review before anything is sent or filed.
Common questions
Straight answers.
What is a managed AI agent for a law firm?
A supervised AI worker that runs one recurring firm task end to end: it reads across your systems, prepares the work, and drafts it for approval. "Managed" means FirmOps builds it, monitors it, and tunes it, so it is not a tool you have to babysit.
How is a managed AI agent different from ChatGPT or a chatbot?
A chatbot answers in a tab with no memory of your firm. A managed agent is connected to your case system, documents, email, and billing, runs a specific workflow on a schedule, and keeps every client-facing or system-changing action behind a human approval gate. MCP is the protocol FirmOps uses to attach those tools. Scoped permissions, logging, and approvals are how FirmOps operates the connection, not guarantees of MCP itself. What is MCP for a law firm?.
Do the agents do anything without a human?
They read and draft on their own. They do not send, file, or change the record on their own. Drafts land in the requesting staff member's email drafts or behind approve, reject, and retry controls. A person always releases the work.
Does MCP itself require a human to approve what an agent does?
No. The MCP specification (version 2026-07-28) says there SHOULD always be a human in the loop with the ability to deny tool invocations, and that clients SHOULD prompt for confirmation on sensitive operations. SHOULD is a recommendation, not a requirement, and the protocol says it does not mandate any specific user interaction model. The approval gate is something the operator has to build and keep turned on. That is the managed part.
What should the first managed agent be?
One recurring bottleneck you personally dislike doing: dormancy digests, records follow-up, client-update drafts, or a weekly owner report. Prove one workflow read-first and approval-gated, then expand from there.
Next step
Bring one recurring bottleneck. We will show it on the brain.
The demo shows how FirmOps decides what should run as a supervised job, what should stay human, and how approval gates keep the work safe.
Put a number on it
Every agent here replaces a subscription or a task. Price the subscriptions with the software cost calculator. It starts from a real firm's stack, tags each line Stays or Consolidates, and nets the $500 a month FirmOps maintenance; what that covers is on the pricing page.