LanternOpsField Guide · No. 001Managed AI Teams
LanternOps Book a diagnosis →
§ 0The operating model

From AI experiments to managed AI operations.

There is a real difference between a person using ChatGPT and a firm operating a supervised AI Team. This is how the second one works, and why the system around the agent is the part that matters.

01
§ 1The ladder

Six rungs from a chat window to a managed operation.

Each rung is more useful than the last, and each one needs more management. Most firms stop at rung one and wonder why nothing changed at the level of the business.

Category ladderAI chat → Managed AgentOps
  1. 01 AI Chat A person asks ChatGPT or Copilot for help. Useful for individuals. Disconnected from operations.
  2. 02 Automation Fixed rules move data between tools. Good for predictable steps. Brittle around ambiguity.
  3. 03 AI Workflow AI is embedded in one business process. Helpful, but often one-off and project-based.
  4. 04 AI Agent AI performs a defined role with tools. Useful, but narrow without supervision.
  5. 05 AI Team Specialized agents coordinate around a function. Needs an operator, permissions, evals, monitoring.
  6. 06 Managed AgentOps Ongoing management of supervised AI Teams. This is the LanternOps lane.
02
§ 2The build

One workflow, designed and operated end to end.

  1. Start with the work, not the model

    We map one high-friction workflow: where information enters, who touches it, which systems it crosses, and where it gets stuck.

  2. Design the AI Team

    Each agent gets a job description, tools, permissions, outputs, and a review cadence. Sensitive actions get an approval gate.

  3. Connect to real systems

    Secure connectors to email, documents, CRM, calendars, and billing, scoped to exactly what the team is allowed to touch.

  4. Supervise and improve

    We monitor the work, manage the review queue, tune the agents, and report on what changed, month after month.

03
§ 3The artifacts

Six artifacts we use to make supervision concrete.

Each one is a working document: how we design, explain, and operate every AI Team, and how you stay in control of what it does.

Artifact 01Workflow map
Inbox Documents CRM Calendar Billing Client msg Owner
Make hidden operational drag visible. Disconnected systems, manual handoffs, and the single managed path that runs through them.
Artifact 02Agent card

Drafting Assistant

Reports to Lead attorney

AI Team
Tools
  • Docs
  • Templates
Permissions
  • Read matter file Allowed
  • Draft documents Draft only
  • File with court Never
Outputs
  • First-draft motion
  • Cite check list
Edit rate 6%
Weekly review
Every agent is a managed role: job description, tools, permissions, outputs, performance, and a review cadence.
Artifact 03Permission map
  • Allowed
  • Controlled actions
  • Approval-gated
  • Never allowed
What the team can read, draft, do under approval, and never touch. Boundaries you can see.
Artifact 04Approval gate
AI-prepared drafts Human approval Approved work Escalation
Sensitive actions pause until a professional approves. AI prepares. Humans decide.
Artifact 05Review queue
Review queue 2 awaiting approval
  • Intake summary — Reyes matter AI draft Pending review
  • Client follow-up email Approved 9:42 Approved
  • Deadline calendar update 1 edit Edited & approved
  • Conflict check — new party Needs review Exception
  • Meeting notes → tasks AI draft Pending review
AI-prepared work waits for human review, then leaves as approved, edited, or escalated.
Artifact 06AgentOps dashboard
Managed AgentOps example data Operating
Items handled 1,284 this month
Hours returned 96 to senior staff
Awaiting approval 7 in queue
Edit rate 6% down from 11%
Edit rate trend lower is better
Stuck points
  • Ambiguous matter type on 3 intakes
  • Billing connector token expired
  • New client template not yet mapped
Connector health
  • Email
  • CRM
  • Docs
  • Billing
  • Calendar
The operating view after launch: work handled, approvals waiting, edit rates, stuck points, and connector health. Proof that someone is operating the system.
§ 4The point

The value is the managed system around the agent.

Tools, permissions, review queues, logs, quality checks, escalation, and ongoing improvement. That is the part that keeps an AI Team useful, safe, and worth paying for.