Personalized AI infrastructure
Who Runs AI Agents After Launch: Ops Handover for Your Company Layer
Learn who should own monitoring, permission reviews and incident response for AI agents after launch, so the company never depends on a single builder.

After launch, your own operations team should run your company's AI agents, with one named owner for each recurring duty: monitoring, permission reviews, incident response and approved changes. The original builder supports that team by tuning, repairing and advising, but does not stand in for it. Writing these owners down before go-live is what keeps the layer running when the builder is busy, away or no longer involved.
Key takeaways
- Name one accountable owner for each operating duty before go-live, not after the first incident.
- Separate the builder role (design, deep fixes, changes) from the operator role (watching, approving, reporting, pausing).
- Give the operator a written runbook for alerts, permission reviews and incident escalation.
- Train a second person to pause each agent, so no single person holds the only way to stop it.
- Review ownership every quarter, using the same evidence you collect for permissions.
Why one builder becomes a single point of failure
At launch, the people who built an agent know its instructions, its tools, the memory sources it reads and the edge cases that were tested. That knowledge rarely lives in a document, so it leaves with the builder. If one person is the only one who reads the logs, adjusts permissions and knows how to stop the agent, the company has a dependency it never chose.
Monitoring only helps when someone reads it. The OpenTelemetry documentation describes observability through signals such as logs, metrics and traces that show how a system behaves from the outside. Those signals are useful only when a named person reviews them on a schedule and knows what a normal week looks like. A dashboard with no owner is a record of problems, not a control.
Which duties need an owner, and who should hold them?
Four duties carry most of the operating risk. Assign each one to a role in your company before the layer goes live, and write down what the holder may decide alone.
- Monitoring belongs to an operations lead who reviews the agent's activity on a fixed cadence, decides whether an alert matters and escalates when it does. The builder tunes the thresholds, but the operations lead owns the response.
- Permission reviews belong to the business owner of the process each agent supports, with a second approver for any tool that can message customers, change records or move money.
- Incident response belongs to an on-call operator who can pause the agent without waiting for approval, plus a named technical contact on the builder side for repair.
- Change approval belongs to someone other than the person who drafted the change. The approver checks the scope, the evidence and the rollback plan before anything reaches production.
How do you write a handover runbook?
A runbook is a short document that lets a capable operator run the agent without reading the builder's notes. Keep it to a few pages and update it whenever the agent changes. Write these parts in this order:
- State the agent's purpose in one paragraph, including what it must never do.
- List every tool, data source and channel it uses, and the account that owns each credential.
- List each alert, who receives it and the first action the receiver should take.
- Describe the pause procedure in plain steps, then have the operator perform it once in a test environment, not only watch the builder do it.
- Name the actions that always require a human decision before they run.
- Set the review calendar and the evidence each review must produce, such as a signed list of granted permissions and a log of changes.
Running permission reviews without a specialist
Permissions drift. An agent receives access during a pilot so a test can run, and the access stays after the test ends. The OWASP authorization guidance rests on a simple principle: deny by default and grant only what a function needs. Apply that principle to agents. For every grant, the review should answer three questions: is it still needed, is it still limited to the smallest useful action, and would the business owner approve it today?
Illustrative scenario: a support agent received write access to a shared folder during a pilot so it could save conversation transcripts. Months later the transcripts go to a different system, and nobody remembers why the folder grant exists. The operator's review finds that the grant serves no current process, removes it and records the change with the reason. The review stays short because the list of grants was written down in the first place.
For a method you can follow step by step, see How to Run a Permission Review for Your AI Agents, Step by Step.
What happens in the first hour of an incident?
Incidents with agents are often ordinary: a wrong answer sent to a customer, a tool that fails repeatedly, or an input that tries to make the agent ignore its instructions. The response should be the same each time. Detect the problem through an alert or a staff report, pause the agent, preserve the logs and the inputs that led to the event, notify the duty owner and the people affected, and then decide with the technical contact whether to repair, roll back or keep the agent paused.
Prompt injection deserves its own line in the runbook. Text from a web page, an email or a document can contain instructions that an agent may try to follow. The operator should know which tools that agent can call after reading untrusted content, because that is where harm would occur, and those tools should have the tightest permissions.
The NIST AI Risk Management Framework is a useful reference for turning these habits into a routine, because it treats risk management as continuing work rather than a single approval.
How does Be Human's managed approach handle the handover?
Be Human designs, implements and operates personalized AI infrastructure for companies, and its managed operations can include monitoring, recurring evaluations, integration maintenance, permission review and incident response. The engagement states which duties stay with your team and which the managed service carries, so the handover is written down rather than assumed. Changes move through a documented request and approval path, and the service is delivered remotely in English and Portuguese.
Keeping ownership healthy after the first quarter
Ownership fades when nobody checks it. Put a standing item on the operations meeting: who holds each duty, whether the holder completed the last review, whether the runbook still matches the current agent, and whether any change went live without an approval record. Rotate a second person through the pause procedure every quarter so the backup is real. When someone leaves the company, transfer their duties in the same week and update the owner list.
Changes deserve the same discipline. The steps are laid out in How to Request and Approve Changes to a Managed AI Agent. For the early period, What the First Months With a Managed AI Agent Look Like describes the rhythm most teams settle into.
If you are not sure which of these duties your company should hand over first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Who should run AI agents after launch? Your own operations team should run them, with a named owner for monitoring, permission reviews, incident response and change approval. The original builder supports that team through tuning, repairs and advice, but the company should not depend on the builder to keep the agent working day to day. Ownership should be written into the runbook before go-live.
Can the builder stay involved after the handover? Yes, and usually should. The builder keeps responsibility for technical repair, configuration changes and evaluation design, under a written agreement. The difference is that the company's operators hold the decisions about permissions, pausing and approvals, so the builder's availability is not a requirement for daily operation.
How often should agent permissions be reviewed? At least quarterly for most agents, and sooner after any change to tools, data sources or the process the agent supports. Each review should confirm that every grant is still needed, limited to the smallest useful action and approved by the business owner. Keep the review record with the list of changes so the next review starts from evidence.
What if nobody on the team wants to own incidents? Assign the role anyway and give it a clear limit: the owner can pause an agent and escalate, while repairs go to the technical contact. Most people accept a bounded role with a written procedure more readily than an open-ended one. If the company cannot staff the role internally, a managed service can carry monitoring and incident response under a written agreement, while the company keeps the decisions that belong to it.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


