Copilot Agents · 7 min read
What Happens After the Agent Is Built? Why Every Agent Needs an Owner
By James Wilkinson 12 September 2026
An agent is a running service, not a finished project. What changes after go-live, the five-point minimum a firm with no IT team can run and who owns it.
TL;DR
- Software you buy gets patched by the vendor. An agent you build is yours, and it sits on a platform Microsoft changes monthly. Built does not mean finished. It means running, and running is a job.
- Six things change after launch, all routine: knowledge sources go stale, connections expire, answers drift, usage moves, the bill moves and Microsoft moves. None is a reason not to build. All are a reason to have a named owner.
- The whole discipline for a first agent is five points: a named owner, a one-page scope, least-privilege access, transcripts reviewed for ten minutes a week and a kill switch plus a test before any change. It is a weekly habit, not an IT function.
- The owner is usually a joint effort at first, then one person emerges, often the one closest to the process the agent took on rather than the most technical. That role is becoming a recognised thing, and Agent 365 is Microsoft's platform expression of it for firms that grow past a handful of agents.
- Embed is the retainer that does this running for firms that would rather it was ours: Starter £750 a month for up to two production solutions, Growth £1,500 for up to ten. New agents, workflows and integrations are always scoped separately.
It is a Tuesday morning, three months after go-live. The letters agent has been quietly producing engagement letter drafts from a form since June, and everyone has stopped noticing it, which is the highest compliment an agent gets. This morning it has produced nothing. The form submissions are sitting there. The Teams channel is empty. The person who built the agent left in August. The practice manager opens Copilot Studio for the first time since the handover and does not know what they are looking at.
The cause, when someone finds it that afternoon, is small. The connection the flow used to post into Teams had been set up under the builder’s account, and that account was disabled the day they left. Nothing was wrong with the agent. Nothing was wrong with the flow. A credential lapsed and nobody was watching.
That is the whole of this post in one scene. Software you buy gets patched by the vendor. An agent you build is yours, and it sits on a platform Microsoft changes monthly. Built does not mean finished. It means running, and running is a job. This is what the job consists of, who in a firm with no IT team ends up doing it and what it looks like when it is done properly.
What actually changes after launch
None of this is dramatic. All of it is routine, and all of it is fixable if somebody is looking.
Knowledge sources go stale. The procedures document gets updated and nobody tells the agent. Or it gets moved to a new library and the agent silently loses it, answering from whatever it can still reach.
Connections expire. A connector’s authentication lapses when the person who set it up changes their password, has their account disabled or leaves. The flow stops. Nobody notices until someone asks why the letters stopped arriving. Our automation maintenance post has been making this point about flows since April; agents inherit it.
Answers drift. Copilot Studio changes its default model, a harness update changes how instructions are interpreted, and an agent that answered well in June is vaguer in September with no error to show for it. This is exactly what a saved question set is for, and how do you know your agent still works explains how to build one.
Usage moves. People find uses the agent was never scoped for, ask it things outside its knowledge and get confident nonsense back. The design has not changed. The demand has.
The bill moves. Credits scale with use. A quiet agent that becomes popular can triple its consumption without anyone deciding to spend more, and a spend threshold set for month one will stop it in month four. The three numbers to check before go-live are the same three to re-check every quarter.
Microsoft moves. This year alone: billing moved to credits, the default harness changed on 3 August, agents built before that date moved to credit billing on 1 September, seeded AI Builder credits go on 1 November and new models keep arriving in the picker. None of it is a reason not to build. All of it means someone has to read the change and decide whether it touches the firm.
What “operations” means when there is no IT team
The word sounds like a department. For a first agent it is five things, and none of them needs a technical background.
- A named owner. One person accountable for what the agent does, who checks it and who gets called when it misbehaves. Not a committee, and not “someone in ops built it”.
- A written scope. What it does, what data it can touch, who can use it and what it must never do. One page. This is what you point to when someone asks whether it can also do something else.
- Data access on least privilege. The agent only sees what it needs. In Copilot Studio that means checking that the connections and knowledge sources actually match the scope, and that the agent cannot reach anything the person asking could not reach themselves. Where your data goes explains why that inheritance rule is the backbone of the whole security answer.
- Logging you will actually look at. Conversation transcripts switched on, and a weekly ten-minute review of what people asked and where the agent got things wrong or refused. This is the feedback loop, and it is the evidence if something ever goes wrong.
- A kill switch and a test before changes. Know how to turn the agent off in under a minute. Never push a change straight to the live version without running a handful of known questions through it first.
That is the whole job for a first agent. It is not an IT function. It is a ten-minute weekly habit plus knowing who to call. The Agent Governance Toolkit post describes the other end of the same spectrum, where rules are enforced in code rather than by a person with a checklist. Most firms of 30 never need that end. They need this one.
Who the owner turns out to be
In firms without an IT team, it is usually a joint effort at first. The practice manager knows the process, the office manager has the admin rights, a partner asked for the agent in the first place, and for a month or two they cover it between them. Then one person emerges. They start opening the transcripts without being asked. They notice when an answer is slightly off. They are the one who says “that document moved, has anyone told the agent?”
That person is often not the most technical member of staff. They are usually the one who was closest to the process the agent took on and who cares whether it keeps working. Agent management becomes part of their role rather than a job they were hired for, and it is a role they gain, not one they lose. It is becoming a recognised thing, and Microsoft’s Agent 365 is the platform expression of it: a control plane for inventorying and governing agents, generally available since May 2026, at $15 per user per month on an annual plan, about £11, or included in Microsoft 365 E7. It needs Business Premium or E5 underneath and is licensed for everyone who interacts with, owns or manages an agent. A free foundational tier gives every Microsoft Cloud customer agent identity and an inventory.
For a 30-person firm with two agents, Agent 365 is probably not yet worth the licence. The five points above do the job. It becomes relevant when the agent count grows past a handful and the owner needs tooling rather than a checklist. Say so to anyone who tries to sell it to you earlier.
What a month of support looks like
For firms that would rather the running was ours, that is what Embed is. Every tier has three strands.
Agent operations. Answer quality checked, knowledge sources kept current, failed answers reviewed, and every agent confirmed to still have a named owner and a working escalation route. This is points one to four of the list above, done on a rhythm.
Optimisation and new use cases. Instruction tuning, widening a job that has proved itself and a backlog of the next builds worth doing. The build work itself is scoped separately.
Async support. A place to take questions, answered within two working days on Starter and one working day on Growth. For route two firms this is the same rhythm continuing after the 90 days included in the build price.
The edges are drawn on the Embed page and they are worth repeating, because a retainer is easier to trust when you know what it will not do. Included: prompt optimisation, knowledge updates, connector troubleshooting, authentication renewal, fixing existing solutions, usage and answer quality reviews and recommendations for improvement. Not included: new agents, new workflows, new integrations, new connectors, new Azure services, new business processes, workshops and training. Anything new is scoped separately, every time.
Starter is £750 a month for up to two production solutions. Growth is £1,500 for up to ten. Enterprise is scoped per engagement. Three month minimum, reviewed quarterly. One solution means one delivered capability, however many flows, tables, connectors and Azure components sit inside it. The letters agent from the opening scene, with its form, its flow and its Teams channel, is one solution. Where the retainer sits in a full year’s budget is set out in what a first agent costs a 30-person firm in year one.
The fair alternative
A firm can run this itself. Route two exists for exactly that: the firm’s own people build, delivered through Enable with two delivery days, two build reviews and 90 days of support, held to the standard the Automate page sets out, and the five-point list is the whole discipline afterwards. What it costs is one person’s attention every week and someone reading Microsoft’s release notes every month.
Some firms have that person. Most firms of 30 do not, and that is the honest reason the retainer is priced the way it is. The work exists whether or not anyone is paid to do it. The only question is who.
Where this leaves you
If your firm has an agent live, or is about to sign for one, the question to settle now is not whether it will need running. It will. The question is who, and what happens on the Tuesday morning when it stops. If the answer is a named person with the five-point list, good, and the Enable page describes the route that builds that capability in-house. If the answer is nobody yet, Embed is built for that gap, and a free consultation is enough to work out which tier fits what you have running.
Sources checked
Last checked: 12 September 2026. Microsoft prices are US dollar list at September 2026 and subject to change.
Related reading
More on Copilot Agents
Common questions