In short
Use Copilot chat for work a person starts and checks: questions, drafts, one-off actions. Use triggers for work that must happen every time an event occurs, whether or not anyone remembers. Many good agents do both: they run on events, and people can ask them about what they did.
Two ways to start an agent
| In Copilot chat or Teams | On a trigger | |
|---|---|---|
| Who starts it | A person, when they think of it | An event: a new item, an email, a schedule |
| Who checks the result | The same person, immediately | Someone later, if at all |
| Whose identity it acts as | Usually the signed-in user | The connection the maker set up |
| Best for | Questions, drafts, judgement calls, one-off actions | Routine processing that must never be missed |
| Main risk | Inconsistent use; people forget it exists | Silent mistakes at scale; cost of runs nobody needed |
When chat is the right answer
- The task needs a human's context. "Summarise what is late in EMEA and draft a note to the owners" depends on who is asking and why.
- The output is a suggestion, not an action. A person reads it before anything changes.
- You are still learning what the agent should do. Every chat is a supervised test.
At the lowest autonomy level this is the natural home for an agent: it proposes, a person approves.
When it must run on its own
- Missing a case has a cost. Every new request must be triaged; every overdue escalation must be noticed.
- The rule is clear and testable. Complete or not complete; past the deadline or not.
- Writes are constrained. The agent can only change the fields you allow, through a controlled flow.
If the work must happen every time, do not make it depend on someone remembering to ask.
The pattern we recommend: both
Our intake and escalation agents run on triggers, because nothing should wait for a person to remember. The natural next step is to also publish them to Teams and Microsoft 365 Copilot, read-only in chat, so a delivery lead can ask "why was WI-24 escalated?" or "which escalations breached today?" and get an answer grounded in the same records.
The rule that keeps this safe: chat may explain and propose; only the triggered path may write, until the agent has earned more autonomy with measured results.
Licences, cost and governance
- Licences. To publish, the maker needs a Microsoft 365 Copilot licence or a Copilot Studio user licence with credits available. People you share an agent with need access to Microsoft Copilot; check the current licensing pages before you promise anything.
- Credits. Agents on the Microsoft Copilot harness are billed by usage in Copilot Credits, in chat and on triggers. Some chat usage by Microsoft 365 Copilot licence holders is not charged; triggered runs generally are.
- Channel policy. The Teams and Microsoft 365 channel must be allowed by the environment's data policy and agent channel settings.
- Sharing. Share with a named security group, not the whole organisation, while the agent is in test. Organisation-wide availability goes through admin approval.
- Authentication. Keep "Authenticate with Microsoft" on, so only signed-in colleagues can use it.
A quick decision guide
- Must it happen every time an event occurs? Yes: trigger. No: chat is enough.
- Does it change records? Yes: only through a constrained flow, and on the triggered path first.
- Would people benefit from asking it questions? Yes: also publish to Teams and Copilot, read-only, to a named group.
- Can you afford the runs? Estimate events per month and credits per run before switching on.
Sources
- Connect an agent to Teams and Microsoft Copilot, Microsoft Learn
- Share agents (Microsoft Copilot harness), Microsoft Learn
- Available channels for agents, Microsoft Learn
- Licensing for agents on the standard harness, Microsoft Learn
- Harnesses in Copilot Studio, Microsoft Learn