The short answer
An AI agent is software that takes a goal in a business process, reads the relevant information in your systems, decides on the next step within rules you set, carries it out or drafts it, and asks a person when the step matters. It is closer to a trained assistant working through a queue than to a chatbot.
The parts of an agent
| Part | What it is | What it means for you |
|---|---|---|
| Trigger | An event such as an email arriving, a lead created or a period ending | You choose what starts the agent |
| Instructions | The goal, the rules and the limits you give it | Your policies become the agent's behaviour |
| Knowledge | The data and documents it may read, such as records in Dynamics 365 | An agent can only use what you let it reach |
| Actions | The things it may do, such as draft a reply, match a transaction or create a record | You decide what is read-only and what can change data |
| Approval | The step where a person reviews and confirms | Risk is controlled where it matters most |
| Audit trail | A record of what it read, decided and did | You can show an auditor why something happened |
A worked example
A supplier emails that a delivery will be late. The Supplier Communications Agent reads the email, finds the purchase order it refers to, notes the new date, drafts an update for the planner and, if you allow it, a reply to the supplier. The planner reviews the draft and approves it. The system records what the agent read and what the planner approved.
Where the controls sit
- Identity and permission. The agent acts with the access you grant, in your Microsoft tenant.
- Data boundary. It works on the data you connect and does not need to move it elsewhere.
- Approval design. Money movements, customer-facing messages and changes to master data are approved by a person until the results justify otherwise.
- Monitoring. Administrators can see what agents are doing, and can stop them.
What can go wrong
- The agent misreads a document or an unusual case. Review catches it, and the exception teaches you what to add to the instructions.
- The data behind it is poor. The agent then makes poor suggestions quickly. Fix the data first.
- Nobody owns it. An unowned agent drifts. Name an owner for each one.
Frequently asked questions
What is an AI agent in Dynamics 365?
Software that takes a goal in a business process, reads the relevant data, takes or drafts the next step within rules you set and asks a person to approve the steps that matter.
Can an agent change our data on its own?
Only if you allow it. You decide what an agent may read and what it may change, and you can require approval for any action.
How do we know what an agent did?
Each agent leaves an audit trail of what it read, decided and did, and administrators can monitor and stop agents.