Handing work to an AI agent is closer to hiring than to buying software. You are trusting it with customers, money or records, so the questions are the ones you would ask about a new team member: what will it do, how will it know our rules, and how will we know it is doing a good job?
Here are the ten we would ask in your place.
1. What exactly will it do, and what won't it do?
Get the job described in one or two sentences, with clear edges. "Answer questions about order status and returns; never discuss billing disputes" is a job. "Handle customer service" is not.
2. Whose rules does it follow, and where are they written?
The agent should follow your policies: your refund limits, your approval thresholds, your tone of voice. Ask where those rules live and who can read them. If the answer is "inside the prompt", changing a policy will mean asking the provider to edit text you cannot see.
3. What does it do when it isn't sure?
This is the most important question. A good agent knows which cases it must not settle alone and hands them to a person with the full history, so nobody starts over. Ask to see what that hand-off looks like.
4. Can we see what it did?
You should be able to look at any task and see what came in, what the agent looked up, what it did and what it replied. Ask whether that record is available to your team directly, or only through the provider.
5. What can it reach in our systems?
An agent needs access to do useful work, but only to the systems it needs. Ask what it can read, what it can change, and whether anything stops it reaching systems it was never connected to.
6. Can it see our passwords?
It should not. The safest setups keep credentials encrypted and hand them only to the connection that needs them, never to the agent's reasoning, where a cleverly worded message could coax them out.
7. What happens to our customers' data?
Ask what is stored about each task, for how long, and whether you can limit it. Conversation history is useful, but it is also personal data, and your privacy rules should decide how long it is kept.
8. How are changes made?
Policies change. Ask how a change reaches the agent, who approves it, and whether you can see what changed and go back if something goes wrong.
9. How is it priced?
Understand what drives the cost: the number of processes, the volume, the systems connected and the AI models used. Ask how AI model usage is billed, and whether you can see what each task costs.
10. How will we know it's working?
Agree up front on what good looks like: how many cases it handles end to end, how many it hands off, how fast, and how often its answers are corrected. Then make sure you can see those numbers yourself.
How Aether answers them
Aether builds, runs and tunes AI agents for businesses. Each agent has a defined job; your rules are kept in Calibration and changed as new, noted versions; anything outside them goes to your team; every task is recorded step by step; passwords stay encrypted and out of the agent's reach; and your team works with its agents in your portal. Book a demo and ask us all ten.