When AI industry leaders call for development to slow down over safety concerns, it is reasonable to wonder what that means for your own business. CNA’s September 14 report examined those concerns and the prospect of stronger testing and tighter limits on agents.

For an owner considering an AI assistant, one practical question follows: what will it be allowed to do with your business tools?

An agent can use connected software to take steps towards a goal. If that concept is new to you, our plain-English guide to how AI agents work explains it through a customer enquiry.

Here, we will use a reply-drafting assistant to explain three limits worth setting. These controls are often called guardrails. They help reduce mistakes and contain their consequences, but they do not make a system infallible.

1. Give it a clear job

“Handle customer enquiries” leaves plenty of room for interpretation. Does that include offering discounts or promising a delivery date?

Give the assistant standing instructions, often called a system prompt:

Prepare a reply using our approved product information. Flag missing prices or delivery dates for a member of staff. Do not invent an answer or offer a discount.

That tells the AI what is expected. It cannot guarantee the AI will comply. Important limits also need to be enforced by the software around it.

2. Give it only the access it needs

This assistant needs the enquiry and relevant product information. It has no reason to access your bank account or delete customer records.

Permissions determine which information it can read and which actions it can perform. Reading a record does not have to include permission to change it.

If the job is drafting replies, keep sending unavailable. Ask whoever sets it up to check the actual access granted by every connection, including whether a broadly authorised account gives it another way to send or change records.

3. Make it stop before the important action

If you later add sending, put an approval checkpoint before it. Show a staff member the recipient and exact proposed reply. Send only after approval, and ask again if either changes.

The workflow should enforce this pause. It should not depend solely on the AI remembering to ask.

Place review where a mistake matters, such as making a customer commitment. Routine product lookups need not interrupt someone each time.

Three questions before you start

  • What can this assistant read?
  • What can it change or send?
  • Which actions require a person’s approval?

Have the implementer demonstrate those limits with sample cases, including missing information and unusual requests. Keep a record of actions and a way to pause the system.

Thinking of connecting AI to your business tools? Tell us what you want it to do, and we can work through the access and checks it needs.

Technical sources: Anthropic on enforced access boundaries; n8n documentation on approval before tool actions.