AI ENGINEERING

What should an AI agent be allowed to do in your business?

The useful question about an AI agent is not how clever it is. It is what it is allowed to do. An agent that drafts a reply is a different risk from one that sends it, and the difference should be a decision you make on purpose.

Sort actions by consequence

List what the agent could do in the workflow, then sort each action by what happens if it gets it wrong.

  • Reading and summarising: low risk, usually safe to automate
  • Drafting for a person to review: low risk, high value
  • Routine, reversible actions such as categorising or scheduling: allow within limits
  • Financial, contractual or customer-facing commitments: require approval
  • Anything irreversible or outside its purpose: do not allow

Give it only the access it needs

An agent should see the data its task requires and no more. Access follows the role, as it would for a member of staff, and the tools it can call are an approved list, not everything the system can do.

Keep a person where it matters

Approval steps are not a failure of automation. For sensitive actions they are the design. The agent prepares the work, a person confirms it, and the record shows who decided.

Make it traceable

You should be able to see what the agent was asked, what information it used, what it did and what it cost. Without that record you cannot improve it, and you cannot explain it when someone asks.

Measure before you widen the limits

Start narrow. Agree what good looks like, test against real examples and widen the agent’s permissions only when the results support it.

An example from customer support

In SoAppy, the HelpDesk uses AI to sort and summarise tickets, find similar past cases and draft replies in context. Staff review the draft before anything is sent, and their corrections improve later suggestions.

YOUR NEXT CHAPTER STARTS HERE

Have a good problem?
Let’s solve it.

Tell us what you’re building, changing or figuring out.

Let’s talk