How businesses can test AI agents before giving them more independence
For small business owners, a routine service request can set off several tasks. At an HVAC company, for example, someone has to look up the customer, check the service history, check availability and book an appointment. An AI agent could help with that, but a human needs to stay in the loop and make sure it does the right thing.
Below, Verizon outlines a three-stage approach — crawl, walk, run — to test that workflow and determine which steps need less oversight. This approach aligns with guidance from the Open Worldwide Application Security Project (OWASP), a nonprofit focused on software security.
Before testing, define what the agent can access
Consider this hypothetical message to an HVAC business:
“Our upstairs AC stopped working last night. It’s 82 degrees up there. We had someone from your company out last summer, but I don’t remember who. Can someone come tomorrow afternoon?”
To help respond, the agent needs limited access to customer records, the service history and the technician’s calendar.
It does not need payroll records, bank accounts, payment-card data or permission to issue refunds. Those systems and permissions fall outside the assignment.
Define access by writing a job description for the agent. That job description spells out its responsibilities, the information it can access and where its authority ends. It also identifies where the agent’s work must stop for human review.
OWASP also recommends limiting an agent’s permissions in the business systems it uses — for example, allowing it to book appointments without giving it access to issue refunds.
Crawl means reviewing every important step
During the crawl stage, the agent can perform much of the work, but a person reviews each important action before it goes on to the next text. That human oversight makes it possible to identify errors and correct the workflow while oversight remains close.
For the hypothetical HVAC request, the crawl review could follow this sequence:
The agent identifies the customer.
Check: Did it pull the right customer?
The agent pulls last summer’s service history.
Check: Is this the right visit?
The agent drafts a response.
Check: Is the information accurate? Would I actually send this?
The agent finds an available appointment.
Check: Is that time really open?
The agent prepares the booking.
Check: Right customer, right time, right technician?
This stage is about checking whether the agent performs the work it was assigned, and where corrections are needed. Verify and improve the workflow until it consistently produces the same result.
Walk means reducing checks on reliable, lower-risk steps
Once two conditions are consistently met — the agent consistently produces the right result, and the consequences of a mistake at that step are manageable for the business — it’s on to the next phase: walk.
Frame this stage around a practical question: Which checks are still protecting the business, and which checks can be removed?
The agent identifies the customer. You don’t need to check.
The agent pulls service history. You don’t need to check.
The agent updates the CRM. You don’t need to check.
The agent drafts the customer response. You still check.
The agent proposes an appointment. You still check if the schedule is unusual or the request is outside the routine.
Run means routine work stays within tested boundaries
During the run stage, that tested routine can work without a person approving every step. But it’s clear that any new or higher-risk requests will still get flagged to the human before the agent takes further action.
For example, if the customer message adds an additional request beyond the scope of that agent’s job description:
“… Also, the technician charged us last time for something we don’t think was fixed. Can you refund that?”
The refund request falls outside the service-and-appointments assignment. The agent has no authority to issue a refund, and the request should be passed on to human review before any refund action occurs.
The same principle applies to billing disputes, unusual pricing decisions or other requests that require authority beyond the work already defined and tested.
The important question during the run phase is whether routine work stays within its assigned boundaries and exceptions reliably reach a person. An agent that completes straightforward bookings but continues into unauthorized decisions has not met that standard.
The goal is a workflow whose limits are understood: The business knows which actions can happen automatically, which still require approval and where the work must stop.
Tia Hopkins is founder and chief execution officer of CTRL+EXEC, where she advises organizations on AI adoption, cyber resilience and turning strategy into execution.
This article was produced by Verizon and reviewed and distributed by Stacker.
