Moving a business task into chat should change how someone requests the work, not who has authority to approve it. Before connecting an AI assistant to a system that can change records, ask your vendor to demonstrate the same assignment under different employee permissions. The useful test is whether the work gets easier without quietly widening anyone’s authority.

That is the operational question inside OpenAI and ServiceNow’s October 28 announcement. The companies describe an integration through which someone could investigate an IT incident and update a ServiceNow record from ChatGPT. OpenAI says customers choose which capabilities to expose, while ServiceNow applies existing business rules and permissions.

In brief: A conversation can become a new way to request business actions. Keep the authority for those actions in the business system, assign an owner to the workflow, and test what happens when a user asks for something outside their role.

The important feature is what stays put

Less switching between applications sounds good. But I would spend most of the demonstration on a less visible promise: according to OpenAI, ServiceNow’s Action Fabric applies the organization’s existing configuration and access when actions come through ChatGPT.

I would treat that as something to verify in your environment, not as a reason to skip the demonstration.

Take a replacement laptop. In this hypothetical company, the employee submits the request, a manager approves the expense, and procurement places the order. Three people have a part in getting that laptop to someone’s desk, even though the employee can describe the whole job in a sentence.

“Replace my laptop and get it approved” expresses an outcome. It should not confer approval authority on the person typing it.

A useful assistant could gather the device information, prepare the request, and route it to the manager. My acceptance test would require it to stop at that handoff rather than turn a convenient sentence into a completed purchase.

Ask for the ordinary employee’s demonstration

For the buying conversation, I would request a controlled test with an ordinary employee account and an approver account. Use a harmless sample record, not a live purchase.

First, have the employee request an allowed change. Inspect the record afterward. Then have the same employee request a change reserved for the approver. The expected result should be a refusal or a routed approval request, depending on your process. Finally, let the authorized person approve it and inspect what changed.

Ask to see the actor recorded against each action. Require evidence that the record distinguishes a proposed change from an approved one. If the vendor cannot show that distinction, keep the pilot read-only until it can.

That is the demonstration I want when a vendor says “guardrails.” Show me where the employee’s authority ends.

OpenAI’s administration documentation separates who can use an app, what actions it can perform, and when ChatGPT asks permission. It also warns that some apps do not offer individually configurable action controls. Ask which controls your proposed connection actually supports. Do not accept an account-wide setting as proof of how a particular action behaves.

Put the process owner in the room

My recommendation is to make the person accountable for the business process own the acceptance test. IT should establish the connection and verify access, but it should not have to invent the purchasing policy while configuring an assistant.

For equipment requests, have the process owner define a complete request and name who approves exceptions. Then make the demonstration awkward. Ask for a purchase above the employee’s limit, leave the approver unassigned, or try opening a record that employee should not see.

Count the work that remains for people. If the assistant prepares a request quickly but the manager must reconstruct its reasoning before approving it, that review belongs in the result. Time the job through approval, including the manager’s review, and compare it with the existing process under the same rules.

This is the same production discipline behind testing whether an AI assignment can resume after a restart. Decide what a completed job means before expanding the assignment.

Buy the available workflow, not the whole roadmap

The rollout dates matter. OpenAI’s availability schedule says the ServiceNow plugin became available on October 28, with expanded capabilities expected by the end of 2026. Voice capabilities are expected in the first quarter of 2027. Security capabilities are expected to enter early access in January 2027, with general availability planned for the second quarter. Those are different purchasing horizons, not one finished package.

There is reason to take this shift seriously without treating every promise as delivered. In its October 28 financial results, ServiceNow reported year-to-date production consumption of agentic AI up more than 13 times year over year. That is a vendor-reported usage measure, not evidence of a comparable productivity gain for your team.

For this week’s decision, ask for one available workflow, the permissions it uses, and the evidence left after an action. Approve expansion only after the ordinary employee’s test works as intended.

The employee should get a laptop with less chasing. The manager still has to approve it.

Research and draft: Mai. Direction and voice framework: John Lipe. Last updated: October 29, 2026.