The AI Pattern: Right or Wrong?

Name the decision before an automated process takes its next step.

An AI agent can now carry a task through several steps. That sounds useful because it is useful.

It also changes the question. Once a system can act, someone has to decide what it may do, what evidence it must show, and when a person needs to step in.

The most useful place to start is the decision the workflow touches.

On October 22, OpenAI announced that Uber was rolling out ChatGPT Work to more than 30,000 employees, following a pilot with 1,000 people. The announcement describes an internal skills library, approved integrations, and training. Those are choices about how the work is governed. The announcement does not yet report measured outcomes.

The Pattern Beneath The Demo

A demo shows the happy path: the agent reads, reasons, writes, and completes the task.

Real work has a second path. A source is missing. A name is similar to another name. A client asks for an exception. The tool presents an answer that sounds certain but rests on the wrong file.

That is where the operating rules matter. Speed makes a workflow visible. Exceptions reveal whether it is ready.

Three Decisions Before You Turn It On

Box 1: What can it prepare?

Drafts, summaries, and comparisons can save a person time. Decide what evidence the draft must carry so the reviewer can check it.

Box 2: What can it change?

Updating a record, sending a message, moving money, or publishing a page affects someone outside the drafting process. Name each action and the person who approves it.

Box 3: What happens when it is unsure?

Give the system a clear way to stop, show the gap, and hand the decision to a person. “I could not verify this” is a useful result. A confident guess can create a week of repair work.

The distinction is simple. The machine can carry more of the process. A person still owns the meaning and the consequence.

A Follow Up Before It Leaves the Draft

Consider a hypothetical agent helping prepare a client follow-up. It may summarize approved meeting notes and draft the email. Before the draft reaches the reviewer, it should identify the source notes, show which commitments it found, and flag any missing date or unclear responsibility.

Suppose the notes say, “Discuss revised scope with client.” The agent should not turn that into, “Your revised scope is approved.” It can prepare the question for review. It needs evidence before treating the discussion as a decision.

The reviewer can then ask: Does the draft reflect what was agreed? Has it added a promise? Is the recipient correct? Who is authorized to approve the next action?

That small example turns a broad instruction to “use human oversight” into a task someone can perform. The agent prepares the work, the reviewer resolves the uncertainty, and the action proceeds under the authority already agreed.

The Architect's Field Notes

Here is the test I keep coming back to: can the person reviewing an AI action see the source, the proposed action, and the reason for it without rebuilding the whole task?

If the answer is no, the workflow needs a better handoff before it needs more autonomy.

Keep a brief record beside the approved draft: what the reviewer checked, what changed, and what action was authorized. “Email wording approved” and “Email may be sent to this recipient” describe different decisions. Make the intended authorization clear enough that the next person does not have to infer it. If the draft changes afterward, identify whether that change needs another review before the action proceeds.

Right Or Wrong?

Before adding an agent to a process, write one sentence that starts: “This agent may...” Then write the sentence that starts: “This agent must stop when...”

If either sentence is hard to finish, keep the agent in draft mode while you work out the boundary.

A Question For You

Where would an AI agent save you the most work, and what action would you still want to approve yourself? Enter your answer in a reply, without confidential business or client details.

Something Funny

The agent finished the task in 12 seconds. The humans spent 20 minutes finding out which task.

May the Profit be with you,

Joy Francis, CFO and Revenue Capture Architect