- A pilot that matches analysts' work shows the software functions. It does not show who answers for the agent's mistakes.
- Michele Grant argued leaders need a baseline, a named owner for each decision and a decision log before they give an agent more autonomy.
An AI agent can match human work in a pilot and still not deserve authority over a core process. Michele Grant made that case on the AI Engineer show. The episode is titled "The Agent Works. Does Anyone Trust It?" and was published on October 13, 2026. Her test is simple. Who owns the agent's failures, and what evidence earned it more freedom?
What was said
Grant used an illustrative scenario, not a real case. A large acquisition has 40 work streams. Each week, analysts check figures against the ledger and build a deck for a steering committee. Engineers build an agent to draft it. In a four-week pilot, its drafts match the analysts' reports line for line.
She said that proves the tech works. It does not settle whether the committee should trust it. Her first demand is a written value case. It needs a baseline for report time, analyst capacity and detection time. Detection time is how soon the committee sees a slip. In her words: "If you can't measure where you started, you can never prove the agent moved the needle."
Her second demand is ownership. The agent prepares the analysis, and people make each decision. As she put it, "The person who makes the decision owns the consequence." She said the one good reason to delay launch is a decision with no owner. Then the agent's flags lead to no action.
Third, she wants evidence built into the agent. That means a decision log of the data used, tool calls, model versions and human sign-off. Core math should run in fixed-rule code that finance has approved. The model's summary counts as explanation, not proof.
Why it matters
Our reading: this moves the buying question from whether the agent works to who answers when it fails. Managers should name a person for each decision the agent feeds, before launch. Engineers should build the log and the baseline as part of the product. A demo cannot show either one.
The other side
The scenario is hypothetical and large. The excerpts give no outcome data showing the approach works in practice. They also do not say how much evidence is enough to grant more autonomy. We cannot say how a small team without finance-approved code would apply it.
Grant also named a risk in her own framework. Once people review only the exceptions, she said, the biggest weakness is whatever the agent failed to flag. Owners and logs help after a miss. They do not guarantee someone spots the miss.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale (2026-10-13)
