Agent trust is now an operating problem, not a model problem

The argument. Michele Grant, an enterprise transformation and M&A integration veteran, says the agent that works is the easy part. The hard part is earning the business trust to run it inside a core workflow. She frames trust as three questions: what value does it create, who is on the hook for it, and what evidence allows more autonomy. It is a governance model, not a technology pitch.

The scenario. She grounds it in a $15 billion acquisition targeting $400 million in synergies across 40 workstreams. An integration management office analyst team spends five business days compiling a weekly steering committee report. An agent drafts the report, reconciles the general ledger, and flags slips. A four-week pilot matches the manual reports line for line. That proves the tech works, but not that the committee should trust it.

The value case. Before production, leadership must write a value case with baselines. Three metrics matter: cycle time from five business days to two, analyst capacity, and detection time from first slip to committee visibility. Four analysts spend half their week on the deck, so the agent can return 80 hours, but that counts only when reassigned to lagging workstreams and netted against compute, licensing, and review costs. No baseline means no deploy.

The accountability. She applies the three lines model: management owns controls, risk and compliance challenges, internal audit assures. Each agent gets one senior analyst with operational boundaries, override authority, and a kill switch. A one-page human work standard maps every decision fork to a human role, required evidence, and escalation path. The agent prepares analysis; people own consequences.

The evidence. Autonomy comes from three artifacts: an auditable decision log, a model inventory record, and an autonomy scorecard. The important engineering claim: core math and reconciliation must run in deterministic, finance-approved code, with the LLM summary treated as explanation, not proof. Regression triggers force full revalidation after model or API changes.

The trade-off. The four autonomy stages track value and control side by side, ending in exceptions-only review. The vulnerability: a falling override rate can mean the agent improved or the human stopped looking. Independent sampling is the check. The evidence is expert opinion with a realistic example, not production data. Teams should borrow the questions and scorecard, then test them on one controlled workflow before scaling.