Trust Is the Real Blocker for Agents at Work

Michele Grant told AI Engineer the agent working is not the test. Trust is. The creator of Leading at Scale framed the enterprise question bluntly: the agent works, does anyone trust it?

That trust has to be engineered before go-live, Grant argued, the same way she did for more than a decade leading M&A integrations across media, tech and consumer goods. An agent earns permission to run only when it delivers budgeted value inside the firm's risk appetite, policies and regulatory guardrails.

Grant grounded the framework in a $15 billion acquisition chasing $400 million in net synergies. The integration management office runs 40 workstreams, from vendor contracts to ERP consolidation, with analysts reconciling actuals to the general ledger every week for a steering committee of the CFO, chief risk officer and integration lead. Engineers built an agent to ingest weekly updates, reconcile the ledger, run variance math and flag lagging initiatives. A four-week pilot matched the analyst reports line for line. Tech works. Value is still unproven.

So Grant put three questions to leadership. What value does it create, who is on the hook, and what evidence earns autonomy. For value, she required a written case and a pre-deployment baseline on three metrics. Cut the five-day report build to a target of two days, count the 80 hours a week freed when four analysts spend half their week on the deck only after that time is redeployed to behind-plan workstreams net of compute and review cost, and most critically, shrink detection time from first data slip to steering committee visibility. If you cannot state the baseline, Grant said, do not deploy.

Ownership follows the three lines model, Grant said. The integration lead owns the output in the first line. Risk and compliance challenge it in the second, calibrated to exposure. An internal draft reviewed by a senior analyst gets lighter scrutiny than an agent publishing to the board. Internal audit decides in the third line when the agent enters the audit plan on material risk. Day to day, one employee, the senior analyst, must understand boundaries, interpret output, override or reverse it, and hit a kill switch. Grant codified authority in a one-page human work standard. The agent prepares analysis. Humans make the decisions. The senior analyst calls timing difference versus cash shortfall, the workstream lead owns remediation, the integration lead flags at-risk initiatives. No assigned owner means no deployment.

Evidence for autonomy comes from how the agent is built, not what it says. Grant looked for three artifacts. An auditable decision log that captures raw data ingested, tool calls, prompt and model versions, output and human sign-off. Core reconciliation must run in deterministic, finance-approved code, not conversational prose, Grant said, and the model's summary is explanation, not proof. A model inventory record holds the accountable executive, lineage, dependencies, validation results, escalation path and kill-switch runbook plus regression triggers that force revalidation after base model or API changes. Then an autonomy scorecard gates four stages: shadow mode side by side, supervised draft with line-by-line review, exception plus random sample, and exceptions only with periodic independent sampling. Stage four is brittle. What the agent failed to flag is the risk, so Grant tracks missed material issues, error severity and override rate. A falling override rate does not mean the agent got smarter. It can mean review got weaker.

The missing piece in most teams is that translation from principle to per-agent controls, and vendors are racing to productize it. Microsoft's Agent Governance Toolkit advertises policy enforcement, zero-trust identity and execution sandboxing covering 10/10 OWASP Agentic Top 10, according to its GitHub listing. That maps directly to Grant's four per-agent decisions: what it is permitted to do, who owns it, what evidence it records and what triggers escalation. Principles alone will not move an agent from pilot to the ledger. A named owner, a baseline, and telemetry you can audit will.