AI Agents Are Deciding—Do You Trust The Output?

This voice experience is generated by AI. Learn more.This voice experience is generated by AI. Learn more.Nalini Garg, Associate Vice President, Deloitte.

getty​I’ve sat in enough enterprise AI reviews at this point to notice a pattern: Leadership teams describe themselves as confident in their AI output, right up until someone asks who actually checked it. That gap between stated confidence and real verification is the problem I keep running into, and it’s rarely about the model. It’s about what happens, or doesn’t happen, around it.

The data backs up what I’m seeing in the field. A 2026 Workiva survey found that 84% of executives are confident in AI-generated content reaching critical reports without human review, while only 11% believe their underlying data quality can actually support that confidence. Grant Thornton found something similar: Seventy-eight percent of leaders lack confidence they could pass an independent AI governance audit within 90 days. Confidence, in other words, is outrunning verification almost everywhere I look.

In my own work advising enterprise leaders on AI adoption, I’ve found that when someone tells me they “don’t trust” an AI output, they’re almost never describing one problem. They’re describing one of three, and conflating them is exactly why so many governance efforts stall out.

The first is a data problem: The model was fed something wrong, incomplete or stale, and no amount of polish in the output changes that. It’s the easiest to spot but rarely the whole story.

The second is what I’d call an articulation problem, and it’s the one I see leaders underestimate most. A system can have good data and a sound answer and still fail to show its work: no source cited, no confidence level attached, no way to tell whether a number was retrieved or guessed. I’ve watched finance teams accept a figure at face value because it was presented with total confidence, then scramble when a board member asked where it came from. As companies move from simple chatbot tools toward agentic systems that plan and execute several steps on their own, this gap gets more dangerous, not less, because a small unverified error early in a chain can compound silently across five or six actions before a human ever looks at it.

What actually closes this gap, in my experience, isn’t a policy memo. It’s what I think of as an agent harness: the engineering layer that decides what tools an AI system is allowed to call, what data it can touch, when it has to stop and ask a person before proceeding and what trail it leaves behind. I tell clients to think of it less as a guardrail and more as a receipt requirement. A model gives you an answer; a harness makes it prove how it got there before anyone downstream is asked to act on it.

The third problem is the one nobody wants to name because it isn’t rational: Even a correct, well-documented answer can still feel hard to trust simply because a machine produced it. This is a well-documented behavioral pattern, sometimes called algorithm aversion: People who see an algorithm make one visible mistake tend to walk away from it for good, even when it keeps outperforming the human alternative. I see the mirror image of this inside teams, too. Atlassian researchers recently found that when colleagues learned identical work had AI assistance, they rated its author as 10 times lazier and were far less likely to recommend that person for a high-visibility project. Clean data and a transparent harness fix the first two problems. Only a consistent track record fixes the third, and that takes time, not tooling.

This is why I push leadership teams to stop treating “trust in AI” as one initiative and start building it as six specific habits, which I usually shorthand as VERIFY:

Visibility: Every output carries traceable data lineage. Executives cannot calibrate trust in a black box.

Evaluation: Continuous benchmarking against ground truth on live data, not one-time validation.

Risk-tiered review, enforced by the harness: The harness mandates a checkpoint before any high-risk action, so oversight happens by design.

Incentives: Reward calibrated skepticism, not adoption volume, so metrics don’t suppress verification.

Feedback loops: Every override and near-miss is logged and routed back to the model owner.

Yardsticks and accountability: A single named owner per AI system, with clear decision rights and audit-readiness.​​

None of that requires new technology; just deciding in advance how much scrutiny an output deserves before a deadline decides it instead. In every organization I’ve worked with, the ones that get this right aren’t the ones with the most confident executives. They’re the ones where someone is quietly expected to check the number before it reaches the board deck, every time, whether or not anyone’s watching.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?