Every organization deploying artificial intelligence has an implicit theory of how the investment pays back. Usually it goes like this: build the capability, put it in front of customers or staff, and the productivity follows. The theory has a missing variable. Whether people use the thing, and use it honestly, decides the return more than the model does, and that willingness is a function of trust.
Trust is treated as a communications problem in most programs, something to be handled after launch with a policy page and a reassuring FAQ. John Margerison, CEO of XFactorAi, has written on the role of trust in enterprise AI, and the subject deserves harder treatment than the comms-plan version it usually receives. Because trust is not a soft variable here. It is the mechanism that converts a working system into a used system, and it can be engineered, measured, and lost.
Adoption, not capability, is where AI returns actually come from
A model that answers well and is consulted by a fifth of the people it was built for is a poor investment. A slower model that everyone routes work through is a good one. That asymmetry is unusual, and it is why capability benchmarks are such a bad proxy for value.
The failure mode is quiet. Nobody refuses to use the system; they simply check its output against the old process, keep the spreadsheet running in parallel, or ask a colleague instead. Every one of those behaviors is a trust deficit expressing itself as duplicated cost, and none of them appear in a usage dashboard as a problem, because the license is still being consumed.
Customers have already begun discounting AI-mediated answers
This is no longer a hypothesis about future sentiment. Gartner’s September 2025 survey finding that 53 per cent of consumers distrust AI-powered search results describes a majority of the market applying a discount to machine-generated information before reading it.
That discount has commercial consequences that are easy to miss. An AI channel that customers half-believe generates a second contact – the call, the chat, the complaint – that the channel was meant to remove. The saving is booked when the system goes live; the cost of verification appears later, spread across the contact center, and is rarely attributed to the deployment that caused it.
Organizations that publish what the system can and cannot do, mark AI-generated output plainly, and give people a fast route to a human being are buying down that verification cost. It looks like caution. It behaves like margin.
Inside the workforce, distrust shows up as workaround, not refusal
The internal picture is more interesting than the customer one, because employees have somewhere else to go. Given a corporate tool they doubt, they will use a consumer one they like, and the sensitive material goes with them.
KPMG’s global study with the University of Melbourne, reported in its May 2025 announcement that trust in AI remains a critical challenge amid the tension between benefits and risks, points at a population that is using these systems while remaining unconvinced by them. Ambivalence of that kind is not stable. It resolves either into confident use, when the organization earns it, or into private use outside the sanctioned tools.
Which means the internal trust question and the security question are the same question, arriving at the same desk with two different names.
Retrofitting trust costs several times what designing it in does
The engineering that produces trust is unglamorous: provenance for the data, logs that record what the system was asked and what it said, defined confidence thresholds, an escalation path when the model is out of its depth. All of it is cheaper to specify before the build than to reconstruct afterwards.
Speed of rollout is exactly the pressure working against that. GeekWire’s 2024 interview with Richard Edelman on the trust risk created by the tech industry’s rapid AI rollouts framed a problem the industry has largely chosen to live with: releasing faster than public confidence can accommodate, then spending on recovery.
Recovery is the expensive path. Once a system has produced a wrong answer that reached a customer, the organization is no longer designing a control. It is proving a negative to a regulator, a journalist or a jury, and doing so with logs that were never built to support the argument.
Trust is becoming a procurement requirement, which is where it risks going wrong
Large buyers now ask suppliers what their models were trained on, who reviews the output, and what happens when the system errs. That is progress, and it also carries an obvious hazard: trust asked for in a questionnaire tends to be answered in a questionnaire.
A completed AI assurance annex proves that someone wrote a document. It says nothing about whether the escalation path is staffed on a Friday afternoon, or whether anybody reads the logs. The organizations that will win here are the ones whose answers can be tested by observation – overrides recorded, error rates published internally, humans genuinely reachable – because that is the standard buyers will move to once the paperwork stops differentiating anyone.
The competitive shape of this is already visible. Capability is converging fast, and it is available on subscription to any competitor with a credit card. Warranted confidence is not purchasable, takes years to build, and can be destroyed in an afternoon, which makes it the scarcer asset by some distance. The firms treating trust as the first design constraint are compounding something their rivals cannot buy in.
