Set aside whether AI vendors deliver what they promise; CIOs have a more revealing complaint. Asked to name their biggest frustrations with the AI vendor community, they most often chose lack of transparency around model performance, data use, or risk (34.3%), in a tight cluster with overpromising (31.2%) and limited integration (30.9%) (Source: IDC CIO Sentiment Survey, May 2026, n = 459).
The ranking is narrow. The composition is not. Opacity and overpromising are one failure seen from two sides: a claim the buyer cannot check.
Transparent comes from the medieval Latin transparere, to show through. The word carries a question that marketing tends to drop: visible through to what, and for whom? The CIOs named their answer: performance, data use, risk.
The industry usually answers with a model card describing the model as trained, a benchmark describing someone else’s test, or a policy describing intentions. Each is accurate within its domain. None shows how this system behaves, with this data, in this environment.
These are not the same thing. Disclosure is what a vendor chooses to show. Transparency is what a buyer can see through to without the vendor’s narration. They coincide only when the evidence is generated by the running system and checkable by someone other than its author.
The sharpest data comes from where enterprises hand AI the most authority. Asked what most increases their trust in autonomous security operations decisions, security leaders chose mechanisms: the ability to roll back actions (54.7%), analyst feedback shaping decisions (50.5%), confidence scoring (46.6%), and transparent logic (43.4%). The one performance claim, proven reduction in false positives, drew 18.5% (Source: IDC Global AI-Driven SOC Survey, June 2026, n = 871).
Trust, in its classic definition, is willingness to be vulnerable “irrespective of the ability to monitor or control” the other party (Mayer, Davis, and Schoorman, 1995). What these leaders describe is the reverse: confidence that arrives after checking, and only because checking and reversing are possible.
Agents make this urgent, because they do not hold still. Models are swapped, tools are added, uses drift. A point-in-time assessment certifies a moment, not a process. Enterprises should expect four kinds of continuous evidence from any AI that acts for them:
- Disclosure at the point of use: which system, model, and controls produced this output, shown where the output appears
- A tamper-evident record of every action: tied to an identity
- Reversibility: designed in before the action is allowed
- Comparative evidence before change: so a new model’s behavior is seen before production, not after
One caution. Evidence can be manufactured badly: checkbox assessments, benchmark-tuned evaluations, disclosures describing the system as designed rather than as running. Transparency that cannot itself be checked is opacity relocated one layer up.
So the useful questions are narrower than “are you transparent?” Transparent about what? To whom? At what moment? Checkable by whom other than the author?
Buyers did not ask to see more of what vendors want to show. They asked to see what the system does.
Message From The Sponsor
Airia is the enterprise AI platform that discovers, secures, governs, and optimizes AI at scale. As agentic AI expands across the enterprise, Airia gives security, compliance, and IT teams complete visibility and control over every agent, model, and workflow in their environment. From surfacing shadow AI and enforcing agent controls at runtime to generating compliance evidence that satisfies regulators and managing model performance and cost, Airia enables organizations to move fast with AI without sacrificing oversight. Trusted by over 500 enterprises worldwide. Learn more at airia.com.
Put these ideas to work.
Schedule a 30-minute walkthrough with our team.