HONG KONG (Realist English). A report by the KYA (Know Your Agent) lab at the University of Hong Kong, published on 3 November at Hong Kong Fintech Week, records that 76% of surveyed companies over the past year have encountered concerns about trust, security, accountability or regulatory compliance when working with AI agents. Of these, 23% encountered this often, 53% — periodically, and only 2% said they had never experienced such problems.
Scale of the gap: adopting, but not trusting
The data paint a paradoxical picture. Forrester Consulting, commissioned by Boomi, surveyed 409 IT executives at director level and above in North America, Europe and APAC: 86% of organisations have already moved beyond the pilot stage of AI agent adoption, yet only 34% said they trust the actions their agents take.
In another survey by VentureBeat Pulse (June 2026, 157 respondents), only 5% of respondents said they fully trust the automated evaluations that are supposed to ensure the safe release of agents into production.
An Okta report points to the commercial consequences: nearly 60% of organisations are rushing to deploy agents that perform actions, but 80% of such initiatives are behind schedule. According to Techstrong TV, 70 to 90% of agents being built today never reach production, and about 20% of this lag is directly linked to trust problems.
What exactly undermines trust
The KYA report identifies two main problems that respondents named as the key barriers to commercial adoption: “low trust from clients or partners” (53%) and “uncertainty over who is responsible for errors” (42%). These two factors outrank difficulties integrating with existing systems (35%) and challenges calculating return on investment (34%).
The consequences are measurable: 41% of respondents reported an increased burden related to compliance and auditing, 40% — delays in launching or expanding projects, 38% — the need for additional manual confirmations, which reduces efficiency, and 23% encountered refusals by clients or partners to use agent-related services.
The “evaluation gap”: autonomy is growing faster than verifiability
VentureBeat describes the situation as an evaluation gap: agents are gaining more freedom, but mechanisms for verifying their reliability are not keeping pace. Half of the enterprises have already encountered a situation where an AI agent or an LLM-based feature that had passed internal evaluations still led to a failure on the client side; every fourth — more than once.
At the same time, 66% of respondents already permit some degree of autonomous deployment without human involvement or are building systems that will do so within the next 12 months.
Among the reasons for distrust in automated evaluations, respondents cited poor alignment with real outcomes (29%), bias or inconsistency (21%), insufficient explainability (18%) and risks of data leakage or privacy (17%).
Why this is a systemic problem, not a “model flaw”
Experts agree that improving model quality alone does not solve the problem. The KYA report puts it this way: “An increase in model capability does not automatically convert into commercial delegability. Only when identity can be verified, execution — checked, and authority — limited, does the agent receive economic authority commensurate with its capabilities.” In other words, trust is not a characteristic of the model but a property of the system: identity, permissions, audit and accountability.
Google Cloud, in practical recommendations for enterprises, directly names trust as the main barrier: employees fear for their value if an agent performs most of their work and wonder who will bear responsibility for an error. The recommendation is to treat agents as junior team members: delegate, define boundaries, verify and start with low-risk scenarios.
What is proposed to bridge the gap
The KYA report proposes a KYA framework — an analogue of KYC for agents — which should provide three things: verifiable identity, auditable execution and limitable authority. 88% of respondents believe the industry needs a unified standard for evaluating or certifying trust in agents, and 60% are ready to adopt such services — either immediately or first in high-risk, critical scenarios.
Forrester separately points to integration as a key differentiator: organisations with a high level of “agent control” are three times more likely to say that reliable, well-managed APIs determine whether they will launch a scenario at all; 46% of them use iPaaS to support agent workflows versus 25% in the “agent chaos” group.
The problem of trust in AI agents is not a temporary technical difficulty but a structural barrier that will determine adoption rates in the coming years. Companies have already deployed agents on a mass scale but are not ready to delegate significant decisions to them, since mechanisms for verifying the agent’s identity, checking its actions and clearly allocating responsibility are absent.
Until these institutions are formed, the gap between declared autonomy and real trust will only grow. The key question is whether the industry can develop unified standards of “agent trust” before accumulated failures undermine trust in the technology itself.
