Across interviews with 70 business and IT decision-makers, the pattern that stalls AI transformation is not capability. It is trust. Two of the named blockers say it plainly: “compliance and security reviews delay expansion,” and “users and customers don't trust the agent, so usage stalls.” The prescription is equally telling — prove safety with humans in the loop, layer sign-off, test safely, and prove results to end users. Nothing about a bigger model.
For anyone deciding how to adopt agents, that reframes the whole problem. The bottleneck has moved. And where the bottleneck sits determines what you should invest in next.
Trust is now the go-to-market surface
In the first wave of enterprise AI, the differentiator was model quality. That era is closing. Frontier capability is increasingly a commodity — several vendors, similar benchmarks. What separates an agent that scales from one that dies in pilot is no longer how smart it is. It is whether the people who must sign off on it can.
That makes trust a go-to-market concern, not a compliance afterthought. The enterprises pulling ahead are the ones treating “can this pass security, compliance, and audit review” as a design requirement from day one — because a review gate cleared quickly is a deployment shipped quickly, and a review gate that stalls is revenue that never arrives.
The stakes are rising, not falling
It would be comforting to think the trust gap shrinks as agents get more familiar. The opposite is true, because agents are becoming economic actors. Agent-to-agent payment infrastructure is arriving fast — the x402 standard, now backed under the Linux Foundation by Visa, Mastercard, Circle, Google, and Stripe, exists to let agents transact autonomously. When an agent can move money, call another company's agent, and act with delegated authority in real time, the question “can I trust it” stops being about user comfort and becomes about liability.
An agent that spends, signs, and delegates without an audit trail is not a productivity tool. It is an uninsured risk. The trust gap doesn't close as agents get more powerful — it widens, unless the evidence layer scales with the capability.
What the research means you should do
Microsoft's own prescription — humans in the loop, layered sign-off, prove results — is a description of an evidence system. Turned into practice, it is four moves:
- Evaluate before authority. Score every agent against a standards-based rubric (NIST AI RMF, CSA guidance, KYA) before it gets access, not after an incident.
- Make the score actionable. A non-technical reviewer needs a number they can act on; a technical reviewer needs the artifacts behind it. Both, or neither works.
- Monitor in production. An agent's behavior changes with every model update. A point-in-time review certifies a version that no longer exists.
- Keep evidence in the format review gates already speak — decision logs, impact assessments, control mappings.
The gap Microsoft's research names is not a reason to slow down on agents. It is a map of exactly where the friction lives — and friction, once located, is something you can engineer away.
That is the entire premise of NextGenIQ. We don't ask enterprises to trust agents. We make agents provide the evidence that earns it. Close the trust gap, and adoption stops being a fight with your own review board and starts being a competitive advantage.