AI Agents Need More Than Capability: They Need Trust

Camunda’s 2026 survey summary reports AI-agent use at 71% of surveyed organizations, while 11% of agentic AI use cases reached production over the year preceding the survey. Those percentages measure different things; 11% is not the share of organizations with a live agent. The findings point to a deployment challenge, but they do not establish a single cause.

Capability is only part of readiness

Demos are impressive. An agent that classifies a request, pulls the right information, and proposes a clear answer in seconds looks ready to ship. But once it moves from a demo to a real decision about a real customer, four questions surface that operations teams often can’t answer:

• How will the agent behave in an edge case?
• What is it doing right now?
• Why did it choose to act that way?
• Is it still performing the way we expect?

An agent’s generated explanation is not, by itself, a reliable record of why an action was taken. For review, the team needs evidence of the inputs, tool calls, policy checks, approvals, and outcome. Model quality matters, but so does the infrastructure that makes those actions traceable.

Why agents get stuck in pilot

When several agent frameworks run side by side without shared identifiers or records, a case can become hard to follow. An agent’s output may sit in one system and a human’s approval in another. The team then has to reconstruct how automated steps, people, and agents contributed to the outcome.

Camunda also reports that 48% of survey respondents use agents in silos. That is consistent with the integration problem described above, but it does not prove that fragmentation caused the production gap. Data quality, model reliability, security, and operational readiness also need to be evaluated.

There is a regulatory dimension too. The European Commission describes logging, documentation, transparency to deployers, and human oversight among the obligations for high-risk systems. As of September 30, 2026, its updated timeline places the Annex III high-risk rules at December 2, 2027. Applicability depends on the system, the organization’s role, and jurisdiction. An Israeli organization should map its own obligations rather than assume that an EU category or deadline applies automatically.

Three foundations of a trustworthy agent

Alongside asking whether the model is good enough, ask whether you can test, observe, and explain its actions. These are design goals to implement and verify:

1. Test it: run repeated tests on representative scenarios and exceptions. Measure correctness, policy violations, and variation, then decide which outcomes are acceptable before exposing customers to the agent.
2. See it: give authorized operators access to the relevant prompts, model version, context, available tools, tool calls, and results. Apply redaction, access controls, and retention limits to sensitive content.
3. Explain it: connect the case record to the policy version, tool actions, escalations, approvals, and outcome. Capture enough evidence for review without treating generated reasoning as a verified account of the model’s internal computation.

What this looks like in an orchestrated process

In a process-first approach, the agent operates within boundaries defined in advance. Camunda’s agent architecture separates the model’s tool selection from BPMN execution. That gives teams a place to model approval paths and handle failures. A complete decision record still requires deliberate instrumentation, correlation with external systems, and retention settings; it is not created merely by putting an agent inside a process.

The idea of bounded autonomy is central: give the agent room to act within a defined part of the process, restrict the tools and data it can access, and route decisions outside its authority to a person. Test that those limits hold when inputs are ambiguous or tools fail.

What this means for your organization

If your agents are still in pilot, ask what operations, compliance, and leadership can verify about their behavior today. Clear approval points, defined authority, and usable evidence improve readiness. They belong alongside tests of model performance and business value, not in place of them.

At NG Workshop, we help Israeli organizations build exactly that layer: AI that operates inside a governed process, not as a black box sitting alongside it.

Sources

• Camunda: AI Agents Don’t Have a Capability Problem, They Have a Trust Problem
• Camunda: Breaking Through the Automation Ceiling - CamundaCon 2026

Summary

• Camunda reports agent use at 71% of surveyed organizations and production deployment for 11% of agentic AI use cases over the previous year.
• The four operational questions help expose gaps in testing, visibility, and accountability.
• Testing, visibility, and decision records support trust, alongside reliable models and measured business value.

Want to check if your processes are ready for a trustworthy AI agent? Let’s talk