Eighty-six percent of enterprises say their AI agents have moved beyond pilot projects. Only 34 percent trust the actions those agents take.
That gap makes sense. We have watched teams connect capable models to messy data, brittle APIs and shared service accounts, then blame the model when the workflow fails. The agent can reason well and still send the wrong invoice, expose a customer record or retry a payment twice.
Forrester Consulting surveyed 409 senior IT decision-makers for Boomi. Among the least operationally ready companies, 77 percent were pushing agents into production anyway. They reported an average US$2.1 million in extra costs from downtime, rework, compliance fines and lost customers.
Some were running up to 200 agents. That is not scale. It is a large experiment register.
Trust lives below the model
When we build an agent that can act, the model is only one component. The harder work sits underneath: clean source data, dependable APIs, explicit permissions, controlled Model Context Protocol connections and an audit trail that shows every tool call.
The survey data backs this up. High-readiness organisations reported 55 percent confidence in agent decisions, compared with 22 percent among low-readiness organisations. They were also almost twice as likely to use integration-platform tooling in agent workflows, 46 percent versus 25 percent.
This does not prove that an integration platform caused the difference. Boomi commissioned the study, so the product-friendly angle deserves scrutiny. But separate research points the same way.
TeamViewer surveyed 4,200 workers and managers across nine markets, including Australia. Sixty-one percent wanted AI to take no independent action. Yet 70 percent were comfortable with autonomous action when they could intervene. Their requested controls were practical: security protections, notice before major changes, access limits, activity logs and rollback.
People are not rejecting autonomy. They are rejecting autonomy without an off switch.
Start with one job
The best production pattern we have found is deliberately narrow. Pick one workflow with a clear start and finish, such as triaging support tickets, reconciling supplier invoices or preparing a sales brief. Give the agent its own identity. Grant only the systems and records needed for that job.
Then split actions into three tiers. Let low-risk actions run automatically. Require approval for anything that changes customer data, commits money or sends an external message. Block actions the agent should never take.
Log inputs, tool calls, outputs, latency and cost. Test rollback before launch, not after the first bad action. Measure completed work, error rates and human review time rather than prompts, demos or agent count.
That last measure matters. TeamViewer found that 56 percent of respondents often or always checked AI output before relying on it, spending an average of two hours each week. An agent that saves 90 minutes and creates two hours of verification is not automation. It is clerical debt.
By early 2027, the strongest mid-market deployments will not have hundreds of agents. They will have a few agents with named owners, constrained authority and enough operating evidence to earn the next permission.
Leave a Reply