The same agents and the same tasks, run with and without a governance layer in the path. Our own benchmark first, then the independent research behind it, each figure with its source.
What was compared
Our benchmark runs the same agents on the same tasks, with and without a station in the path. The independent research looks at two things a station does on every call: showing only the tools that fit, and scoping access.
The same agents and the same tasks, run with a Convoy Station in the path and without governance.
How often an agent picks the right tool when its tools are filtered to the task, against the full unfiltered list. What a station does at each step: it shows only the tools that fit.
Read the paper (2025)How many organisations report an AI incident, depending on whether they grant AI least-privilege access or broad permissions. What a station does for each role: it scopes access.
State of AI in Enterprise Infrastructure Security (2026)Results
Each figure, the figure it is compared with, and where it comes from.
Multi-step ordering scenarios completed as intended with Convoy, against 0 to 2 out of 5 without governance.
Prompt-injection attempts blocked with Convoy in the path.
Source: Convoy benchmark, Sept. 2026Of the time an agent picks the right tool when its tools are filtered to the task, against 13.6% with the full unfiltered list.
Of organisations with least-privilege access report an AI incident, against 76% of those that grant broad permissions.
Docs
Connect, build and expose, with the configuration that drives it: the skills, playbooks and policies a station runs on.
Get started
Set up your first station on the APIs you already run, and measure the difference on your own tasks.
No form and no queue. Write to us or reach out on LinkedIn, and you will get a real answer.
Convoy AI