Skip to content
Ascenda

How much of our team's AI agent work goes through the approved model or vendor?

Short answer

Declare which vendors and routes are approved, such as a specific cloud provider or an internal gateway, then measure the share of agent time that went through them across the team. Ascenda reports this as a cohort figure, so you see policy coverage without anyone's individual usage.

Why it matters

Many organisations approve AI agents on condition that traffic goes through a particular route: a cloud provider in a chosen region, or an internal gateway. Whether that holds in practice is hard to see from the platform side, because work that goes elsewhere doesn't show up there.

How Ascenda helps

Flow reads, on each engineer's Mac, which model vendor each session used and how it was served: direct, through a cloud provider (Bedrock, Vertex, Azure), a gateway, a custom endpoint, or on device. The organisation declares its approved vendors or routes, and Ascenda reports the share of the cohort's agent time that went through them.

  • Route is a category, never a hostname or a country.
  • The figure is an aggregate of ten or more, with no individual view.
  • Route is read from Claude Code and Codex; vendor also from Cursor and Windsurf.

What else you'd need

Your gateway or cloud provider logs show the traffic that reached the approved route. Ascenda adds the engineer's side: the share of all agent work that went through it, including work that went elsewhere.

Next step

Start with a 30-day Report

The Report works from your engineers' imported history for a cohort of ten or more, with no integration, and ends in a written deliverable. The 90-day Study is the next stage and runs live.

Book a StudyGet Flow for Mac

Related questions

How can I see my AI coding work split by model vendor and route?

Flow shows your own AI coding work split by model vendor and by how it was served: direct, through a cloud provider, a gateway, a custom endpoint or on device.

How do we test whether switching AI model or vendor helped our engineers?

Roll the switch out to groups in turn. Ascenda compares failed tool calls per active hour with groups that haven't switched yet, and says when it's too small to tell.

How can we measure AI engineering without monitoring individual employees?

Ascenda reports only aggregates of ten or more consenting engineers, with no individual view for managers. Prompts, code and model outputs are never sent.