Atlas. Together we make an impact
Interactive map. Click any box, or play a scenario.

How a governed AI agent actually works

The agent never touches your systems directly. Everything goes through a gateway you control. Play the three scenarios from the demo and watch where each one is allowed, blocked, or held for you.

Gateway: the agent only sees this Decision ownersets the boundaries, signs the decision AI agentreads, reasons, draftsasks the gateway for everything Identity and permissionswho the agent is, what it may touch Budget and run limits€ per run, max steps, time Tool allow-listwhat it may call; the rest does not exist Approval holdactions with impact wait for a person Audit logevery call, outside the agent's reach Data sourcesproject, history, capacity, PTRS Language modelwrites the narrative, does not classify Rules engineclassifies risk, cites the rule Tools with impactchange gate, move people, send mail Decision recordwhat was allowed, what happened, who signed, what it cost request escalate / approve
Back to the summary
Built for the Atlas roundtable at PPM Summit Frankfurt, 23 September 2026. Companion to the decision-record demo. Illustrative and simplified.
Miloš Dordevič · milos.dordevic@chooseatlas.cz  ·  Jonas Zoulik · jonas.zoulik@chooseatlas.cz