What can your AI actually do autonomously — can it act, decide, spend, or trigger anything without a human — and where exactly are the human-in-the-loop points?

The full executive answer

The platform is architecturally advisory: no module executes a business action, moves money, changes a system of record, or communicates externally on its own. Every generation terminates in a stored, human-readable artifact — a verdict, a register, a roadmap — and the highest-stakes output, the Vouli Verdict, is deliberately shaped as a proposed board motion, which by definition requires humans to adopt it. The human is not "in the loop" as a courtesy checkpoint; the human is the loop’s entire endpoint.

Even the machinery around generation runs under bounded autonomy. Generations happen on request or on scheduled jobs, but every one passes the same pre-flight gates before a single paid model call: data-readiness checks, per-organisation rate limits, tier spend allowances, and a monthly budget ceiling that fails closed. Retries are bounded — one corrective attempt per gate, then rejection with nothing saved. There is no unbounded agent loop anywhere in the product; the closest thing to autonomy is a queue worker with capped retries and exponential backoff, which is 1990s-grade discipline applied to 2020s models, deliberately.

Where it gets interesting for a CAO: we also advise customers on agentic autonomy, and we practice the doctrine we sell. Our agentic-architecture module prescribes phased rollout — shadow mode, then assist, then autonomous-low-risk, then autonomous-high-risk — with explicit safety boundaries and a human-in-the-loop role type in every blueprint it produces. If we ever add autonomous actions to our own product, they enter through that same ladder, starting in shadow. Roadmap honesty: today there is nothing to escalate past "assist", and that is a design position, not a limitation we apologise for.