Is a bot ever fully unsupervised?

No. FULL_AUTO means the bot runs and you audit anytime; every action is still logged, and you can demote it at any time.

What drives the confidence score?

Your approvals and edits across runs. The bot earns latitude by what you actually approve: not by asking for it.

Can I skip straight to FULL_AUTO?

You set the trust level, so technically yes, but the design intent is that autonomy is earned, because reviewed bots are bots you can trust.

What happens when a bot is trained?

Feedback patches its kernel set; every bot sharing a kernel improves with it, and the change is a dated, reviewable event.

AI Agent Lifecycle: Hired, Promoted, Fired

Most AI agents are fire-and-forget: you prompt them, they answer, and you never see them again. Raisy treats bots differently: as employees with a lifecycle. Hired into probation, reviewed on every output, promoted as confidence grows, and trained from your feedback. The difference is not a feature list. It is trust you can scale.

What we set out to change

Two things make people reluctant to hand work to AI: unreliability and opacity. An agent that acts without oversight is a liability, and an agent with no audit trail is a black box. The standard answer in the market is to keep everything in manual mode forever, which defeats the point of automation. (This lifecycle is one piece of the Raisy company operating system; the cost model explains what a promoted workforce is worth. Browse the blog.)

The idea / how Raisy fixes it

The fix is a lifecycle with a review gate where it matters. A bot starts in MANUAL mode: it prepares, you review every run, and nothing reaches the outside world without you. As its confidence score climbs, you promote it through SEMI_AUTO: AI prepares, you approve before it acts externally , and then to FULL_AUTO, where it runs and you audit anytime. Autonomy is earned, not assumed.

How the lifecycle works

Hired: a set of kernels is created for the role.
Probation: MANUAL approval mode, every output is reviewable before it acts.
Reviewed: a confidence score is tracked per bot, per run.
Promoted: you move it MANUAL → SEMI_AUTO → FULL_AUTO. Your call, not the machine's.
Trained: feedback patches the bot's knowledge: teach one bot, and every bot sharing that kernel improves.
Fired: archived with its full audit trail.

Every step of that lifecycle writes a log row. Promotion, training, and firing are events you can point to.

What the user actually sees

Your roster shows each bot with its role, its current approval mode, and its confidence score. You approve a run, the score goes up, you promote the bot, and it starts handling more on its own, while the audit log keeps growing underneath. You never have to choose between "unreviewed AI" and "AI that never does anything."

Figure 1. Autonomy is earned: bot autonomy rises from 0% at hire to 90% at full-auto while human review falls from 100% to 10% — review scales with trust but never disappears. Hover for values.

Enlarge

Autonomy is earned: autonomy rises and human review falls as a bot progresses through its lifecycle — Figure 2. Autonomy is earned, not assumed: autonomy rises and review need falls as confidence grows.

Why this matters beyond convenience

Trust is not binary. A new hire does not get the keys on day one; a reliable one earns more latitude over time. If AI is going to run real parts of a company, it should earn trust the same way: measurably, reviewably, and reversibly. That is the difference between an agent you experiment with and a workforce you rely on.

Deploy your first bot and watch it earn its promotion at raisy.team