The inbound → approve → execute loop
Business-Gym runs on one operating loop: inbound → approve → execute. Every item that arrives is triaged and routed, turned into a proposed action, checked against confidence and risk, and executed only after a person approves anything consequential. It is the operational heart of the platform — the same loop every head uses.
Parts of that loop are live today and parts are still being built. This page marks which is which, so you know what you are relying on.
What happens when something arrives?
Section titled “What happens when something arrives?”It is ingested and classified — what kind of item it is, which area owns it, how urgent it looks. The system also does identity resolution: working out who the sender is and which customer, supplier or colleague they map to, so the item lands in the right place with the right context attached.
That first layer is live. Everything after it is where the grades start to differ.
What does the whole loop look like?
Section titled “What does the whole loop look like?”The five layers of the inbound → approve → execute loop, with the human approval gate drawn explicitly: every external send passes through it, and the learning loop back to layer 3 is planned rather than live.
What decides if it needs me?
Section titled “What decides if it needs me?”A confidence × risk gate. It weighs how sure the system is about the proposed step against how much damage that step could do. Reversible, low-risk, high-confidence prep can proceed if you allow it.
The gate is partly built — treat it as a filter that catches the obvious cases, not as a finished judgment layer.
One rule sits above it and does not move: anything that sends externally is approved by a human, every time. That holds regardless of how confident the system becomes, and it will still hold once the learning loop exists.
How much of a job can it do in one go?
Section titled “How much of a job can it do in one go?”One step. Today an action is a single verb — draft this reply, create this task, update this record — not a multi-step plan it works through on your behalf. A job that needs three things done in sequence comes back to you between them.
Matching an item to a playbook — a reusable “when this arrives, do these things” recipe — is the layer that would turn single steps into sequences. It is largely not built yet. Do not plan work on the assumption that a playbook will carry a case end to end.
What happens if a connector is not set up?
Section titled “What happens if a connector is not set up?”The step tells you the truth: it renders a plain “not connected” state and the item queues until you configure the connector. It does not skip the step quietly, and it never reports a send, a filing or an update that did not happen.
This matters more than it sounds. A fabricated success is worse than no automation at all, because you find out weeks later that a customer never got the reply. Business-Gym would rather stop and show you a gap.
Does it learn from my decisions?
Section titled “Does it learn from my decisions?”Not yet. The intended design is a closed loop — execute the action, capture the outcome, feed that back so the next proposal is closer to what you would have done. That is the differentiator the platform is being built toward, and it is not live today.
What happens now is simpler: actions are executed and their outcomes are recorded. Your approvals and edits do not yet change future proposals.
When the learning loop does arrive, it will tune the drafts. It will not lower the approval floor — external sends stay human-approved.
Which parts are live today?
Section titled “Which parts are live today?”- Ingest, classify, route, identity resolution — live.
- Playbook match — mostly not built.
- Multi-step action plan — partial: a single action today, not a plan.
- Confidence × risk gate — partial, with the external-send rule always enforced.
- Execute → outcome → learn — planned, not live.
- Honest “not connected” handling — live.
- Human approval for external sends — live, and permanent.
Video script
Section titled “Video script”Every piece of inbound work runs through one loop: inbound, approve, execute.
It is classified, matched to the right area, and tied to the person it came from.
A confidence-and-risk gate decides what can proceed — and anything that sends externally waits for a person, every time.
If a connector is not set up, the step says so and the item queues. It never claims a send that did not happen.
Next: meet the department heads that run this loop, or connect your integrations.