Autonomy is a dial your team turns.
Agents read live state, plan, act through your APIs, then check what changed — one notch at a time.
Sense, plan, act, verify.
The gate sits on act — never on sense or plan. The agent may think as hard as it likes.
Sense
Primary records on your side. No scraped dashboards.
Plan
Steps committed up front. Below your bar, it halts and says why.
Act — gated
A reversal path exists before it can fire.
Verify
Checks what changed. A copilot stops; this closes the loop.
Nothing arrives at full autonomy.
Each class of action starts at the bottom and climbs only on your own numbers.
Recommend only
The agent proposes; a person acts. Every disagreement logged.
Execute on approval
Staged and quantified. A named approver signs once.
Run within policy
Unattended inside your thresholds. Cross one and it parks.
Moving up a rung is your team's call. gothink.ai never widens an agent's authority on its own, in a release note, or by default.
Reading is free. Changing state is not.
Anything that only observes
Detection, attribution, forecasting, drafting — continuous and unattended.
Anything that changes state
A purchase, a posted entry, a lifted payment block — held for a named human.
Budgets and blast radius
Caps per run, allow-listed operations, and a kill switch you own.
Replayable, step by step
Inputs, tool calls, cost and the approver's name — in your systems.
Scored before it ships
Your ground truth reruns on every model or tool change.
A confidence gate on every field.
You set the bar per field and document type. Doubtful values go to a steward, not the database.
- Raise the bar: more goes to review
- Lower it: more goes straight through
- Either way, the trace records which happened
Better on its own — inside your guardrails.
Gates decide what an agent may do. The Loop decides what it may improve.
Maker
Proposes one small, reviewable change
Checker
Reviews it independently
Scored test
A test that can fail the work decides
Memory
Lessons carry into the next run
Hard stops
Every unattended run is bounded. Cross a limit and it halts.
Provably right, not confidently wrong
Accepted because it passed the test — not because the model said so.
Three gates. Ninety days.
Then priced per unit (document, decision, receipt), never per seat, plus an annual platform fee; quoted after scoping.
Prove
Your documents, labelled by your experts. Your straight-through rate is measured here; we publish no figure until it is measured.
Expand
More document types; gate thresholds set by your risk function.
Industrialize
Into your own estate. We stop at day 90 if targets are missed.
Turn hand-found work into a governed capability.
Walk through the loop, the rungs and the gates with our engineers.