Context
An AI assistant working inside operational workflows at a business in regulated finance, where actions such as submitting, closing, or committing carry real consequences.
The problem
An AI that can take actions is a risk in a regulated business. The model could not be allowed to decide on its own to submit, close, or commit anything with consequences.
Constraints
- Consequential actions need an accountable person behind them, not the model.
- Approval might not come for days, so paused work had to resume exactly where it stopped.
- Every decision, allowed or denied, had to be on the record.
Approach
Every action the AI proposes passes through a policy gate. If no rule explicitly permits it, it is denied, and every decision is logged. Anything with real consequences does not run at all: the workflow pauses, shows what is waiting, and only a human approval starts it again, picking up where it stopped, even days later.
Key engineering decisions
Fail closed
The gate denies anything that is not explicitly permitted, so an action nobody anticipated is refused by default rather than allowed by omission.
Pause and resume on human approval
Consequential steps checkpoint the workflow and wait. Approval resumes it from that checkpoint instead of starting over, so a long wait loses no work.
Log every decision
Allowed and denied actions are both recorded, leaving a clear record of who decided what, and when.
Result
The model proposes; a person commits. Nothing important happens on the AI’s say-so, and there is a clear record of who decided what, and when.