Most enterprise AI projects die not because the model was weak, but because nobody could answer what happens when it gets something wrong. This is our answer.
Four constraints we do not negotiate
Not approved, not sent
Checked when the API is called and again in the database as the row is written. No setting turns it off.
A workspace is a real boundary
Enforced in the data layer, not hidden in the interface.
Every change leaves a mark
Adjustments create new records. History is not overwritten.
States are never faked
Live, demo, error, waiting and blocked all show honestly. No pretty screens covering things up.
Where a customer-facing message stops
Four stages, and the third is the one no setting turns off.
- 01
An agent drafts
With a source behind every number
- 02
Policy gates
Off-policy pricing, the spend ceiling, who may read what
- 03
A person approves
Checked at the API, and again in the database as the row is written
Needs approval - 04
Send and record
Who approved, when, and the data as it stood
Who approves what
Real permissions live in workspace grants. Roles only change what the interface puts first.
| Task | Who can | Where it stops |
|---|---|---|
| Draft a quote | An operator | Off-policy pricing gets flagged |
| Approve and send to a customer | Someone with approval rights | The API and the database |
| Change the pricing policy | A manager | Version and author recorded |
| See raw financial data | Account owner and admin | Checked in the data layer |
Questions we get
What if we want the AI to send on its own?+
We do not offer that. It is a runtime constraint rather than a configuration option, and it is the reason operations teams let AI touch a live order at all.
Is our data used for training?+
No. It stays in your workspace and is not used to train models.
What do we show an auditor?+
Every run leaves a trail with timestamps, the accountable person, and the data as it stood. From a line on a report back to the source document.
Show this to your compliance lead
We would rather take the hard questions now than have the problem surface in month three.