Data, Access and Human Review
AI tools touch real business data. This page explains how we limit risk. Specifics are agreed in writing for each engagement.
Access model
- We work in your own accounts and with your own API keys where possible.
- We ask for least-privilege access. If a tool only needs to read a folder, it gets read access to that folder, not the whole drive.
- Keys belong to you. You can revoke them at any time.
- We do not use client data to train anything.
Human review
The more costly a mistake, the more review it gets.
- Low cost of error: the tool may act on its own, with logging.
- Medium cost: the tool prepares, a person approves.
- High cost: the tool only suggests, and a person does the action.
Money, legal commitments, customer-facing messages and deletions normally sit in the second or third group.
Review queues
Uncertain or sensitive items go to a queue. A reviewer sees the input, the AI output and the source. They approve, edit or reject. Rejections are recorded so we can see where the tool is weak.
Failure handling
AI output can be wrong, empty or badly formed. We plan for this:
- Outputs are checked for the expected format before they are used.
- If a step fails, the item goes to a person. It is never silently dropped.
- Retries are limited, so a fault cannot loop and run up cost.
- The tool is told to say when it is unsure instead of guessing.
Kill switch
Every tool has a clear way to stop it. This might be a single setting, a revoked key or a paused automation. The named owner knows where it is and how to use it.
Logging
We log what came in, what the tool produced, who approved it, and what it cost. Logs let you answer "what happened and why". We agree with you what is stored and for how long, and we avoid storing sensitive content we do not need.
Ownership
A named person on your side owns each tool after handover. They decide changes, review logs and use the kill switch if needed.
What we cannot promise
AI will sometimes be wrong. Our aim is to catch errors before they matter, not to claim there are none.