AI safety
HighSign is built around one rule that doesn't bend: no AI model sits in the answer path. This page explains what that means and how the optional AI features are scoped around it.
The core invariant
Every green, amber or red verdict is computed deterministically from the policy rules your governance team writes. Given the same rules and the same facts, HighSign always produces the same answer. There is no AI model between your policy and the verdict a staff member sees, so an answer can always be traced back to a specific rule, not a model's guess.
Optional AI features are drafting aids only
HighSign has optional AI features, such as researching a tool's data-governance profile or mapping a workflow's steps. They are off by default, run only once your company has switched them on, and each one runs only when someone on your team starts it, for example by clicking "Research with AI" or "Generate map". They work the same way every time:
- They produce a draft, never a final answer.
- A human confirms the draft before it becomes part of your policy or register.
- They can never set a verdict. Verdicts only ever come from your confirmed policy rules.
Off by default, switched on per company
AI features are off for every new workspace. An owner or admin can turn them on, or off again, under AI features in Settings. The switch covers your whole company, everyone on your team can see whether it is on, and each change is recorded in your activity log with who made it and when.
While AI features are off, nothing is sent to an AI provider: the AI controls are replaced by a short note pointing to Settings, and HighSign refuses an AI request on the server even if one is attempted. Research and workflow maps that were drafted earlier stay in your workspace, and you can still review or discard them.
What gets sent to an AI provider
The AI provider is Anthropic, which processes what each feature sends as our sub-processor. What each feature sends to it differs:
| Feature | What is sent to Anthropic |
|---|---|
| AI tool research and profile drafting (Tool Register) | The tool's name, vendor and public website, so the model can read the vendor's own terms and policies. Nothing from your workspace. |
| Map from a description (Workflows) | The text you type, paste or dictate into the box, up to the first 8,000 characters. This is your own content, so the panel reminds you to leave out client names and other personal or confidential details. |
Apart from the name of a tool you choose to research, no feature sends your policy text, your register, your staff list or anything staff submit. Nothing is sent while AI features are switched off, nothing is sent until you start a feature, and the workflow panel says what will be sent before you click.
No training on your data
HighSign does not use your workspace content to train third-party models.
Why it's built this way
A governance product that let an AI model quietly decide verdicts would be governing AI risk with more AI risk. Keeping the verdict engine deterministic means your risk team can stand behind every answer, and an auditor can check the reasoning without needing to trust a model's judgement.
Related
- Security and hosting for hosting, isolation and the activity log
- Privacy and data use for what's collected and how aggregated benchmarks work
- /security for the full statement this page summarises
This page is a summary for convenience, not legal advice.