AI intelligence
AI that shows its working.
An analyst over the record, not a chatbot over the internet. Ask why integration health fell this week, or what would be affected if a system failed, and get an answer built the way a good analyst builds one.
How an investigation runs
- Find what the question is aboutIdentify the objects the question is about.
- Gather the evidenceRead their health and its evidence, walk their dependencies, list what is open on them, note what changed.
- Hand over the evidence, and only the evidenceHand the evidence, and only the evidence, to the model, fenced as data, with every person's name replaced by a token.
- Check the answer against what was readTake the answer back through the grounding check: a fact that cites nothing read is dropped; an answer with none is refused.
- Show the workingShow facts, inferences (marked, with confidence), uncertainty, next steps, and the full list of what was read.
What an answer looks like
Every answer separates what the record shows from what the AI infers from it, and says what it does not know.
QuestionWhy did the customer feed from Harbour's CRM fail this week?
The customer feed from Harbour CRM to the group platform is failing. Feed health, read on Tuesday
The field customer_ref changed from integer to text on Tuesday. Schema change on Harbour CRM, Tuesday
The change of type is the likely cause of the failure. An inference, medium confidence.
Whether the group platform rejects the new type or reads it wrongly. No reading covers that yet.
Ask Harbour's Head of Technology to confirm the change, then review the field mapping.
The feed's health and its evidence, the schema events on Harbour CRM, and the incidents open on this feed.
An illustration on the fictional Northstar Group.
What the model never gets
It reads assembled evidence, nothing else.
Only what the asker's own permissions allow.
Names, emails and phone numbers are tokens before anything leaves.
No fact without a citation. The page says it is AI, every time.
Off until you turn it on
The provider sits behind one interface and can be changed without touching the application. Customer data reaches a model only under zero-retention terms, and only after your administrator has turned the layer on.
Ask it a question yourself
The demonstration ends with an investigation on the Northstar scenario. Bring a hard question.