AI
What the assistant is allowed to do, and who decides
“It has AI” is no longer a differentiator. The question a buyer should ask is: what can this assistant change without my approval?
5 min read
The right question is not about accuracy
Discussion of AI in business software usually revolves around accuracy: how often is it right? That matters, but it is not the first question.
The first question is: when it is wrong, how far does the mistake travel? A system that only proposes has its errors seen and rejected by a person. A system that executes has its errors written into the record.
Four boundaries
Four things to ask any vendor
Whose eyes does the assistant see with?
This is the sharpest question and the least often asked. If the assistant reads with the whole firm's access, the first time someone in sales asks it about salaries, it will answer. Ask whether the access filter lives inside the query or is applied after it — because "after" means a narrow-access person's results go to rows they may not see, and they come back empty-handed.
What is proposed, and what is executed?
The answer should be specific and enumerable. If a vendor cannot tell you exactly which actions happen without approval, nobody designed that boundary.
Who sets that boundary?
If the vendor set it and it is fixed, you are accepting a policy rather than buying a tool. Capability control belongs to the workspace administrator.
What gets sent to the model?
Ask whether there is a usage ledger you can actually look at — what text went, to which model, at what cost. Without it, a privacy claim cannot be checked.
Our answer to all four
The assistant sees with the asker's own eyes. Engagements, contacts, meetings and records are all filtered to that user's access, and the filter lives inside the query — so anything you are not allowed to see neither appears in the answer nor takes the place of a result you were allowed. Writing goes through the same door people use, with the approver's eyes.
The assistant proposes by default and does not execute. Actions are confirmed before they happen, and agent capability control (SPEC §18.5, deployed 2026-07-31) sits with the workspace administrator rather than with us.
The AI usage ledger records every call — what was sent and what it cost — and is available to you. Self-hosted, the language model can sit inside your own network, in which case nothing leaves at all.
And every answer the assistant gives carries its citation. That is slower than a confident answer, and it is deliberate: in work whose output goes to a client, an uncited answer has no practical value.
Common questions
Is our data used to train models?
No. Your data is not used to train any model. Hosted, text is sent to a model for processing and the usage ledger shows what went; self-hosted, you can keep the model inside your own network too.Can the AI be turned off entirely?
Yes. The free plan has no AI credits at all, and Pulse, feeds and the pipeline work fully within it — because none of those make a model call.
Try it on your own data
The free plan has no time limit. Bring the hardest document you have.