The AI drafts. You send.Never the other way round.
The most useful thing we can tell you about the AI is the list of things it cannot do, because that list is written into the product as a rule rather than left as a setting somebody switches on later.
It summarises, drafts and ranks. It does not send, close a deal, delete a record or contact anybody. It cannot read a record the person asking is not permitted to see, and it will not invent a fact to fill a gap.
6 things, and what each one is for.
Customer and lead summaries
Ask for a summary of a lead or an opportunity and you get one, together with the specific CRM fields and activities it read to produce it. It runs inside the asker's permissions, so it cannot summarise a record they are not allowed to open.
A seller enters the next conversation informed without rereading every activity.
Summary · fields and activities read · permission filtered
Priority and next-best-action
It tells you which work looks most worth doing next, so the queue is ordered rather than alphabetical, and proposes the next action on a record.
The day starts from a considered list instead of the loudest item in the inbox.
Ranked queue · proposed next action · human decision
Follow-up drafting
Follow-up emails, campaign copy, the message you have been putting off since Tuesday. Every draft is marked as AI-generated and stays unsent until an authorised person edits it and approves it. There is no setting that turns that off, because it is not a setting.
Writing time goes down without handing the customer relationship to a machine.
Draft · marked AI-generated · unsent until approved · no autonomous send
It says when it does not know
When the data is thin or out of date it reports that the evidence is insufficient. It will not invent a company size, a revenue figure or an intention the customer never expressed. Text arriving from outside the system, in an imported email, a note or an integration payload, is treated as data and never as an instruction, so a message engineered to manipulate the assistant does not get to.
That rule is the difference between an assistant and a liability.
Insufficient evidence response · external text treated as data
The lead score is not the AI at all
Worth separating, because people assume otherwise. Your lead score is a deterministic calculation from criteria you configure. It is arithmetic, not a judgement, and you can open any score and see how much each factor contributed to it.
You are never asked to trust a number you cannot take apart.
Configured criteria · per-factor contribution · recalculated on change
AI work queue and audit trail
Recommendations land in a decision queue with accept, reject and failure states, and every AI action writes an append-only audit event carrying who ran it, what it touched and whether it succeeded.
AI work is managed like accountable work rather than hidden magic.
Queue · accept/reject · terminal failure · append-only audit
Four steps, in this order, every time.
- 01
Ground
Load only the CRM evidence this user may already see.
- 02
Propose
Return a summary, a ranking or a draft.
- 03
Review
A person edits, approves or rejects it.
- 04
Record
Write the audit event, or fail without changing anything.
The next three questions.
Thirty minutes, and you leave with three answers.
Which package fits, what we would migrate and how, and what the fixed fee would be. Written down, so you can put it in front of a partner or a board.