The Console
No rule builder, no admin queue. Describe what should be true of your CRM, and it shows you every run it intends to start before anything is written.
The gap this closes. Everyone already knows what their CRM should look like. What nobody has is the afternoon it takes to make it look that way, or any confidence that the thing doing it understood the instruction.
You type a sentence you would say out loud on a handover call. Dedupe the accounts we inherited, standardise them, fill the gaps, tag them as acquisition sourced, route them by our territory rules, and tell each owner what landed in their patch.
Reading happens straight away. How many records match, what your rules currently say, what a normalization scheme would change, what an enrichment would cost. Those are answered in the conversation, because reading your own data is free and asking permission to look at it wastes your time.
Anything that would change a record stops. It becomes a run in a plan, with its scope and its record count, and nothing runs until you confirm. Eight clauses in one sentence become eight runs you read in order, rather than eight things that already happened.
The line it draws
Every turn sorts itself into one of these two, and the sorting is structural rather than a judgement the model makes again each time.
Reads. They touch nothing, so they need no permission and no confirmation.
Writes. Each becomes a run in a plan, scoped and counted, waiting on you.
Honesty
The distinction matters more than the reassurance, so the product says it out loud on the surface itself.
Holding every write for confirmation is a property of this surface. It is the right default for a person typing a sentence, because a sentence is ambiguous and the cost of acting on the wrong reading is somebody else's CRM.
A credential calling the tools directly is a different situation and gets a different answer. What an agent may do on its own, and what it must bring back as a proposal, is a permission you set at the tenant level, not a prompt it can talk its way around. Creating runs and dispatching them are separate permissions, so an agent can be allowed to propose everything and execute nothing.
What it leaves behind
A chat window that forgets is a demo. Everything here outlives the tab it was typed in.
The conversation itself is saved too. It is still there after a reload, in another tab, or tomorrow.
For agents
Everything the Console composes is available to your own agent over MCP. Reads are free; writes are a grant you issue.
count_crm_records how many a scope selects readlist_filterable_fields the fields and operators your CRM offers readcreate_plan compose an ordered sequence of runs writeget_plan read one back, run by run readdispatch_run start a run, when that grant is held writeWhere to start
The readiness scan reads your CRM and reports what it finds, before you describe a single change.