THE UI IS THE AI.
OURS HAS FIVE KEYS.
Two CRM vendors spent the summer arguing about where the screen should live. Neither answered the question underneath it: what a person actually touches once the agent has done the work. CMP42 answers it with five keys and one signature.
On 31 August, every HubSpot portal in the world woke up grey. The new theme had been optional since June; on the last day of August it stopped being optional, and the announcement promised there was "nothing to relearn". Five days earlier, in a press release about something else entirely, Marc Benioff had written a sentence that made the whole exercise look odd: "Here, the UI is the AI."
Both companies are answering the same sentence. One is bringing the model into its own screen. The other is shipping its CRM into somebody else's. We think both answers skip a step, and that the step they skip is the only part a person in a regulated company actually cares about.
Everyone agrees the UI is the AI. Then they hand you a chat box.
Language is excellent for intent and terrible for movement. Nobody wants to type a full sentence to mean "next row". Nobody wants to wait two polite seconds for an assistant to confirm that it understood "down". Spend a working day in a pure chat interface and you notice what the demo never shows: the old software made you learn everything, and the new software makes you say everything.
The second problem is harder than the first. When every instruction is a sentence, every instruction is also an interpretation. :archive and "could you archive that one" are not the same claim about the world, and in a company that gets inspected, the difference is the whole point. "The assistant probably did what I meant" is not a line anyone wants to read back out of an audit trail.
So the question is not whether the interface becomes intelligent. It already has. The question is which parts of it are allowed to be.
Five keys, and two of them never ask the model.
CMP42 has a grammar with five ideas in it, and only five. Slash asks. / dormant accounts in Bavaria. / why is this deal stuck? You do not need to know which tab hides the answer, or which three objects have to be joined to produce it. The model works that out, and the table in front of you rearranges itself around the question.
Colon does. :merge. :assign Sarah. :archive. The execution may be clever. The meaning is not negotiable, and it reads the same for every user, in every tenant, to every reviewer.
J and K move. No model, no round trip, no waiting. Moving through a hundred thousand rows triggers no API call at all, because the intelligence is not allowed anywhere near your thumb. Enter opens a record, a derived value's lineage, or the path the model took to an answer. Escape leaves a dialog, a level, or an answer you have already stopped needing. That is the whole vocabulary: small enough to learn in an afternoon, small enough that after a week you forget you learned it, and small enough to survive the keyboard, because a slash drawn with a finger means ask and said out loud, "ask" means ask.
Underneath it, the same boundary applies to agents. By command they rearrange views and columns. Commands that write or delete values are closed to them. And every write names the tenant, the person, and whether a person or an agent made it. The speed is on the record too, with its machine: 1,000,000 records in 2.1 seconds, and 16,000,000 contacts, both on a small two-vCPU machine.
The agent writes the report. A person still presses the button.
Watch what happens to the interface when the work moves to an agent. The rep no longer fills in forty fields. They read a draft and confirm it. The screen does not disappear; it shrinks to the one moment where a human says that what the machine wrote is true.
In a regulated company, that moment is the product. Which is why the release on 15 October is built around evidence rather than around assistance.
Consent is an event, not a checkbox: the time, the channel, a fingerprint of the address, and a reference to the exact version of the text the person agreed to. Confirmation is a person pressing a button, because opening a link confirms nothing and never did. Withdrawal is another event, and the proof of it survives the erasure of the person it belonged to. What the contact shows you are derived columns, locked against hand edits, so nobody can quietly improve the history. And channelSENDER sends over the list the check has already cleared, rather than over the list somebody exported last Tuesday. CMP42 keeps the evidence; whether a given campaign is lawful stays the sender's call, which is the honest division of labour and also the only one that survives contact with a works council.
One more piece ships with it, and it is the part that makes the opening argument concrete: a remote MCP connector, with the tools generated from the customer's own model. So the assistant your team already uses can ask CMP42 questions and take the operations the command boundary allows it to take, in your own vocabulary, without a screen in the middle. The UI can be the AI. The record of who did what stays here.
What is not in this release
Everything above goes live on 15 October. The following is in release planning or on the roadmap, and is not in the product on that date:
- Erasure that reaches into the audit trail
- Retention and legal hold as an enforced refusal, not a policy document
- Tenant separation enforced by the database itself
- Access control per Model, with fields masked per role
- Personal values encrypted per person, so deletion destroys the key
Treat that list as intent, not as a commitment. Statements about future functionality, sequence and timing are forward-looking: they describe what we are currently planning to build, they depend on what we learn from the releases in between, and they may change or be dropped. Nothing on it should be part of a decision you have to defend later. If a feature on that list is the reason you would buy CMP42, say so in writing before you buy, and we will tell you where it actually stands.
The screen is not going away. It is getting smaller and sharper.
The vendors arguing about whose window the work happens in are arguing about the wrong half. Over the next few years, most CRM work will leave the CRM: it will happen in chat windows, in agents that run overnight, in connectors inside assistants that belong to other companies. What stays behind is review, approval and the exceptions.
Which means the remaining interface carries more weight per pixel than any screen that came before it, and that a grammar with five keys in it is not nostalgia. It is what is left when the rest has been automated. If you want to see it against your own data model, start here, or write to us and bring the awkward requirement first.