CRM UIS ARE DYING.
AI IS THE UI.
But AI without context is autocomplete on a hallucination. CMP42 is the context: a CRM built knowing that the primary reader will be an agent, and the primary writer a human helped by an agent.
CRM IS A GRAPH PROBLEM NOW.
Something quiet has happened in the last twelve months. Ask a sales manager how they check the pipeline. A year ago the answer was “I open HubSpot.” Today the honest answer is “I ask my agent to summarise it, and if the summary looks off, I open HubSpot.”
Ask a partner in a consulting practice how they draft the follow-up note after a client call. A year ago it was Word. Today it is Claude, with a paste of the meeting transcript and a “make it sound like me.” The database has not moved. The workflow has.
The CRMs we have — HubSpot, Salesforce, Pipedrive — were designed on an assumption that turns out to be temporary: that a human would open the UI. When an agent tries to read the same schema, it treats it like a chat history. It hallucinates the missing bits, invents fields that were never declared, fabricates a stage that is not in the pipeline. Not because the model is bad. Because the schema was not built for the reader.
CMP42 is what happens if you build a CRM knowing the primary reader will be an agent. Three deliberately small parts — Context, Models, Processes — and a fast, boring API that any agent can read and write against.
BECAUSE THE WORKFLOW ALREADY MOVED.
Every honest CRM user we speak to has one of two setups:
Wired up, inconsistent
The CRM is connected to an AI agent through Zapier and a lot of Python. The results are inconsistent — because the schema was not built for agent reading.
Given up, copy-pasting
The integration never happened, or never worked. Transcripts and exports get pasted into ChatGPT by hand, one conversation at a time.
Either way, the CRM is not the tool that runs the day. It is the database that gets consulted after the fact. If that is the honest role, let us build it for that role.
“CRM UIs are dying. AI is the UI now. But AI without context is autocomplete on a hallucination.”
BUILT FOR CLICKERS. BUILT FOR AGENTS.
A conceptual comparison of two design assumptions — not a feature comparison with any specific product.
| Built for human clickers | Built for agents | |
|---|---|---|
| Primary reader | A person opening the UI | An agent reading structured context; a human helped by an agent writes |
| Why a field exists | Because someone will look at it or click on it | Because it is declared in the ontology — defaults lean, extension explicit |
| Relationships | Inferred by a person from layout, labels and free text | Typed edges: a contact knows a company, owns a deal, authored a doc |
| Schema changes | Settings behind tabs, owned by whoever administers the tool | Models in git-versioned YAML, migrated with one command, backwards-compatible |
| Invalid input | A person notices and fixes it later — or does not | Validated at write-time with clear errors, so agents can self-correct |
| Interface | Dashboards, pipeline boards, forms | No forced UI — ChatGPT, Claude, Slack or the app you build next |
| AI | Features layered on top of the interface | Bring your own model; the CRM supplies grounded context |
| Accountability | Designed around who clicked what | Every write attributed to a human or agent; process runs replayable, token cost visible |
SAME DATA. DIFFERENT SHAPE.
CMP42 does not ask you to collect more. The contacts, companies and deals you already have are enough to start. So are the documents, meeting recordings, product briefs, playbooks and messy Notion pages that never made it into the CRM.
What changes is the shape. Instead of five silos — people in the CRM, content in a drive, knowledge in a wiki, schema in someone’s head, process in someone’s habits — the same data becomes one connected object graph. Every object versioned, every edge typed, every field addressable by URI.
Who you work with
Contacts, companies, deals, tasks, activities — as typed objects with relationships that are edges, not free-text fields.
What you produce
Documents, recordings, briefs and proposals, linked to the people who authored them and the accounts they are about.
What you know
Playbooks and frameworks your team keeps re-explaining — as context an agent can use on the next engagement.
MORE IMPACT FROM WHAT YOU HAVE.
Impact, in CMP42’s sense, is what agents can do reliably with your data — and what you can verify afterwards. We publish no figures until Design Partner case studies provide real ones.
Answers with citations
An agent walks the graph and answers with references to source objects — not with vibes. Humans can follow every citation.
Drafts from real history
The follow-up or renewal note starts from the actual account, the actual engagement, the actual last conversation.
Work you can audit
Processes log every transition and record what each agent read, what it wrote and at what token cost. Replay any run against a past snapshot.
Agents you have not chosen yet
The same graph serves ChatGPT, Claude, your team’s Slack bot and the app you build next month — without a new integration for each.
NO PLATFORM BLOAT. NO HIDDEN MAGIC.
The list is deliberate. It is also the reason CMP42 exists.
What we refuse to build
- No forced UI. The admin surface is a data table and a migrations CLI. That is it. Your team already has an interface — ChatGPT, Claude, Slack, or the app your engineers build next month. CMP42 does not compete with it; it feeds it. Headless →
- No “AI insights” layer. We ship no proprietary model. If we shipped a black box, we would be part of the problem we set out to fix. Bring your own model ↓
- No 200-field contact object. Defaults are lean. Extension is explicit. Silence is a feature. Lean defaults →
- No drag-and-drop pipeline builder, and no dashboards that go stale on Tuesday — until you write them yourself.
- No cookie banner, no third-party scripts. Same posture as everything else protagx ships. Zero tracking →
What we ship instead
- MCP tools, day one. Every model becomes a set of MCP tools automatically. MCP tools →
- REST + GraphQL parity. Every operation for non-MCP clients, with the same auth and the same audit trail. API →
- Git-versioned models. Your ontology is a first-class product artefact, not a hidden setting. Models →
- Replayable processes. Audit an agent’s work the way you audit a git branch. Processes →
- Self-hostable. Managed by us in the EU, or on your Postgres and object store. Same binary, same models. Self-hosting →
- A boring API that gets out of the way.
BRING THE MODEL YOU TRUST.
Most software now arrives with an “AI insights” panel: a proprietary layer that reads your data and tells you what it thinks. CMP42 deliberately does not. We give your agent grounded context; your agent produces the insight.
There are three reasons. Transparency: an insight layer you cannot inspect is a black box, and a CRM that hallucinates behind a friendly panel is the problem, not the fix. Longevity: models change quickly — the context layer should outlast whichever agent ships this month, and feed the ones you have not chosen yet. Governance: which model provider may see your customer data is your organisation’s decision, not your CRM vendor’s.
On the managed EU tier, CMP42 does not train on your data, and nothing leaves except what you explicitly send to an agent you connect. Self-hosted, nothing leaves your environment at all.
Works with anything MCP-compliant
- Claude Desktop and Claude Code
- Cursor and Cline
- Any custom agent built on the MCP SDK
- REST + GraphQL for non-MCP clients — same auth, same audit trail
If this describes the mess on your desk
Come and Alpha it with us. Twelve Design Partners, six months from October 2026, €6,900 one-time.
→ Apply to be a Design Partner