Most small sales teams already pay for HubSpot or Pipedrive. The records go stale within weeks, because updating them is nobody's job, and the team works the pipeline from memory and inbox search. Another tool rarely changes that. A few agents with narrow jobs can: they read the same records the team does and put the next step, with a draft attached, in front of a person who approves it.
The three-agent model
Three agents, each with one job, all reading from one retrieval layer. Splitting the work this way makes each agent testable on its own, which a single do-everything prompt never is.
- BaseRetrievalCRM records, call notes, documents, signals.
Broken whenAnswers mix up accounts, or miss what was said on last week's call.
- Agent 01ScoutEnrichment: a typed context block per account.
Broken whenSomeone opens a record and immediately opens LinkedIn.
- Agent 02ScribeDrafting: first versions of every outreach and follow-up.
Broken whenDrafts are rewritten from scratch more often than they are sent.
- Agent 03SentinelMonitoring: a ranked morning list, drafts attached.
Broken whenA deal goes quiet and the team hears it from the customer.
Agent 01Scout, enrichment
Scout reads each record and builds a context block: company signals, recent activity, what the account shipped last quarter. The output is a typed object the next agent can use, not prose, refreshed weekly or when a deal changes stage.
Agent 02Scribe, drafting
Scribe writes the first version of every outreach, follow-up and proposal, from Scout's context, the deal history and the rep's own voice. Each draft is checked against a schema and graded against the rubric before anyone sees it.
Agent 03Sentinel, monitoring
Sentinel watches the pipeline for changes in behaviour: deals that went quiet, deals about to close, signals nobody picked up. It never moves cards. It produces a short ranked list each morning with drafts attached, and a person approves or redirects each item.
Underneath, all three read the same retrieval layer: a vector search returns candidate passages, a reranker keeps the best few, and the agent writes against those. Retrieval is tested separately from generation, with its own fixtures and its own gate.
Starting targets per agent
Each agent ships with its own eval harness: real labelled examples, a written rubric, and a regression gate that blocks a release when a score drops. These are our starting targets. We tune them per workload in the Discovery Sprint and reset them after the first month of live data.
- Scout. Enrichment accuracy of 0.92 or better, no context older than seven days, structured output valid every time.
- Scribe. Draft quality of 4.2 out of 5 or better on an LLM-graded rubric, tone match of 0.90, faithfulness to the source of 0.95.
- Sentinel. Precision of 0.80 and recall of 0.75 on the signals it raises, surfaced within four hours of the change.
Three scales, one table
The agents stay the same from a founder selling alone to a team of fifteen. What grows is the boundary work: voice profiles per rep, views per role, and keeping one person's notes out of another person's drafts.
| Layer | Founder plus assistantsScale 01 | Sales team of 3 to 15Scale 02 | Several teams or regionsScale 03 |
|---|---|---|---|
| Scout | Enrichment on every new lead, weekly refresh on active ones | Shared enrichment plus each rep's relationship history | A retrieval index per team |
| Scribe | Drafts every outbound and follow-up | A voice profile per rep, drafts routed to the right person | Voice profiles and templates per team and jurisdiction |
| Sentinel | A short morning list with drafts | A team view and a view per rep | Alerts per team, an anonymised view across teams |
| Surface | One web dashboard, readable on a phone | Web plus Slack for urgent items | A dashboard per team and an admin console |
| Cost band | CHF 12,000 to 18,000 | CHF 15,000 to 25,000 | Scoped per case |
| Build time | A focused MVP can be ready in two weeks. A scoped product can launch in around two months. We agree the timeline after reviewing your requirements. | A focused MVP can be ready in two weeks. A scoped product can launch in around two months. We agree the timeline after reviewing your requirements. | Scoped per case |
| Live example | No named engagement yet | No named engagement yet | Not built yet |
CRM intelligence stack at three scales
Scale 01Founder plus assistants
- Scout
- Enrichment on every new lead, weekly refresh on active ones
- Scribe
- Drafts every outbound and follow-up
- Sentinel
- A short morning list with drafts
- Surface
- One web dashboard, readable on a phone
- Cost band
- CHF 12,000 to 18,000
- Build time
- A focused MVP can be ready in two weeks. A scoped product can launch in around two months. We agree the timeline after reviewing your requirements.
- Live example
- No named engagement yet
Scale 02Sales team of 3 to 15
- Scout
- Shared enrichment plus each rep's relationship history
- Scribe
- A voice profile per rep, drafts routed to the right person
- Sentinel
- A team view and a view per rep
- Surface
- Web plus Slack for urgent items
- Cost band
- CHF 15,000 to 25,000
- Build time
- A focused MVP can be ready in two weeks. A scoped product can launch in around two months. We agree the timeline after reviewing your requirements.
- Live example
- No named engagement yet
Scale 03Several teams or regions
- Scout
- A retrieval index per team
- Scribe
- Voice profiles and templates per team and jurisdiction
- Sentinel
- Alerts per team, an anonymised view across teams
- Surface
- A dashboard per team and an admin console
- Cost band
- Scoped per case
- Build time
- Scoped per case
- Live example
- Not built yet
What not to build
Most CRM AI demos sell one of these. Each looks good in a meeting and costs the team time within a month.
- A chatbot in the corner of the CRM. Asking a model questions one at a time is a side quest. It adds steps instead of removing them.
- A weekly lead score. A number from 0 to 100, written by a batch job once a week, stops being checked by the third week.
- One agent for everything. A single prompt that enriches, drafts and monitors drifts with every model update and cannot be tested.
- Drafting without a faithfulness check. Drafts slide toward fluent and wrong, and you find out when a customer does.
- Switching CRMs to get the AI. The intelligence layer reads and writes; it should never require a migration.
Where we stand
We have no named, live CRM intelligence engagement to point to on this page. The shape above is what we propose and build. When a client agrees to be named, the engagement goes here with its numbers.
The place to start is a two-week Discovery Sprint. We sit with the real sales motion, find where deals leak, and write the three agent specs and the rubric before anything is built. If the answer is that you do not need agents, the sprint says so.
Self-audit
Twelve questions for the CRM you run today. The score updates as you tick; what the bands mean is explained once on the stacks index.
Method and pricing
The layer model, how to read the cost bands, the notes on Swiss and EU data protection and the EU AI Act, and how these reports are versioned are shared across all five stack reports: how we write the stack reports. The eval method behind every agent is the Eval Spec.