Turn account research into CRM field updates

Stand up a **CRM Mapping** agent — ingest notes, call summaries, or read-only CRM context — output **proposed field values** (next step, pain, stage hint, and competitor mentions) in a paste-friendly shape — **you** apply or approve writes — optional MCP ([Add an MCP server](/tutorials/add-an-mcp-server)) stays narrow.

Plus: three Admin-Agent passes — build Auxot→CRM column mapping from your vendor's export headers, draft **PROPOSED_ONLY** JSON block contract, and rehearse rejection when sources conflict.

Audience	Everyone · Admins
Time	~12 min
Prerequisites	Sales context exists ([Add your first context file](/tutorials/add-your-first-context-file) — ideal customer profile, stages, and when to disqualify). Building agents ([Create an agent from scratch](/tutorials/create-an-agent-from-scratch)). Helpful: lead rhythm ([Triage and follow up on inbound leads](/tutorials/triage-and-follow-up-on-inbound-leads)), call prep ([Prep for your next sales call](/tutorials/prep-for-your-next-sales-call)), MCP CRM read ([Add an MCP server](/tutorials/add-an-mcp-server)). If automation writes CRM rows without humans, pair gates ([Require human approval before risky actions](/tutorials/require-human-approval-before-risky-actions)).
You'll end up with	One CRM Mapping agent — outputs mapped fields + confidence + source bullets — plus a one-page schema cheat sheet your reps paste beside their CRM until integrations mature.

When a tutorial shows italic text in quotation marks, it usually mirrors a label or helper string inside Auxot. Product copy changes between releases — if something reads differently in your workspace, trust what you see on screen.

Callouts with a Worth knowing gold accent are meant as must-read context before you move on. Blockquotes that open with Tip are lighter, optional depth.

Why this matters

Research piles up in Chat and Slack; CRM stays stale. The failure mode is not missing automation; it is missing structured handoff. Reps copy paragraphs into Notes until nobody searches them.

A CRM-mapping agent does not replace judgment. It translates messy narrative into field-shaped proposals: Suggested next step, Pain hypothesis, Role clarity, and Risk flags, aligned to columns you defined. Humans paste or approve; CRM stays canonical.

Lead triage (Triage and follow up on inbound leads) optimizes first-touch mail. Call prep (Prep for your next sales call) optimizes the meeting brief. This lesson keeps CRM rows current after either; same pipeline, different artifact.

Nothing writes CRM on its own — you confirm proposals, you own duplicate activity rows.

Quick start

Inventory CRM fields — export header row or screenshot allowed dropdown values: store cheat sheet context file (Add your first context file).
Create CRM Mapping — job description mandates:
- Inputs: pasted notes or MCP-fetched account snippet (Add an MCP server).
- Outputs: markdown table CRM_FIELD | PROPOSED_VALUE | CONFIDENCE high/med/low | WHY (<=15 words).
- Never claim sync happened: end with STATUS: PROPOSED_DRAFT.
Dry-run one account — real yesterday conversation; verify proposals fit allowed CRM values (no invented stage names).
Paste ritual — reps block five minutes Friday: apply the CRM-mapping agent output; or push through integration only after human tick (Require human approval before risky actions).
Audit trail — link the Chat thread or the Slack paste ID in the CRM note so anyone reviewing the row can find the original conversation (View your audit logs when Auxot executed tools).

Done? CRM-mapping agent selectable; cheat sheet published; one account updated honestly end-to-end.

The agent can do that?

1. Schema translation

Chat → Admin Agent:

CRM export headers pasted: […]. Our internal vocab: [bullets]. Produce mapping table CRM column ↔ Auxot question the CRM-mapping agent must answer — markdown — max 12 rows.

Why it’s non-obvious: Vendors love opaque API names; translation table prevents garbage-in columns after you paste headers.

2. Conflict guardrails

The CRM-mapping agent saw conflicting revenue signals in two pasted paragraphs — rewrite instructions — two sentences — demand **CONFLICT** row listing both interpretations — human picks — no merge.

Why it’s non-obvious: Models tend to merge contradictions; forcing explicit conflict preserves trust until you edit prose.

3. Batch sanity

We're updating 30 accounts from spreadsheet exports ([Batch spreadsheet rows through a workflow](/tutorials/batch-spreadsheet-rows-through-a-workflow)) — ordered QC steps before trusting the CRM-mapping agent — bullets — assume junior rep runs paste.

Why it’s non-obvious: Volume amplifies confident mistakes; checklist slows disasters after you admit batch mode.

Go deeper

CRM write paths

Pick the path that matches your CRM. Salesforce: official hosted MCP, generally available April 2026, Enterprise Edition or above. HubSpot: official MCP, free tier covers the connector. Pipedrive: no first-party MCP; use the Composio wrapper for managed OAuth.

What keeps any of them safe is a field-level permission set on the CRM side, configured before you wire the MCP. The agent reads everything; the agent writes only to one custom field, like auxot_research_summary. Stage and revenue stay read-only to the agent because a wrong write there moves forecast.

Pair the agent’s tool policy with the CRM-side field permission (Define a tool policy, Manage your Credentials): two locks, one on each side. The full per-CRM wiring (connected app + integration user + field-level permission set, plus the operating note) lives in Lock CRM writes to a single custom field.

Workflow handoff

Research → the CRM-mapping agent → human CRM step mirrors maker/checker (Use two-person rules for high-impact actions) when writes touch revenue.

Slack

Drop the agent’s output tables into deal channels (Connect Slack to your agents): async approval thread.

Regression

If CRM-mapping agent output shifts after model changes, add rows (Catch regressions after you change an agent).

Walkthrough

Step 1: Cheat sheet v1

Pick seven fields max; depth beats covering forty poorly.

Step 2: Author negative examples

Show the CRM-mapping agent deliberate bad input; instruct refusal pattern.

Step 3: Pilot three accounts

Revise confidence wording until reps trust low labels.

Step 4: Train paste chore

Short screen recording or wiki GIF: the CRM-mapping agent’s output → CRM clicks; under three minutes.

Step 5: Review monthly

Spot-check five records: fix cheat sheet; retire stale dropdown lists (Run a quarterly review of your agents tone).

What’s next

→ Triage and follow up on inbound leads. First-touch upstream; the CRM-mapping agent cleans CRM after conversations mature.
→ Stress-test your pitch with simulated buyer personas. Rehearse objections before live calls once the CRM-mapping agent keeps facts honest.
→ Prep for your next sales call. Briefs feed the CRM-mapping agent’s inputs; paste talking-point deltas post-call.
→ Add an MCP server. Read CRM context without export gymnastics; keep write paths rare.
→ Lock CRM writes to a single custom field. The canonical field-level safety pattern when the CRM-mapping agent does write back: read everything, write only auxot_research_summary.
→ Batch spreadsheet rows through a workflow. When imports arrive as CSV waves, row_key keeps accounts traceable.
→ Require human approval before risky actions. Before any bot pushes CRM commits unattended.

Reference

Manual: MCP servers
Pages in Auxot: Chat, Settings → Agents, Settings → Tool Connector Keys
See also: Turn incident lessons into updated runbooks, Stress-test your pitch with simulated buyer personas, Add your first context file, Define a tool policy, View your audit logs