The gap between how sales teams are supposed to work and how they do work is mostly administrative. Call notes get written days late or never; CRM fields sit empty until pipeline-review day; follow-up emails and proposals get rebuilt from scratch each time. This is prime territory for Claude because the raw material — call transcripts, email threads, past proposals — already exists as text, and because the consequence model is forgiving: a rep reviews every output before it reaches a customer or a system of record.
The problem: selling time lost to writing time
Ask your sales operations team where deals stall and data goes missing, and the answer is usually after the call: the summary that never got written, the next step that never got logged, the proposal that took a week because it started from a blank page. The aim is not to automate selling — it is to make the administrative residue of selling nearly free, and to make CRM data complete enough that forecasts and handoffs stop depending on memory.
A reference design
The division of labor follows the same rule as every good LLM workflow. Claude does the language work: turning a call transcript into a structured summary (situation, needs, objections, next steps), proposing values for CRM fields from what was actually said, drafting the follow-up email, and assembling proposal first drafts from your approved content library against the prospect's stated requirements. Deterministic code does the consequential work: pulling transcripts and account context, writing to the CRM only after the rep confirms, and inserting pricing from the price book — never from the model. The rep reviews everything: a suggested CRM update is a diff to approve, not a silent write; a follow-up email is a draft in the compose window, not a sent message.
Model tiering is natural here: Haiku 4.5 handles field extraction and classification cheaply at volume; Sonnet 5 drafts customer-facing prose; complex multi-document proposal work can escalate to Opus 4.8 when it earns it.
Where humans stay in charge
Three review gates keep this safe. Nothing writes to the CRM without the rep approving the specific field changes — this keeps reps accountable for data quality rather than blaming "the AI." Nothing goes to a customer unsent-by-a-human — follow-ups and proposals are drafts until a person presses send, and reps must check that the model has not "helpfully" paraphrased a commitment the rep never made. And pipeline judgments — deal stage, close probability, forecast — remain the rep's and manager's call; Claude can summarize evidence for the call, but the judgment is a human one that reps must own in front of their managers.
Rollout advice
Start with call summaries and follow-up drafts for one team: it is the highest-frequency task, the value is felt the same day, and it requires no risky integrations. Put the output where reps already work — the CRM record, the email client — because a separate tool will not get used. Measure honestly: time from call to logged summary, CRM field completeness, and whether reps keep using it after week four. Expand to proposal drafting once your content library is curated; like proposals anywhere, drafts are only as good as the approved material behind them. Everything this design needs — Messages API, tool use, streaming — runs on all four platforms, so build wherever your CRM integrations and cloud identity already live.
Pitfalls
The common failures: auto-writing CRM fields without rep confirmation (data quality problems become invisible instead of solved); letting drafts state prices or commitments; transcribing and processing calls without checking your recording-consent obligations by region — a legal question to settle before the pilot, not after; and measuring adoption by logins rather than by whether notes actually get logged faster. Customer conversations are also confidential business data: keep the pipeline inside your governed cloud boundary and confirm data-handling specifics with your provider.
Where to go next
Call summarization mechanics are covered in Meeting Notes and Action Items, Automated. If your team also produces outbound content, Marketing Content Workflows with Human Review shows the same draft-review pattern applied to campaigns. For first steps on your cloud, see the quickstart.