LangChain's sales agent built 1,540 deal documents in 12 weeks. Deals that got one advanced 47% of the time versus 16% without
The most useful line in LangChain's write up is that most of the quality came from the context they gave the agent, not from prompts.
What happened
LangChain says its internal GTM Agent generated 1,540 sales artifacts over the last 12 weeks: account plans, proof of value documents, proposal decks and onboarding plans that reps used to build by hand. Deals where a rep sent one of these documents moved to a later stage 47% of the time, compared with 16% for matched deals without one. Artifacts that were sent to customers accounted for 337 hours of saved rep time.
The agent reads the Salesforce record, product usage, call transcripts, risks, decision criteria and next steps. It copies a leadership owned template into a sandbox, fills it, renders it to PDF, looks at the pages and fixes layout problems, then files it as a native Google Slides, Docs or Sheets file in the account's shared Drive folder. The rep gets a preview in Slack, with flags on anything a human still needs to fill in. If inputs like an economic buyer or success criteria are missing, the agent asks the rep first.
One fix mattered: early versions kept re-reviewing pages that were already fine. A rule to ship once a review finds no real defect took a one pager from 18 minutes to 7.
My take
This is the pattern I push on every CRM build: the agent is only as good as the deal data it can read. If stages, decision makers and next steps are not in the CRM, no model fixes that.
What I would copy for a smaller team on HubSpot or GoHighLevel:
- Tie each document to a pipeline stage, so the agent knows what is missing before a call.
- Let it ask for missing fields instead of guessing.
- Keep templates owned by a person, so one update reaches every rep.
- Put a stop rule on any self review loop.
- Track whether documents get sent, not just generated.
That 47% versus 16% is LangChain's own data on matched deals, not a controlled trial, but sent rate and stage movement are the right two numbers to watch.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
