The challenge
The client’s onboarding was thorough, and that was the problem. A twelve-step process with five different CSMs touching each account meant every new customer waited four days to see the platform’s first piece of value. New hires could not run onboarding without two weeks of shadowing. As the customer base scaled past 1,800 accounts, the CSM team was burning out on what was, structurally, a checklist.
Three previous attempts to automate had failed. One was a no-code workflow that broke whenever a customer had a non-standard data source. One was a chatbot nobody used. One was an internal Notion playbook nobody followed.
The constraint: the new system had to work for the long tail of unusual customers, not just the median. It had to handle “this customer has Snowflake but also five Excel spreadsheets that get emailed in monthly” without escalating to a human every time.
The approach
A multi-agent system, not a single chatbot. We mapped the onboarding into four agent roles: Discovery (interviews the customer), Configuration (sets up the platform based on discovery), Verification (runs test transactions and validates output), and Handover (writes a summary doc and pings the assigned CSM only when the agents detect ambiguity).
A supervisor agent routes between them based on state. Each agent has its own eval suite: Configuration’s accuracy is measured against a golden set of 80 historical onboardings; Verification’s recall is measured against known-bad inputs.
Memory is shared across the four agents through Mem0, so the Configuration agent does not re-ask things the Discovery agent already learned.
What we built
We built a LangGraph state machine running on the client’s AWS account, with custom MCP servers wrapping their internal product API, their data ingestion service, and their Salesforce instance. Responses stream through their existing customer portal so the experience feels like one conversation, not four handoffs. A human CSM receives a clean handoff payload for any onboarding that the agents flag as ambiguous, and approves any irreversible account configuration before it is applied.
Eval suite: 180 examples covering all four agents, with regression gating in CI. A cost dashboard wired to LangSmith shows per-onboarding cost, which settled around $4.20, versus approximately $220 in CSM time for equivalent human work.
Build time: 6 weeks. Hardening: 2 more weeks.
Results
- 48× faster onboarding, average time-to-first-value down from 4 days to 2 hours
- 92% self-serve completion rate, only the genuinely weird 8% escalate to a CSM, with a clean handoff payload
- 3 CSM weeks reclaimed per month, redeployed to expansion accounts
- Cost per onboarding: $4.20, vs ~$220 in equivalent CSM time
- NPS unchanged, customers actually preferred the agent (“I did not have to schedule a call”)
Tech used
LangGraph · Claude Sonnet 4.6 · Mem0 · custom MCP servers · LangSmith · AWS Bedrock · Pinecone · their existing Salesforce + product APIs
Takeaway
The outcome was not replacing CSMs, it was letting CSMs do the work that actually needs a human. Production AI agents worked here because the four roles in the onboarding were genuinely different jobs. One agent trying to do all four had been the failure mode of the previous attempts.