---
title: "What it took to build an agent our CSMs actually trust - Atlan Frontier Labs"
url: "https://atlan.com/frontier/essays/an-agent-our-csms-actually-trust/"
description: "Four attempts in under two years, three of them failures. How customer success at Atlan went from a chatbot you had to ask to an agent that puts the draft in your inbox before you ask - and what it cost to earn that trust."
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

An essay from Atlan Frontier Labs (Operating Notes, essay 006) by Dhruv Saharya, Customer Success, Atlan, dated 2026.09.18, 2,735 words. Four attempts in under two years, three of them failures: how customer success at Atlan went from a chatbot you had to ask to Relay, an agent that puts the draft in your inbox before you ask. Part of [Becoming Frontier](https://atlan.com/frontier/).

## Excavation work before every call

An account with Atlan for three years means something like 100 calls, 20 key people across the customer's organization, and 300 support tickets, seven from the last week alone. Before a call, a CSM dug through it by hand: calls in one tool, tickets in another, usage numbers elsewhere, plus whatever the previous CSM remembered. The size of that dig capped how many accounts one person could carry and how much attention each customer got. Earlier fixes from 2024 (a tool for searching internal information, another for pulling account data against saved questions) helped narrowly, but neither remembered anything between conversations.

Now: a few minutes after a customer call ends, the Customer Success agent Relay puts a finished draft email in the CSM's drafts ("Based on your last conversation, here's what I'd suggest sending: ..."), ready to read and send. Relay is part of Atlan's work to become a [frontier company](https://atlan.com/frontier/essays/becoming-a-frontier-company-in-the-open/), using AI to get more ambitious, not just more efficient; in Customer Success that means less time on account notes and manual work, more time with the customer.

## Two agents on WhatsApp

When [OpenClaw](https://openclaw.ai/) launched in late 2025, the author and Himanshu (who leads field engineering at Atlan) connected their own Claws to their messaging accounts, gave them context about who they were, and pointed them at the excavation problem. The agents exchanged about a hundred messages, worked out from Atlan's org chart that Himanshu was the author's skip-level manager, and Himanshu's agent started assigning tasks to the author's agent until it was satisfied. Then, on their own, the two agents decided the result was worth showing to the CEO for approval. That meeting request was not carried out.

It was a wakeup call: the technology was good enough to build something autonomous inside customer experience, and a company helping customers become AI-native should walk the walk. The author flew to India to meet Himanshu; two days of a mini hackathon at his house, in harnesses and at a whiteboard, produced the foundation of Relay.

## Our previous attempts

1. Early 2025: a "talk to the data" chatbot in a Slack channel. The CSM picked the data source and days of history and got an answer in the thread. Not agentic: every question was a hand-assembled pile of context; too much and answers got worse, further and it could not answer. A parallel experiment auto-posting a summary note to a shared doc after every call failed almost immediately, because no one reads a note they never asked for.
2. June 2025: an agent framework on [n8n](https://n8n.io/) let the agent choose what to pull. Same wall from the other side: limited context window, and a cost every time it looked at one more thing.
3. A couple of months later: a multi-agent system. An orchestrator called Hermione routed questions to specialist sub-agents (product data, strategy, commercial context, and so on). It worked decently; launched at a company-wide meetup, most of the team used it within months. Limits: slow, flat over-explained AI tone, and still something a CSM had to ask. "It was a better copilot, but it was still a copilot."

What all three did: each got better at answering questions about customers. What none did: anything with the answer. Atlan tries internal tools in the open, lets them fail and keeps what worked. The lesson matched one from building the [context layer](https://contextandchaos.substack.com/p/what-an-enterprise-context-layer): getting AI right depends less on model intelligence and more on getting the underlying context right.

## Teaching Relay to remember

Relay, the fourth attempt, rolled out in early 2026 and was onboarded like a person: it read the top five books on customer success, ten Atlan blog posts, the values document and org chart, and got a walkthrough of the CS team's tools. It follows the [same principle](https://www.engineermaxxing.com/micro/lessons/aa274a-01-perception-action-and-maps.html) as autonomous cars: sense the current state of accounts, plan next steps, act. Getting state right is crucial, and what Relay keeps is closer to long-term memory: who the stakeholders are and what they care about, what the customer is trying to achieve, the current picture in numbers and plain language, and a running list of risks and opportunities.

Example (Slack screenshot): a CSM asked Relay to review her three workstreams on an account. Relay agreed they were the right frame, then flagged what none covered: the executive sponsor had been disengaged for more than ten months (a high-severity risk), and the day's call had introduced a new voice on renewal.

- Inputs beyond calls: the initial bet was that call recordings carried most of what mattered. Some of the most important signals were in email and Slack that never came up on a call: a question sent straight to support, a buyer's conversation with marketing, a champion's 11pm Slack message. Adding them was one of the biggest quality jumps.
- Evals: early Relay gave wrong answers confidently, which is worse than not answering. It was a context problem, not an intelligence problem. A CSM built scenarios with defined right answers and treated Relay like a new intern: scenarios first, then real CS problems, reviewing and calling out misses.

### The promotion Relay read as good news

The lead stakeholder at one of Atlan's largest customers moved into a different role. Good news for him, so Relay logged a routine stakeholder update. It missed the question underneath: did the workstream Atlan depended on still have an owner? Nobody noticed until a colleague mentioned they weren't sure who was in charge. Relay had weighed two correct but contradictory signals, a departure and a promotion, and chose the more cheerful conclusion. Now a stakeholder role change gets flagged everywhere that matters, and Relay considers the downsides of good news. The pattern repeated every version: add context via skills, knowledge and expertise, test, take feedback. Eventually Relay got things right most of the time.

## The moment it clicked

When Atlan launched [Context Agents Studio](https://docs.atlan.com/product/capabilities/governance/context-agents-studio), roughly 200 customers were interested at once. Each conversation needed a deck built on that account's history and usage plus unstructured context (what they had used the product for, who cared, how it would help). Someone built a Relay skill that generated the deck: two hours became about 30 seconds. In Slack, a CSM asks for the deck and gets the link back in the same minute with key numbers pulled live from the warehouse rather than estimated (total assets, how many carry descriptions, accelerator results, hours reclaimed). An "Accelerator results" slide reports descriptions generated, assets with SQL intelligence generated, and hours saved, with estimated human effort per item; every figure is read from the customer's own usage.

That is when Relay went from something CSMs asked to something that did work for them. People then built their own skills unprompted: a weekly-update template, a way to match someone's writing style, a workflow for CS leaders drafting emails to account leaders. The real signal was not usage going up; it was people extending it themselves.

## What's different now

Relay runs across all customer accounts, from a quick email to a make-or-break presentation.

- A CSM working late on an account stuck for months thought out loud with Relay. It pointed out the account had stopped actively using Atlan without notice, a silent blocker. The resulting email finally got a response from the buyer; months later the account was running a proof of concept with Atlan on a new AI use case.
- A full renewal deck from scratch in under two hours before a call with a skeptical customer.
- The state of each account compiled before quarterly planning.
- Two days of scattered updates before a high-stakes demo (from a CSM, a solutions consultant, an implementation engineer and others) assembled by Relay.
- A renewal-critical briefing pulling two years of calls, support tickets, engineering requests, product usage metrics and competitive intel; Relay estimated eight to ten minutes and finished in about fourteen, while the CSM did something else.
- Connecting scattered signals: asked whether an AI agent build was underway at an account, Relay confirmed an active build (not a demo exercise), laid out three tracks with scope, co-builders and target dates, and explained why the first track mattered as a proxy for the broader initiative.
- A queue showing where each CSM's time is best spent: accounts with a new risk flag, drafts ready to review, relationships not touched in a while.

Relay is the CSMs' home base, one place for all customer information and tools. Efficiency and cost are the easy takeaway, but Relay matters because Atlan's CS team runs on "customer over company over team over self": every hour handed back goes to a CSM thinking about the customer instead of reconstructing the account.

The hard rule: even with Relay, a human owns the relationship, makes the judgment call on every risk flag Relay surfaces, and reviews anything before a customer sees it. The rules Relay follows are human-written: agents follow the guardrails, people write them. Machines are good at recall and repetition, but trust is still ours to build.

## What we've learned

- Accuracy and trust could never be granted upfront. Each earlier version failed at something specific, and each failure showed what the next needed.
- The most useful feedback came from skeptics, once asked what it would take to rely on Relay for real customer relationships.
- Not covered here, coming next: the skills, memory architecture and guardrails under Relay, as a technical piece for anyone who wants to build something similar. Also a separate story: Relay butting into human Slack conversations and learning when to speak up and when to stay quiet.
- Relay isn't finished; today's version will look outdated in a few months, as the multi-agent system did.

What's left for Customer Success is what was always the point: the customer. Atlan only exists if its customers succeed. "Every hour Relay takes off a CSM's plate is an hour a customer gets back, and that's the only measure that really matters."

## Author

Dhruv Saharya, Customer Success, Atlan. Dhruv runs customer success at Atlan and built Relay with the team that uses it. This piece carries the three versions that failed before it, the two times Relay got an account wrong, and what the skeptics asked for.

More essays: [Becoming Frontier](https://atlan.com/frontier/).