Build vs Buy an AI Sales Assistant? Why RevenueCat Built Its Own on Slack and n8n
By Nathan B., Senior GTM Operations / RevOps Leader at RevenueCat ·
Most teams reach for an off-the-shelf AI sales tool and move on. RevenueCat built its own. In this session, Nathan B. walked through a Slack-native AI sales assistant he built on n8n and OpenAI that pulls only from internal company data, prepares reps for calls in seconds, and is deliberately locked down so it can never pollute the CRM. This recap covers what the assistant does, how it is wired together, what a live meeting-prep query actually returns, and the build-vs-buy logic that justified a custom build.
What is the AI sales assistant?
It is a Slack-native bot that lives where reps already work. A rep messages the bot, and it pulls context from across RevenueCat's internal systems to answer sales questions and prepare for meetings. The core design constraint is that it draws from company data only, with no open web access, which keeps answers grounded in real account history rather than generic internet content.
The assistant is built on n8n for orchestration and powered by OpenAI for reasoning. Crucially, it is scoped to sales use cases. It will not run open web searches and will not answer off-topic queries, which keeps it from drifting into a general-purpose chatbot.
What data sources does it connect to?
The assistant reads from six internal systems, each chosen for a specific slice of account context.
| Source | What it provides |
|---|---|
| Salesforce | Accounts, contacts, pipeline, and activities (read-only) |
| Gong | Call summaries, routed via the Salesforce account and domain rather than a direct Gong integration |
| Slack | All accessible channels |
| HockeyStack | The prospect's website journey |
| App Figures | Mobile app download and revenue data |
| RevenueCat internal docs | In-call documentation lookup |
App Figures is the source that makes this build distinctive. Because RevenueCat operates in the mobile subscription space, knowing an app's download volume, revenue signals, ratings, and review counts is core qualification data, not a nice-to-have. That kind of mobile-app signal is exactly what generic sales tools do not surface.
How is it architected?
The flow is straightforward. An n8n webhook receives the Slack message and triggers an AI agent. The agent runs on tool-priority prompt logic, meaning the prompt tells the model which tools to consult and in what order rather than letting it guess. The agent then queries the connected systems, assembles the context, and returns an answer in Slack.
How was the workflow itself built and refined?
Nathan used Claude to audit and improve the n8n workflow via MCP. Rather than hand-tuning every node, he let Claude inspect the workflow and suggest improvements, which is a useful pattern for anyone iterating on a complex automation: the AI helps build the AI.
Why is it read-only?
Salesforce access is read-only by design. Write access was deliberately held back to avoid polluting CRM records. This is the kind of constraint that separates a production-grade internal tool from a demo: the assistant can read everything it needs to prep a rep, but it cannot silently create or overwrite data and degrade the source of truth everyone else depends on.
What does a live query actually return?
In the session, the assistant was run against a demo account to show end-to-end meeting prep. From a single request, it:
- Pulled the inbound context for the account, including a custom pricing request, a user base of roughly 500,000 users, and a planned paid launch in September
- Surfaced a Gong call summary from a 47-minute call on May 26
- Retrieved App Figures data for the app, including a 4.5 iOS rating, roughly 3,100 Android reviews, and that the app is currently free
- Fetched the relevant internal docs and drafted a follow-up email complete with a Calendly link
The takeaway is the compression. Work that would normally mean opening Salesforce, Gong, an app analytics tool, and a docs library in separate tabs collapses into one Slack message and one response.
Does it actually get used?
Adoption is honest, not hyped. Some reps use the assistant well and lean on it for real prep. Others treat it like generic ChatGPT and get less out of it. That gap is a recurring reality for internal AI tools: the technology working is only half the battle, and how reps frame their questions determines the value they extract. Scoping the bot to sales use cases helps, because it steers users back toward the queries it is actually good at.
The next planned step is a proactive push mode. Instead of waiting for a rep to ask, the assistant would surface context on its own, for example a nudge fifteen minutes before a meeting with the last call recap attached. That shifts it from a tool you remember to use into one that shows up when you need it.
When does building your own beat buying?
The build-vs-buy verdict was refreshingly unromantic. Off-the-shelf tools work well for standard stacks, and most teams should use them. The custom build was justified here by two things specific to RevenueCat: its mobile-app niche and the App Figures integration that no generic sales assistant provides. The lesson is not "build everything." It is that a custom build earns its keep precisely where your data and your niche fall outside what packaged tools cover.
Key Takeaways for GTM Teams
- Ground the assistant in your own data. Restricting the bot to internal systems with no open web access keeps answers tied to real account history instead of generic content.
- Make it read-only until you trust it. Holding back write access protects the CRM as the source of truth and removes the biggest risk of an autonomous agent.
- Use tool-priority prompting. Telling the agent which tools to consult, and in what order, beats letting the model improvise.
- Let AI build your AI. Auditing an n8n workflow with Claude via MCP is a fast way to iterate on complex automations.
- Build only where you are different. Buy for standard stacks; build where your niche and data, like mobile app analytics here, are not served by off-the-shelf tools.
- Adoption is a framing problem. The same tool delivers very different value depending on whether reps treat it as a focused assistant or a generic chatbot.
Want to Attend the Next Session?
The GTM Clawllective is a community of go-to-market practitioners building and sharing real AI workflows. Sessions are live, hands-on, and not recorded, so these recaps are the only written record of what was shown. If you want to see builds like this one as they happen and trade notes with people shipping them, join the community.
About the Speaker
Nathan B. is a Senior GTM Operations and RevOps leader at RevenueCat, where he works at the intersection of revenue operations, Salesforce architecture, and AI agents. He holds six Salesforce certifications spanning Sales Cloud, Platform App Builder, Marketing Cloud, and Pardot, and brings a product manager's instinct for clarity and focus to the systems he builds.
His work centers on turning go-to-market data into tools reps actually use, as shown by the Slack-native AI sales assistant he walked through in this session.