Sonora's Customer Knowledge Graph: Architecture for the AI Era

Most CS platforms are databases with a reporting layer on top. You connect your customer data, they store the records, they let you build dashboards. The “AI” is usually a sentiment classifier bolted onto support tickets, or a health-score formula you configure in a UI. The data sits. Your team interprets it.

That’s not what we built.

Sonora is a single, unified graph of everything you know about a customer. Every account, contact, ticket, meeting, and feature request lives as a connected point in that graph, and AI-generated signals (a churn risk, an expansion opportunity, a sentiment read) attach directly to those points. The key design choice: those signals aren’t hardcoded. You define new ones in configuration, not in code.

That matters more than it sounds. In a typical system, adding a new signal — say “expansion readiness” — is an engineering project: define a field, write a classifier, wire it in, ship it. In Sonora, expansion readiness is just a new entry in the config, pointed at data you already have. It produces a scored, explainable output, with no deploy, no engineer, and no waiting on a roadmap.

Getting your data in cleanly

There’s a problem nobody mentions in the demos: getting data in cleanly is harder than it looks. Complex data models, custom objects, non-standard fields — most platforms handle the straightforward stuff and quietly struggle with the rest. Implementation drags. Data-quality issues pile up. By the time health scores are running, they’re already built on an incomplete picture.

It’s worth being clear about why this is so hard to fix. For most platforms it isn’t a tuning problem. It’s foundational. The data model was built to expect data in a certain shape, and once a system is committed to that shape, reworking it to handle the messy reality of a real customer’s stack means rebuilding the foundation. So it doesn’t get fixed; it gets worked around, and the incomplete picture becomes permanent.

This is exactly where the graph earns its keep. Because the structure is flexible by design, the custom objects and one-off fields that stall other platforms land cleanly. They become their own points in the graph instead of getting forced into a rigid template or dropped on the floor. The hard part of every CS implementation is the part we built the system to absorb.

That’s also why Sonora integrates with complex, customized systems in days, not months. There’s no foundation to bend the data to; the graph just takes the data as it actually is. The implementation that stalls other vendors for a quarter is the part we move through fastest.

What happens after the data lands

Before Sonora scores anything, it reads the full history of the account across every connected source: Salesforce, Zendesk, Intercom, Slack, Gong, email, and more. This isn’t “feed one support ticket into an LLM.” It’s closer to having an analyst who has actually read the last six support tickets, Tuesday’s Gong call, the 90-day usage trend, the open feature requests, and the account’s exec history before forming a view.

The difference shows up in three ways:

• The reasoning is grounded in your real records, not a summary of them
• Every score traces back to the specific inputs behind it
• It gets sharper as more history accumulates

Signal from outside your four walls

There’s one category of signal no CS platform touches: what’s happening outside the account. A funding round. An exec change. A competitor moving into your customer’s space. That context lives on the web, not in your CRM, and for most teams it never makes it into a health score at all. Someone reads it in the news, mentions it in a QBR, and it evaporates.

In Sonora, that external signal flows into the same graph as everything else. So a risk score isn’t just ticket volume and Slack tone. It can also weigh the layoff announcement that landed yesterday or the new VP who just replaced your champion. The system reasons over what’s happening inside and outside the account at once, because to the graph they’re not different kinds of data.

One picture, not four dashboards

A do-it-yourself stack gives you one tool per source. Gainsight scores your Salesforce data. Gong reads call sentiment. Zendesk trends your ticket volume. None of them talk to each other, so you’re left stitching it together by hand, or pasting exports into Claude and hoping.

The more modern version of build-vs-buy runs into the same wall. You can connect a model to your sources directly — through MCP, a custom agent, or your own prompts — and it works impressively for a single account. But it answers one query at a time. Ask about one account and you get a sharp answer; there’s no practical way to run that same reasoning across your entire book, see the results side by side, and act on them. Analyzing one account is a demo. Visualizing and acting across thousands at once is the actual job, and that’s where the DIY approach hits its ceiling.

The advantage isn’t the AI; it’s that everything feeds one graph underneath it. Sonora runs an agent per account, so every account gets its own continuous, full-history analysis, rather than one query you point at the book a slice at a time. So you can see your whole book at once and act on it, with each account already reasoned through individually. And on any single account, Sonora sees the full picture before it forms a view: escalation risk isn’t ticket volume alone. It’s ticket velocity and Slack sentiment and exec-level email going quiet and a usage drop, read together. Not four signals on four screens. One read, four inputs, across every account at once.

You can see why, and you can change it

Every score is stored with the exact setup that produced it: the thresholds, the definitions, the model behind it. When you want to change how health scoring works, you change the setup and re-run. The old scores don’t disappear; they’re kept, so you can compare the new approach against the old one on the same accounts.

This solves a real operational problem. Most in-house health scores turn into tribal knowledge within 6–12 months. The person who built the formula left, the thresholds got tweaked informally, and now nobody can say why account X is green when it feels red. Sonora keeps the methodology as something you can actually inspect, not institutional memory that walks out the door.

The data compounds

Every interaction you bring in — every call, ticket, email, Slack message, and feature request — becomes a scored, explainable data point tied to the right account and contact. At month one, that’s useful. At month twelve, it’s a complete record of your book of business that knows not just what happened, but what it meant.

We built Sonora this way on purpose. Customer success doesn’t sit still: your product ships new features, your business moves into new segments, your customers’ priorities change underneath you. A system with a fixed data model and hardcoded signals starts drifting out of date the moment any of that happens. Sonora was designed to move with it. New signals are configuration, not engineering. New sources join the same graph. The analysis re-runs as the picture changes.

The goal wasn’t a platform that demos well today. It was one that still fits your business a year from now, and is sharper for having been there the whole time.

See it in action →