In healthcare, an answer you cannot verify is worse than no answer. Abridge built its product on that principle. Its platform turns a clinical conversation into documentation a clinician can trace back to what was actually said, auditable and checkable against the source. Today it powers more than 100 million medical conversations a year across health systems including Kaiser Permanente, Johns Hopkins Medicine, Duke Health, and Yale New Haven Health.
The same standard governs how Abridge uses AI internally. Leadership’s charter to Chris Humphries, a GTM Systems Engineer, was to build the most innovative and impactful go-to-market system he could, and to do it responsibly. Speed and quality, he says, are no longer two separate things.
Useful AI depends on access. A model is only as good as the data it can reach and the context it has to reason over. Abridge’s revenue picture lives in Salesforce, and a generic, off-the-shelf AI connection to it carried two risks at once. The answers could not be verified, and the access could not be seen.
“If I connect every one of our sales users to a generic server and say go for it, you get as many different answers as you have reps. We can’t afford that kind of confusion.”
For a company whose reputation rests on trustworthy AI, that risk was unacceptable. The job was to make go-to-market AI genuinely useful while keeping it governed, auditable, and trusted.
A governed layer for go-to-market AI
Abridge runs its go-to-market AI on the Tray AI Orchestration Platform. Chris works across several surfaces, including his AI coding tools, and Tray is the governed layer that ties them to the business.
Tray Agent Gateway gives the team governed access to Salesforce data, with full visibility into who is querying what. On top of that access, Chris layers curated context about how the organization actually operates. Salesforce holds a great deal of information that means little without that context, and a raw connection can confidently point someone in the wrong direction. The governed, context-aware version returns answers the revenue team can act on.
Responsibility led the rollout
Before broadening AI access for the sales team, Chris got the security foundation right. Using Tray’s governed connection to Salesforce, he spent days querying the org’s own data to find the gaps in profiles and permissions, then closed them, bringing a sprawling setup to a consistent posture built to scale. By hand he estimates it would have taken months. Working through Tray, it took weeks.
The discipline extends to the data itself. Abridge focuses on keeping its internal systems data clean, so its AI tooling has a strong foundation to work with.
The work that runs without him
Tray iPaaS also runs the work itself. When the native option could not sync conversation data the way the team needed, Chris built a custom integration that maps calls from their call recording vendor into Salesforce, then ran an 18-month historical backfill through the same build. Tray orchestrates agentic workflows too, on two fronts: the administrative and architectural work of managing Salesforce itself, and the business intelligence work of answering questions and taking action on the org’s data for the commercial teams.
What matters is the work that runs without him, while he is at his desk or away from it.
“All of these tools still rely on me typing into a computer. The real power is what can run without us, and Tray gives us that orchestration layer.”
What it adds up to
The leverage shows up in what a lean team can now ship. Governed AI runs work on its own, so a small go-to-market systems team supports a company of more than 500 without the headcount that would normally take. Tray, Chris says, makes 24 hours feel like 72.
None of this is reckless speed. Each expansion of what AI can touch followed the security and governance work that made it safe, the same order of operations Abridge applies to the AI it ships to clinicians. A healthcare company with an exacting bar for responsible AI trusts Tray to keep its go-to-market AI safe and governed. For any regulated business weighing how to adopt AI, that is the signal worth noticing.
The backbone
Chris calls Tray the backend of his go-to-market operation, and mission critical to it every day. The reason is structural.
An idea no longer stays an idea, and it no longer becomes an unsupervised script running where no one can see it. It becomes governed infrastructure. For a company building responsible AI for healthcare, that is the only version of fast worth having.
