Snyk ran its support agent internally for a year before customers saw it. Over 85% of sessions now end without a ticket

The fastest way to a customer facing AI agent, judging by Snyk's numbers, was to make its own support team the first users.

What happened

Snyk, a developer security company, published a case study on LangChain's blog about Snyk Assist. Its support team handles thousands of cases every couple of weeks, each one read, classified and routed by hand. For roughly a year, the agent ran only internally, as a case triage tool and virtual agent for support staff. It launched in the customer support portal in April 2026 and moved into the core product on September 1. Snyk's docs currently list it as Early Access on the Enterprise plan.

The design choices are the useful part. One LangGraph agent sits behind a Slack app, a web app and an API. Tools are attached at request time based on the signed in user's permissions, so the agent can only reach data that user could already see. Conversation history lives in PostgreSQL. Every pull request runs the real agent against test suites, including real questions with known good answers and red team prompts, and blocks if scores fall below agreed thresholds. A scheduled job grades production runs on two questions: was it about Snyk, and did it answer? That gives a measured deflection rate instead of a guess.

Snyk reports more than 60,000 customer queries handled, over 85% of sessions resolved without a support ticket, and 250+ cases auto detected and escalated to the right team. This is a vendor case study, so treat the numbers as Snyk's own.

My take

This is the order I would follow for any service business:

1. Point the agent at your own team first. Mistakes stay inside, and staff feedback arrives in hours.
2. Scope tools to the person asking. If a client could not see a record in the portal, the agent should not either.
3. Keep one agent and change only the surface: internal Slack first, then the help widget.
4. Write the test set before launch, from real tickets with known answers.
5. Grade live conversations, so "deflection" is a number you measured.

For a team on HubSpot or GoHighLevel, step one can be modest: an agent that reads new tickets, tags them and drafts a reply for a human to send.