Vellum sells workflow automation to engineering teams — meaning their customers are engineers, their bugs are intricate, and their support load is technical. They embedded AIVA directly into their product. Six months later, the entire on-call rotation got shut down.
We embedded AIVA inside the workflow editor itself. When an engineer hits a wall, AIVA is right there — with access to their actual workflow, their actual error, their actual logs.
The challenge
Vellum's customers ship production workflows that touch payment APIs, internal systems, and live data pipes. When something breaks, they're not asking where the unsubscribe button is — they're pasting stack traces, JSON payloads, and asking why a webhook fired twice.
Generic chat widgets are useless here. By the time a customer has explained the context, a senior engineer has already lost 20 minutes. And every reply requires reading their workflow first.
Until mid-2025, Vellum had four engineers on a rotating on-call schedule, splitting the support inbox alongside their day job. Average first response: 47 minutes. Worst-case: 14 hours overnight. And growing.
4 engineers on rotation, each losing ~6 hours/week to support triage.
Tickets were technically rich — workflow JSON, error traces, account state — but had to be reconstructed by hand every time.
47-minute median first response. Customers either gave up or posted in the public Slack.
Generic chatbots couldn't see the workflow — they'd just hand off immediately, defeating the point.
The setup
The Vellum team didn't want a chat bubble in the corner. They wanted AIVA to live inside the workflow editor — same surface the customer was already looking at, with full access to the workflow definition and recent execution logs.
AIVA's in-product embed made this trivial. Their senior engineer wrote a 40-line integration that (a) embeds the AIVA pane next to the editor, (b) passes the active workflow JSON as context, and (c) wires three custom tools: fetchWorkflowLogs,retryExecution, and escalateToHuman.
From clone to demo: one afternoon. They piloted internally for two days, then rolled to 5% of customers, then 100% within a week. The team picked the model and escalation rules; AIVA handled the rest.
The outcomes
AIVA pays for itself before lunch. Our engineers got their week back. Our customers got answers faster than they expected. The only complaint we've had is that people now expect this from every other tool they use.
— Priya Subramanian, Head of Product, Vellum