Workflow-First Vertical AI Adoption

Diving deeper into

Levi Lian, CEO of Raycaster, on why vertical AI is workflows first & chat last

Interview
less than 1% of manual workflows have been replaced with AI
Analyzed 8 sources

The real bottleneck in regulated industries is not model intelligence, it is trust at the exact step where a person must change an important document. In biotech, legal, and finance, most work still lives in spreadsheets, quality records, regulatory packets, and internal source files that need traceable edits, approvals, and system integrations. That is why AI adoption stays tiny until the product can slot into the workflow itself, not just answer questions in a chat box.

  • Raycaster is built around document heavy biotech work like tech transfer packets, CMC authoring, QA processes, and CRO or CDMO RFPs. It plugs into systems like Salesforce, Google Drive, Veeva, IQVIA, SharePoint, and LIMS, then captures document diffs, user corrections, and tool calls so the product improves inside the customer workflow.
  • The pattern matches other regulated vertical AI winners. Harvey grew by packaging model orchestration with legal workflows and high touch implementation, including customer success staffed with ex lawyers. Hebbia spread in finance by indexing sensitive internal documents and partnering with firms like FactSet and Fitch to become part of live investment work, not a standalone research toy.
  • Benchling shows what broad replacement looks like once a workflow system becomes the system of record. It reached about $210M ARR and 1,200 customers by replacing paper notebooks, email, and Excel for experiment design, documentation, and sample tracking. AI workflow companies are earlier in that same migration, moving one regulated task at a time from manual review into software.

The next phase is moving from copilots into audited systems that draft, route, compare, and log every change across high stakes workflows. As more vertical AI products earn trust at the document layer, adoption should spread from narrow pilots into the operating core of firms, the same way systems like Benchling became standard infrastructure once they proved they could hold the record of work.