Merge co-founder Gil Feig says AI token bills are brutal for CFOs, so Merge is investing heavily in cost savings through its Gateway product.
Merge's Shensi Ding says a large deal that would have consumed about 90% of resources paused for a month, pushing the team to adopt AI coding tools like Claude Code and Windsurf.
Gil Feig predicts many AI companies will fail or be unable to raise their next round, while Merge is glad it avoided inflated valuations.
Merge bets agents need security and cost controls
Merge is moving from integration plumbing to infrastructure for AI agents. Agents multiply output but also risk, because they write more code than humans can review and can leak data once connected to tools. Merge's bet is that guardrails, governance and cost routing become the layer companies need before they adopt agents at scale.
Takeaways
A paused deal pushed Merge into AI. A deal that would have consumed 90% of resources paused for a month. The team used the pause to learn AI coding and finalize Agent Handler, which gives agents hundreds of integrations through one MCP server.
Agent-written code makes security the central risk. Code change requests to GitHub are soaring, so humans cannot review the agent-written code, and supply chain attacks spread through shared open source packages. Feig says Merge answers with hard guardrails that block sensitive data leaving through agents and jailbreak detection.
Attackers now have AI leverage. Feig says AI gives attackers fluent English and good code, with unlimited manpower. Over a thousand bot sign-ups in about an hour scanned Merge's endpoints before bot detection blocked them. He thinks a capable model finding twenty or thirty year old bugs is realistic.
Token costs are now a CFO problem. Feig says AI token bills reaching CFOs are brutal, and Merge has invested heavily in savings through Gateway, which routes between models under security and cost policies. Ding notes a growing margin focus and cites Cursor's margin improvements.
Revenue accelerated while headcount barely grew. Ding says headcount did not grow much last year while revenue accelerated significantly, and AI use counts toward performance. Feig cites Keith Rabois's barrels and ammunition, arguing AI lets one person build so Merge needs more barrels.
Agents will choose vendors with headless access. Feig says a headless approach lets agents work through APIs or a CLI instead of a user interface, and Salesforce's open API has been a moat for twenty years. Ding predicts vendors will need easy sign-up and payment because agents will choose them.
AI buyers move fast but want control. Merge has over 20,000 self-serve organizations and over 400 enterprise customers. Large AI companies buy much faster than financial services firms, though many customers still want AI running in their own infrastructure.
What the founders expect next
Feig predicts companies will start using agents this year.
Feig expects many AI companies to fail or be unable to raise their next round, forcing layoffs and slowdowns. Merge is glad it avoided inflated rounds.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Feig argues English instructions will make agents repeatable.
Ding argues companies overengineer machine learning by training custom models when generic models would be enough.