Deploying Agents at Scale: What Enterprise Adoption Actually Looks Like

I spend my days inside enterprise AI rollouts — standing up the platforms, watching what agents do at scale, and cleaning up after them. This talk covers where the deployments and the research line up, and where they contradict each other.

The evidence doesn't agree with itself. METR's 2025 randomized controlled trial found experienced developers 19% slower with AI while they believed they were 20% faster — and its 2026 follow-up couldn't measure the gap cleanly, because heavy AI users now refuse to work without it. Google's DORA study of nearly 5,000 practitioners says strong engineering foundations protect you. Faros AI's telemetry across 22,000 developers found the opposite: throughput up 66%, incidents per pull request up 243%, in mature organizations that should have been safe. I'll show you which read holds up.

Then the enterprise layer — JPMorgan's 200,000-employee rollout, and the MIT-versus-Wharton disagreement over whether any of this returns money. The organizations pulling ahead didn't buy a better model. They redesigned the work around the tool, added guardrails, and measured against a real baseline before scaling. AI amplifies whatever system it lands in. Point it at a mess and it produces the mess faster.

You will leave with a clear understanding of what levers increase enterprise AI success, and what levers cause it to fail.