I build AI systems and run the adoption loop.
Agents in production. Humans in the loop. Every output evaluation-gated.
Ten years of getting organizations to actually use what they buy, now applied to AI: autonomous marketing-ops agents running on live businesses, with review queues, eval gates, and cost instrumentation.
I'm Jason Leinart. I build AI systems for real businesses and stay until people actually use them: agents that draft the work, humans who approve it, and evaluation gates that keep the output honest.
The receipts: an LLM reporting pipeline gated by automated graders, live for paying clients. Unattended agents with audit trails. A platform running across three production sites. Behind that, a decade of enterprise integration, including six systems across a 44-store PE-backed portfolio. Oxford-certified in AI governance, NIST-aligned, and fluent from the C-suite to the front line.
In Motion
- Agentic operations: Standing agents on live client businesses — scheduled runs, review-ticket approvals, output that ships only after it clears an evaluation gate.
- Evals and economics: Rejected outputs become golden regression sets; every agent run captures token usage so cost per completed task is a number, not a guess.
- Enablement: Training non-technical operators to run their own AI review queues, with prompt libraries and SOPs that outlive the engagement.
- Governance: NIST-aligned risk classification for every deployed agent. Oxford AI Governance certified.
- Speaking: "I Know Kung Fu" — PMI Great Lakes Spring Symposium 2026, on modeling AI adoption (delivered April 2026).
Previously: a decade leading marketing and operations technology across multi-location portfolios — the pipelines, integrations, and adoption work that make a rolled-up operation run as one business.
Recent Analysis
View allThe Deployment Gap: What Peer-Reviewed Research Reveals About Healthcare AI's Readiness Problem
A systematic analysis of peer-reviewed healthcare AI papers reveals that deployment failures are organizational readiness problems, not technology problems. Evidence-based framework for healthcare AI leaders.
Enterprise AI Transformation
Why 61% of AI initiatives fail to deliver EBIT impact—and the 10-20-70 framework that separates high performers from the rest.
Recent Notes
View allWhat 510 Contracts Taught Me About Training Data
My first ML model predicted the same 9 labels for every input. The fix wasn't a better model — it was a better data pipeline. The difference between 510 training examples and 15,700.
When Consistency Beats Intelligence
A 70,000-person field experiment proved AI doesn't need to be smarter than humans to outperform them. It just needs to be more consistent. Better inputs, not better decisions.
Stack & Methods
Claude & Claude Code • agentic workflows with human-in-the-loop • evaluation harnesses and golden regression sets • MCP (consumer and server) • prompt libraries and enablement SOPs • workflow orchestration • GA4 • Search Console • Google Ads • Cloudflare Workers & D1 • Astro • Sanity • Stripe • Brevo • Python & TypeScript • NIST AI RMF • OWASP LLM Top 10
Problems I Think About
Why AI pilots die before production • What LLM output needs before a business can trust it: evals, gates, audit trails • Token economics as unit economics — the cost of a completed task, not a token • Why adoption is a training problem before it is a technology problem • How far human review can safely recede as evidence accumulates.
Hiring for AI enablement or delivery?
I'm looking for the seat where this work scales: AI enablement, adoption, and forward-deployed delivery roles. Everything on this site is a live system running on a real business, and I'm glad to walk through any of it.