Wade Foster paused every project at Zapier, declared a "Code Red," and put the entire company, not just engineers, into an AI hackathon. Within one week, company-wide AI usage jumped from under 10 percent to over 50 percent (Zapier). That single data point should change how founders think about rolling out AI internally. The bottleneck isn't a training deck. It's a forcing function.
The compression is the point
Zapier's later numbers show why the hackathon format compounds rather than fades. Daily AI adoption climbed from 63 percent in late 2023 to 77 percent by the end of 2024, then to 97 percent by the time the company published its retrospective. Zapier credits hackathons and recurring internal builder sessions, not top-down mandates, as the most effective driver (Zapier). The mechanism is straightforward: a hackathon has a deadline, a demo, and public stakes. A training module has none of those, so it gets deferred.
Scale doesn't dilute this. Tata Consultancy Services ran what it calls the world's largest AI hackathon, the tcsAI Hackathon 2025: more than 281,000 employees across 58 countries, working on 21 AI themes. Twenty-nine percent of participants had no technical background (TCS Newsroom). At the other end of the size spectrum, STIM, the Swedish music-licensing organization representing more than 109,000 rightsholders, took 40 to 50 employees with zero AI background and had them build three working AI systems, including an internal FAQ bot and a sentiment-analysis tool for customer emails, inside a 48-hour hackathon run with Devoteam (Devoteam). Same format, opposite ends of headcount. Both moved non-technical staff from spectators to builders in under three days.
Non-technical staff need something to build, not something to watch
The pattern across every example here is that non-technical participation rises when the format requires output, not attendance. Canva's second company-wide "AI Discovery Week" pulled in more than 5,300 employees across 64 sessions, logging 25,940 hours of AI learning, but the number that mattered was what came after the learning: a two-day hackathon that generated 467 registered ideas, up from 330 the year before (Canva Newsroom). Learning hours are a vanity metric until they funnel into a deadline with a deliverable.
IBM and Microsoft's third annual agentic AI hackathon followed the same logic at enterprise-client scale: more than 900 participants across 225 teams from clients in 25 countries, running over an eight-week structure, submitted more than 200 prototype solutions for judging (IBM Think). Eight weeks is longer than a three-day sprint, but the design principle holds: a judged deliverable beats an open-ended mandate to explore AI tools on your own time.
What Meta's pushback exposes
The format has a failure mode, and it shows up when the deadline gets stacked on top of an already full workload instead of replacing it. Meta scheduled a company-wide, three-day AI hackathon for July 14 to 16, 2026, and the plan drew public pushback from staff who said they were already stretched too thin to step away for it (HR Grapevine). Zapier's version worked because Foster paused everything else first. A hackathon bolted onto normal workload reads as one more task on the pile. A hackathon that replaces normal workload for three days reads as a mandate worth taking seriously.
That's the operating detail founders skip. The hackathon isn't the intervention. The pause is. Without a real pause, the non-technical staff you most need to reach, the ones already at capacity, will treat the event as optional and skip it, which is exactly what undercuts the 29 percent non-technical participation TCS achieved and the 40 to 50 non-AI-background employees STIM converted into builders.
What to do this week
If you run a company with more than 20 people and less than universal AI usage, don't schedule a lunch-and-learn. Pick a three-day window, tell every function it's a full stop on normal work, pair non-technical staff with whoever on the team is already fluent, and require a working demo by day three, not a slide deck. Judge it publicly. The organizations that ran this experiment, from a 40-person rights body to a 281,000-person services firm, report the same result: adoption moves when people build something real under a deadline, not when they're handed another tool to explore on their own time.
Zapier's jump from under 10 percent to over 50 percent adoption happened in one week, the week normal work stopped, not the week training started.
Paul Worrell is a contributor to The Founders Report. He runs Rvysion, a design and growth agency for startups.
Disclosure: The Founders Report's editor runs BrainVaultAI, referenced in this piece.