Arriving at ‘Hello World’ in enterprise AI

Brendan Falk didn’t set out to become a cautionary tale. Three months after leaving AWS to build what he called an “AI-native Palantir,” he’s pivoting away from enterprise AI projects altogether. In a widely shared X thread, Falk offers some of the reasons: 18-month sales cycles, labyrinthine integrations, and post-sale maintenance that swallows margins. In other words, all the assembly required to make AI work in the enterprise, regardless of the pseudo instant gratification that consumer-level ChatGPT prompts may return.

Just ask Johnson & Johnson, which recently culled 900 generative AI pilots, keeping only the 10% to 15% that delivered real value (though it continues to struggle to anticipate which will yield fruit). Look to data from IBM Consulting, which says just 1% of companies manage to scale AI beyond proof of concept. Worried? Don’t be. After all, we’ve been here before. A decade ago, I wrote about how enterprises struggled to put “big data” to use effectively. Eventually, we got there, and it’s that “eventually” we need to keep in mind as we get caught up in the mania of believing that AI is changing everything everywhere all at once.

Falk’s three lessons

Though Falk has solid startup experience (he cofounded and ran Fig before its acquisition by Amazon), he was unprepared for the ugly stodginess of the enterprise. His findings:

Enterprise AI sells like middleware, not SaaS. You’re not dropping an API into Slack; you’re rewiring 20-year-old ERP systems. Procurement cycles are long and bespoke scoping kills product velocity. Then there’s the potential for things to go very wrong. “Small deals are just as much work as larger deals, but are just way less lucrative,” Falk says. Yep.

Systems integrators capture the upside. By the time Accenture or Deloitte finishes the rollout, your startup’s software is a rounding error on the services bill.

Maintenance is greater than innovation. Enterprises don’t want models that drift; they want uptime, and AI’s non-deterministic “feature” is very much a bug for the enterprise. “Enterprise processes have countless edge cases that are incredibly difficult to account for up front,” he says. Your best engineers end up writing compliance documentation instead of shipping features.

These aren’t new insights, per se, but they’re easy to forget in an era when every slide deck says “GPT-4o will change everything.” It will, but it currently can’t for most enterprises. Not in the “I vibe-coded a new app; let’s roll it into production” sort of way. That works on X, but not so much in serious enterprises.

Palantir’s “told-you-so” moment

Ted Mabrey, Palantir’s head of commercial, couldn’t resist dunking on Falk: “If you want to build the next Palantir, build on Palantir.” He’s not wrong. Palantir has productized the grunt work—data ontologies, security models, workflow plumbing—that startups discover the hard way.

Yet Mabrey’s smugness masks a bigger point: Enterprises don’t buy AI platforms; they buy outcomes. Palantir succeeds when it shows an oil company how to shave days off planning the site for a new well, or helps a defense ministry fuse sensor data into targeting decisions. The platform is invisible.

Developers, not boardrooms, will mainstream AI

In prior InfoWorld columns, I’ve argued that technology adoption starts with “bottom-up” developer enthusiasm and then bubbles upward. Kubernetes, MongoDB, even AWS followed that path. Falk’s experience proves that the opposite route—“top-down AI transformation” pitched to CIOs—remains a quagmire.

The practical route looks like this:

Start with a narrow, high-value workflow. Johnson & Johnson’s “Rep Copilot” is a sales assistant not a moon shot. A narrow scope makes ROI obvious.

Ship fast, measure faster. Enterprises are comfortable killing projects that don’t move KPIs. Make it easy for them.

Expose an API, earn love. Developers don’t read Gartner reports; they copy code from GitHub. Give them something to build with and they’ll drag procurement along later.

What’s next

Falk says his team will now “get into the arena” by launching products with shorter feedback loops. That’s good. Build for developers, price like a utility, and let usage (not enterprise promises) guide the road map. The big money will come from the Fortune 500 eventually, but only after thousands of engineers have already smuggled your API through the firewall.

Enterprise AI transformation isn’t dead; it’s just repeating history. When visionary decks meet ossified org charts, physics wins. The lesson is to respect that and abstract away the integration sludge, price for experimentation, and, above all, court the builders who actually make new tech stick.

Falk’s pivot is a reminder that the fastest way into the enterprise is often through the side door of developer adoption, not the fancy lobby of the boardroom.