The bitter lesson of workplace AI

Simple, worker-driven AI adoption beats elaborate, IT-engineered solutions. Every time. And the endgame is more radical than "stop engineering": AI can route around large parts of the cognitive scaffolding humans built over centuries to get real output. But that's not the same as saying the scaffolding stops mattering, and it's not a claim about today—see the corrections below.

Published: September 17, 2025—Original post

The original bitter lesson (Rich Sutton)

Throughout AI research history, researchers kept learning the same painful truth: simple methods that leverage computation always beat sophisticated algorithms designed by clever humans. Chess engines that searched more positions beat ones with hand-crafted strategies. Neural networks that processed more data beat systems with carefully engineered features. The lesson: stop trying to be clever. Scale and simplicity win.

The workplace version

The same pattern is playing out in enterprise AI right now:

• Companies pour millions into custom models, complex integrations, and elaborate governance frameworks
• Workers get better results with a $20 ChatGPT subscription
• IT's response: "That's just shadow AI. Our real solution will replace it."
• The workers have it right every time

This isn't workers being lazy or IT being incompetent. It's a fundamental truth about how technology spreads. The simplest path to value wins.

The prescription: Stop engineering a better AI tool. Start enabling the AI tools workers already found. Apply governance at the workspace layer, not the tool layer.

The radical extension: AI erodes the invisible 80%, it doesn't dissolve it

The invisible 80% framework established that 80% of knowledge work is invisible—judgment, tacit expertise, pattern recognition, reasoning. Enterprise AI can only see the 20% (outputs, documents, observable actions) and fails because it can't reach the rest.

The current position, corrected three times since the original argument (twice on the conclusion below, once on the timing—see the third correction): AI doesn't need to replicate human cognitive processes to deliver real value—it can and does route around large parts of the scaffolding humans built over centuries to get from inputs to outputs. But the invisible 80% doesn't dissolve on its own, and it doesn't shrink to zero even as AI erodes it. Whatever remains invisible still has to be deliberately captured—which is exactly the job the knowledge factory does.

Where the original version overshot, and how it got corrected:

The first draft of this framework claimed AI simply doesn't need the 80% at all—that the invisible cognitive scaffolding was purely human infrastructure AI could route around entirely, arriving at outputs without ever needing to capture what humans do beneath the surface. Think of it this way: if you're getting SVP-level strategic support from AI, how much of the invisible work humans do beneath that level—the years of pattern recognition, the accumulated judgment, the institutional memory—did the AI need? It didn't build those layers. It just arrived at the output. That observation is still true. Two corrections landed on the conclusion drawn from it, from different directions:

AI still needs the 80%, captured deliberately, not routed around. Skills are all you need is an entire argument for capturing tacit knowledge as text files. The crawl/walk/run post argues agents fail precisely because nobody did that capture work—the opposite of AI routing around it. And the knowledge factory states outright that it revises this framework: the shared departmental second brain is an engineered apparatus for capturing exactly the tacit knowledge this section originally said not to bother capturing.

2026-08-28—the "dissolves" language implied an endpoint that doesn't exist. It suggested a clean shift from 20% visible / 80% invisible to eventually 100% visible / 0% invisible. That's the wrong shape. The boundary between visible and invisible moves as AI erodes into the 80%—the split might become 60/40, or 40/60, whatever the real number turns out to be for a given kind of work—but there's no version of this where it hits zero. Something always stays invisible, and that remainder still needs the same deliberate capture work, permanently, not as a transitional phase before AI eventually absorbs everything.

The connection to factory electrification

This is the Phase 2 → Phase 4 jump from factory electrification. Phase 2 companies are engineering AI into existing human workflows (new tech, old processes). The bitter lesson says: stop. You're engineering around human constraints that AI doesn't share. Phase 4 is redesigning work around what AI makes possible—which means a lot of the invisible cognitive infrastructure humans needed simply isn't part of the new design.

The third correction: it's a sequencing claim, not a standing one

2026-09-04. The cleanest way to reconcile this framework with the knowledge factory is to stop reading the bitter lesson as a claim about today and start reading it as a claim about later. Both are true, in order:

Now: if you want AI inside the systems and processes you actually run, you have to build the factory. Forward-deployed engineers, deliberate capture of the tacit knowledge, curated canon, governance, the whole apparatus. There is no shortcut where scale alone gets you there, and the crawl/walk/run evidence is that skipping the capture is precisely why agent deployments disappoint.

Later: once the factory exists and progressively automates itself, the system starts cutting pieces out. The scaffolding that had to be hand-built to get the thing running turns out not to be needed to keep it running, and the parts that were engineered around human constraints get dropped first. That's the bitter lesson arriving—not as a reason to skip the build, but as what happens to the build afterward.

Read that way, the original claim wasn't wrong so much as early. The mistake was stating a long-horizon endpoint as present-tense advice, which turned a real prediction about where engineered scaffolding ends up into a reason not to do the work that gets you there.

Using this framework

Deploy when:

• An enterprise is building elaborate AI integrations that workers ignore in favor of simpler tools
• IT is engineering a better AI tool when workers have already found one—the bitter lesson is about the tooling, not about the knowledge
• Someone argues "AI can't do X because it lacks human judgment/experience/intuition"—the bitter lesson says AI might not need to replicate the human path to achieve the human outcome
• A company is in Phase 2 of the factory electrification pattern and needs to be pushed toward Phase 4 thinking
• The debate is about which AI tool to standardize on—the answer is: enable the ones workers already chose, govern the environment
• Someone assumes the knowledge factory's scaffolding is permanent—the long-horizon read above says the factory automates and thins over time, even though it has to be built first

Do not deploy it as an argument against capturing institutional knowledge. That was the original version's deployment advice and it's been retired: the capture work is the prerequisite, and the bitter lesson only applies to what happens to that apparatus after it exists.

The prescription for the tooling is unchanged: stop engineering, start enabling. The prescription for the knowledge is the opposite: enable the pioneers first, then industrialize the capture. And the philosophical anchor holds either way: don't assume AI needs to work the way humans work to achieve what humans achieve.