The fatal mistake behind 80 percent of AI failures—and the 5-step playbook that prevents it

Adopting the process behind SpaceX’s decision making can expose costly flaws most leaders never see until it’s too late.

Last month, Andreessen Horowitz published a sweeping profile of SpaceX, detailing the five-step operational process Elon Musk has embedded into every program the company runs. SpaceX calls this process “The Algorithm.”

The Algorithm runs in exact sequence:

1. Question every requirement 
2. Delete any unnecessary part or process
3. Simplify and optimize
4. Accelerate cycle time
5. Then automate

Tim Berry, who spent a decade leading Falcon 9 and Falcon Heavy upper-stage production, described the process as “drilled into our minds.”

After 10 years, the Raptor 3 engine is the physical proof of the algorithm’s utility: The engine produces 22 percent more thrust than its predecessor, according to Andreessen Horowitz, and requires no heat shield because its plumbing has been 3-D-printed directly into the engine’s metal structure.

What happens when the algorithm is not followed? According to reporting by CNBC, Musk admits that “excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated.”

Andreessen Horowitz notes that the problem wasn’t just excessive automation but automation out of order: “The mistake at Tesla’s factories was that automation was attempted first, before requirements had been questioned and processes deleted.”

The Fatal Flaw in Most AI Projects 

Research from MIT’s Project Nanda found that 95 percent of organizations deploying generative AI saw zero measurable return, and the failure was almost never the model. 

Why? 

Most companies are funding step five without completing steps one through four. The processes they’re automating were never questioned, never simplified, and in many cases, never should have existed at all. 

According to S&P Global’s 2025 Voice of the Enterprise survey, 42 percent of companies abandoned most of their AI initiatives, up from 17 percent the year before. Rand Corporation estimates that more than 80 percent of AI projects fail to deliver their intended business value, roughly twice the rate of comparable technology projects without AI. Every major research organization studying this pattern identifies the same culprit: companies deploying AI on processes they’ve never questioned or simplified. 

What Getting It Right Looks Like 

Amazon CEO Andy Jassy’s 2025 shareholder letter described how six engineers rebuilt the entire Bedrock inference engine in 76 days using Kiro, Amazon’s agentic coding service. The project was originally estimated at 40 engineers and a full year. The resulting engine, Mantle, became the backbone of Amazon Bedrock’s rapid scaling. The compression came from a team with the judgment to rebuild from first principles—the kind of judgment that only comes from having rebuilt systems before. 

Here are Algorithm’s steps slightly reframed, so any leader can apply them to their AI (or other innovation) efforts: 

1. Question every requirement. Every requirement needs a named owner—a specific person, not a department. Smart people’s requirements are the most dangerous ones, because no one thinks to challenge them. Make it someone’s job to question even those, and to keep pushing until the list of requirements is as lean as possible. 
2. Delete every part or process you can. Cut aggressively. If you don’t end up adding back at least 10 percent of what you removed later, you didn’t cut enough. 
3. Simplify and optimize. Do this only after step two. A common error is to streamline something that should have been eliminated entirely. 
4. Accelerate cycle time. Speed up every process, but only after completing the first three steps. Accelerating a flawed process just produces faster failure. 
5. Automate. This step comes last, not first. At Tesla’s factories, Musk made the mistake of automating before the earlier steps were complete, locking in a flawed process at machine speed. Automation amplifies whatever you give it, so the process needs to be clean before scaling with technology. 

Underneath the SpaceX story is a narrative about the limits of tools without judgment. AI excels at executing what’s been specified. The judgment to know what should have been deleted before automation began can only come from people. 

This Week 

Before your next AI deployment decision, confirm the processes you’re about to automate have actually been questioned, with steps or features deleted where possible and simplified to their absolute minimum. Then pick one workflow and commit to running steps one through four before anyone touches a tool focused on scaling. 

The sequence is the strategy. You can’t automate the work of defining what really matters. 

Soren Kaplan helps leaders leapfrog to what’s next by cutting through noise, aligning faster, and making smarter decisions. He speaks at corporate events and conferences worldwide. Learn more at sorenkaplan.com. Go even deeper on this topic with Soren’s podcast on Apple, YouTube, or Spotify. 

This post originally appeared at inc.com.

“Click here to subscribe to the Inc. newsletter: inc.com/newsletters"