Share from 0:00
0:00
/

Tesla Automation Mistake: Are You About to Repeat It With AI?

Innovation Unpacked: before you put AI on a process, ask if the process should exist.

Note: links to the free checklist and playbook are at the end


Tesla Pulled Robots Off Its Model 3 Line. Musk Called Too Much Automation His Mistake.

In April 2018, Elon Musk posted this about his own company: “Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated.”

Tesla had pulled robots off its Model 3 line and brought people back in to do the work.

Tesla had set out to build a line run almost all by machines. It hit trouble fast. Musk told CBS the plant had a complex network of conveyor belts that wasn’t working, so they got rid of the whole thing. On an earnings call, he talked about a robot built to place fiberglass mats on the battery pack. He called it the “fluffer bot.” People did that job better.

To be fair to Tesla, most of the line stayed automated. Musk said so on the same call. So this isn’t a story about robots being bad. It’s a story about the order you do things in.

Here’s my thesis. Tesla automated steps before it knew which steps should exist. Most companies are about to do the same thing with AI. They’ll take a slow, messy process and make it run faster. Then they’ll pay to keep a faster version of the waste.

The same mistake, with AI

Here’s the pattern I see. A company has a slow, messy process. Someone buys an AI tool and puts it on top. Now the messy process runs faster. It also costs more, breaks in new ways and is harder to change, because now it’s built into software.

Here’s an example. Take a sales team that isn’t closing big deals. Leadership buys an AI dashboard that scores leads. Six months later, nothing has changed.

Why? The reps already knew where the good leads were. Their pay plan punished them for chasing big, slow deals. The problem was the pay plan, not a lack of data. The dashboard solved a problem nobody had.

You can find this in most companies. A chatbot sits in front of a support process that already sends people in circles. An AI writer gets added to a review cycle with six sign-offs. The tool works fine. The process around it is the problem.

That’s what happens when you automate first and think second. You lock in the workaround. Every step you automate is a step you now have to maintain, test and fix. If the step shouldn’t exist, you’re paying for it twice.

Why the usual approach fails

Most AI projects start with the tool. A vendor shows a demo. A team picks a process that looks slow. Someone asks how much faster the tool can make it.

That question skips everything that matters. It assumes the steps are right and only the speed is wrong. It never asks who wanted each step, or why. It never asks what the result should cost.

I think this happens because speed is easy to measure and easy to sell. “We cut the time in half” fits on a slide. “We found out half the steps didn’t need to exist” is harder to say in a meeting. It means telling someone their process, or their rule, was waste.

So teams start at the end. They automate. And because the automated process now works a little better, nobody goes back to ask whether it should have been there at all.

Cut first, automate last

Musk’s own fix, as Walter Isaacson laid it out in his 2023 biography, is a set of steps in a fixed order:

  1. Question every requirement. Find out who asked for it and why.

  2. Delete any part or step you can.

  3. Simplify what’s left.

  4. Speed it up.

  5. Automate it, last.

Most teams start at step 5. Some start at step 4. Almost nobody starts at step 1, because questioning a requirement means asking someone senior why they wanted it.

The first rule has a detail I like. Every requirement should come with the name of the person who made it, not a department. “Finance wants it” isn’t an answer. A department can’t explain why it wants something. A person can. Often the person who asked left years ago, and the reason left with them.

The delete step has a test I like even more. If you don’t end up adding back at least 10% of what you cut, you didn’t cut enough. It sounds odd, but it’s honest. Adding a step back isn’t a failure. It’s proof you cut close to the bone. If nothing you removed turned out to matter, you were too careful.

Price the gap before you build

Once you’ve cut, there’s still a question. Is what’s left worth automating?

Here’s how I check, using the Venture Proof method. Take the result the work produces, not the steps. Ask what it costs today. Then ask what it would cost if nothing were wasted. That’s the floor. It holds only what the work truly needs: compute for digital work, materials for physical work and the checks a law or a contract requires. It holds no waiting, no handoffs and no extra sign-offs. It can’t be zero.

Divide today’s cost by the floor. I call that ratio the Physics Gap.

Here’s an example with made-up numbers. Say a company pays outside lawyers $1,200 to review a standard contract. The compute to do a first pass on that same contract might cost around $0.31. That’s an example, and token prices change fast. But the gap is thousands of times. That’s not a process to speed up by 10%. That’s a process to rebuild.

A small gap tells you something too. If today’s cost is close to the floor, the market already runs lean. Automating it won’t buy you much.

A worked example, start to finish

Here’s a made-up invoice process to show the whole order. The numbers are examples, not client data.

A company pays about 24,000 supplier invoices a year. One run is one invoice paid correctly. Today it takes nine steps. A clerk opens the email and saves the file, keys the invoice into the finance system and matches it to the purchase order. Then the clerk prints it and walks it to a manager. A manager signs it. A second manager signs it. The clerk scans the signed copy back in, enters the approval and schedules the payment. That’s one hour of clerk time and 12 minutes of manager time on every invoice.

Question every requirement. The team asks who wanted each step. The second signature was added years ago after one bad payment. Nobody remembers who added it. The written policy only asks for a second sign-off on invoices over $25,000. That’s about 2% of invoices. The print, walk and scan steps exist because approvals used to be on paper. The finance system has had an approve button for years. And the policy only asks for one manager approval on invoices over $5,000, about 20% of them.

Delete. The print, walk and scan steps go. Manager approval moves into the system, and only over $5,000. The second sign-off stays, but only over $25,000. A month later, the team adds one check back. A supplier’s first invoice now gets a manager look, whatever the amount. That’s the add-back the 10% test predicts.

Price it. Say the clerk costs $40 an hour, fully loaded, and the manager costs $90. Before the cut, each invoice costs about $60: $40 of clerk time, $18 of manager time and about $2 of rework. That’s about $1,440,000 a year. After the cut, it’s about $30 an invoice. The floor is about $2, mostly the approvals the policy requires, plus about 5 cents of compute. So the gap was about 30 times before the cut, and about 15 times after.

Look at that. Cutting alone saved about $720,000 a year, and no AI was involved yet.

Rank and test. Of the steps left, keying and matching hurt most. They’re the slowest, and they cause the rework. So the team tests an invoice-reading tool for two weeks on 200 invoices, with a clerk checking every one. They set the rules first. It’s a go if errors stay at or below today’s 1 in 10, and keying plus matching drops under 15 minutes. The result is about 1 in 25 errors and about 6 minutes. That’s a go.

Automate last. The tool now opens, keys and matches each invoice. The clerk handles only the ones it flags. The new cost is about $9 an invoice, or about $216,000 a year.

Now look at where the money came from. Cutting saved about $720,000. The AI saved about $504,000. If the team had put AI on the original nine steps, it would have sped up keying and kept the paper loop and the extra sign-offs. It would have paid for a faster version of the waste.

Test before you spend

The last step is the one teams skip most often. Before you fund a big build, run a small test that does the work by hand, or close to it, for a few real customers.

Here’s an illustration of how that can go, not a client story. A team plans a $2.4M software build for claims handling. Before they start, they run a 4-week test that costs $9,200. The test shows the build would solve the wrong problem, so they stop it.

A cheap test that kills a bad build is the best money you’ll spend all year. Write your go and no-go rules before the test starts. If you write them after, you’ll bend them to fit what you got.

The counter-case

The strongest pushback goes like this. “Tesla still automated most of the line. Automation won. You’re just telling people to go slow.”

I agree that automation won on most of that line. That’s why the story matters. Tesla didn’t fail because it used robots. It hit trouble where it automated steps nobody had questioned. My argument isn’t against automation. It’s about the order. Automation is step 5. It’s the right step, once steps 1 through 4 are done.

The second pushback is about speed. “AI is cheap now. Just try it on the process and see what happens.” Trying it is fine. That’s a test, and I’m all for tests. The problem is when the trial turns into the build. Once a process is wired into software, it gets much harder to delete a step. People stop seeing the step as a choice. It becomes part of the system.

The third pushback comes from regulated work. “We can’t just delete steps. Some of them are the law.” Right. Keep those. Mark every step a law or a contract truly requires, and leave it alone. In my experience, the steps people defend as required often turn out to be habit. The only way to know is to ask who made the requirement and find the written rule.

What to do Monday

Pick one process you’re thinking about putting AI on. Then:

  1. Write one sentence on what one run of the process is, and the result the customer pays for.

  2. Write down every step, the way it runs today. Note who does it and how long it takes.

  3. For each step, get the name of the person who asked for it, not a department.

  4. Mark the steps a law or a contract truly requires. Those stay.

  5. Cross out every step you can’t tie to the result.

  6. For what’s left, estimate today’s cost and the floor. Divide.

  7. Score the surviving steps from 1 to 5 on how much they hurt.

  8. Write your go and no-go rules, then test the top step for two weeks on real work.

  9. Automate only the steps that survive, show a big gap and pass the test.

  10. A month later, check what you’ve added back. If it’s nothing, cut again.

Where this breaks

This order has limits.

First, the floor is an estimate. If you set it too low, you’ll think there’s more room than there is. List its parts so someone can argue with each one.

Second, deleting steps can upset people. Every step has an owner who thinks it matters. Go in with the written policy and the name behind each requirement, not with opinions.

Third, compute prices move. The contract example uses a number that will be wrong soon. Re-price the floor each time you run this.

Fourth, some processes are already lean. If the gap is close to 1, don’t force a rebuild. Speeding it up may be the right call.

The close

Tesla pulled robots off a line it had built to be run by machines. Musk called it his mistake. The lesson isn’t that robots are bad. It’s that robots went in before anyone asked the questions.

You’re about to make the same choice with AI. Cut first. Price the gap. Test small. Automate last.

Get the free download

Grab the free checklist, with the playbook included. The checklist gets you started this week. The playbook walks you through the full invoice example, step by step.

This is the kind of work we do in the Innovation Unpacked Mastermind.

If you want to talk through a process of your own, book a call.


Sources

  • Musk on X, Apr 13, 2018, via TechCrunch (also the CBS conveyor quote); see also The Guardian, Apr 16, 2018

  • Fluffer bot and “vast majority” automated, Q1 2018 earnings call, via BGR

  • Musk’s five steps, the name-not-a-department rule and the 10% rule: Walter Isaacson, Elon Musk (Simon & Schuster, 2023), chapter 46, “Fremont Factory Hell,” via FS Blog


Book an appointment: Click here
Email me: mike@pjtbd.com
Call me: +1 678-824-2789
Follow me on 𝕏: https://x.com/mikeboysen
Articles - jtbd.one - De-Risk Your Next Big Idea
Website - pjtbd.com

Slaying the Sacred Cows of Innovation

Discussion about this video

User's avatar

Ready for more?