We Have It Backwards: Supervised AI Is the On-Ramp to Autonomy

Robotaxi didn't come first. Supervised driving did. And most companies are rolling out AI in the opposite order. If my car is driving itself and I get pulled over, I get the ticket. Not the car. Me. I don't argue with the officer and say, give it to the software. I was supervising. I own it. Hold onto that, because I think it's the whole argument for how to bring AI into a company.

Picture two roads. The first is the Robotaxi road. In AI terms, that's autonomy first. Agents act on their own, and no person is accountable for each call. So the way I see it, you have to max out the prerequisites before you get any production value. The ontology and the semantic layer, yes. Autonomous agents need those. But they're only part of it. You also need governance, guardrails, evals, exception handling, a plan for liability, and a scope narrow enough to prove. That's how real robotaxis roll out, too. Tesla started in Austin in June 2025. About ten cars, inside a fenced-off part of town, with a safety monitor in the passenger seat. It was about seven months before the first public rides with nobody in that seat. As I record this, it's eight cities. Waymo has gone city by city, too. One geofence at a time. Every new patch has to be proven before it opens. And this part is my guess, not a fact. Some rural roads may never get it. I think autonomous AI plays out the same way. One narrow use case at a time.

The second road is supervised self-driving. The car will drive you almost anywhere. But you're responsible. Tesla's own fine print says it plainly. It needs an attentive driver, ready to take over at any moment, and it doesn't make the car autonomous. In AI terms, this road means you meet people where they are. You give the worker an agent, right in their real, messy environment. The spreadsheets. The workarounds. The stuff that lives in one person's head. And the person stays accountable. The controls you already have still apply. Least privilege. Monitoring. Audit trails. None of that goes away because a bot is helping.

Most companies pick the first road. So how's it going? McKinsey's 2026 State of AI survey looked at this. About 37 percent of respondents said AI has touched their bottom line at all. About 6 percent said it's made a big difference. And that number hasn't moved in a year. A survey can't tell you why. So this next part is my opinion. I don't think data is the whole story. I think they're doing things in the wrong order. Here's the difference. Autonomy only counts as a success if it's near perfect. Supervised AI just has to beat a person working alone. That gap is why AI is stalled in so many companies. And I think there's something underneath it that no amount of data work fixes.

Let me tell you a story. At a gaming company, they'd spent fifteen million dollars on gaming analytics infrastructure. I asked the executive team, how are you using it? The answer was, "We look at all the analytics, but in the end we just go with our gut." That's been true at almost every company I've been in. And it's not just me. KPMG asked CEOs about this back in 2018. Two out of three said they'd ignored what their data models told them when it went against their gut. Billions get spent across the industry on analytics, data quality, data management. If the people making the call don't trust it, it doesn't matter. That's a trust problem. And no ontology fixes trust. So how do you build trust?

You start where people actually work. Every company has a Rube Goldberg machine. A spreadsheet feeds a tool, the tool feeds another spreadsheet, and a human stitches it all together at the end. Most of that is meat robot work. Copy, paste, check, chase. People hate it. The Robotaxi approach looks at that and says, let's automate the whole department. The supervised approach asks a smaller question. What sucks the most? Go after that first. The mundane and important stuff. The work that has to be right, and that nobody wants to do.

Here's mine. Email. VIP email dread. My inbox is about 95 percent noise and 5 percent signal. It reminds me of Voyager. That probe talks to us with about 23 watts. Roughly a fridge light bulb. By the time the signal reaches Earth, NASA says it's a fraction of a billion-billionth of a watt. That's what finding the one email that matters feels like. So I handed it to Grokbot. It took a few weeks to build trust. Now I don't even go into my inbox anymore, and I don't worry about missing the signal. If I do miss something, that's my fault, not the bot's. Same as the ticket. I'm still the driver.

Now here's the part I think we have exactly backwards. Robotaxi didn't make supervised driving possible. Supervised driving made Robotaxi possible. Years before the first Robotaxi ride, Tesla was learning from ordinary owners just living their lives. Back in 2019, Andrej Karpathy, who led Tesla's Autopilot AI work then, laid out how. The cars watch for interesting moments. A new model runs quietly in the background, in what they call shadow mode. It makes its guess, and that guess gets compared to what the human actually did. When a driver takes over, that's a signal too. The drivers were the labelers. Tesla's own counter now shows more than 17 billion miles driven on FSD Supervised. That's Tesla's number. And every one of those miles had a human responsible behind the wheel. Now bring that back to work. When your people work with an agent, they make corrections. They fix the data. They override the bot. You can see what comes out the other end of their Rube Goldberg machine. All of that is labeled data. That's the raw material for an ontology.

And this isn't either-or. I don't buy the tyranny of the or. It's a layer cake. Layer one is morale. When you take away the work people hate, they feel it. A CFO can't put that in a spreadsheet. It's still real. Layer two. The people overseeing these bots have the domain knowledge. They're closest to the information. So you amplify them. You don't try to pull what's in their heads out and file it away somewhere. Layer three. What they do with the bot gets captured. And that captured work feeds whatever you automate next.

There's a meme Elon Musk likes. He calls it the Jose meme. One guy digging the hole. Eight people watching him dig. Every company has its Joses. A small number of heroes who end up as the choke point on every launch. So go to them first. Arm them with Grokbot. While you amplify their work, you capture what they do. That feeds right back into how we should be doing this, and what we can automate next. That's the loop back to Robotaxi. And sometimes it shows up fast. A financial analyst was in his initial onboarding onto Grokbot. He had a job due the next morning that was going to take hours. He deployed a bot, and it did the work for him immediately. Zero data prep. The onboarding was the value delivery.

Now I want to be honest about the risk. Handing a bot to a human is not risk-free. Bots move faster than people. A person can use poor judgment when they deploy one. So you need detection. You need to be able to see when something's off. And the hard part is, we don't know yet what normal looks like with these agents. There are a lot of new problems to solve. Even with cars, blame doesn't always land in one place. In 2025, a Miami jury split the fault in an Autopilot crash, and put a third of it on Tesla. So yes, the human owns it. Build the guardrails anyway. In my view, the Robotaxi path for AI means a very long wait before you see success. And the two roads aren't mutually exclusive. Run your data initiative on its own ROI. At the same time, arm your highest-trust, most productive people. Do both.

So here's where I've landed. We have it exactly backwards. If you want to build an ontology, you'd better be arming the people working in the information workflow with Grokbot. Because you can be capturing what they're doing, and the corrections they're making to data. You can look at the outputs of that Rube Goldberg machine. It's not supervised versus autonomous. It's the order. Go autonomy first, and you have to max out every prerequisite before you see any value. Go supervised first, and you get value now. And the supervised work builds what autonomy will need. And then, when you're ready, Robotaxi arrives.