AI Doesn't Need the Car Keys Yet — It Needs a Learner's Permit

Companies handed AI the car keys. Now, quarter after quarter, they're calling the old driver back.

Ford is rehiring hundreds of experienced engineers this year to fix quality problems its automated systems couldn't catch (Forbes, 2026). IBM's AI now handles 94% of routine HR requests — and quietly failed on the other 6%, the cases requiring judgment rather than pattern-matching, prompting the company to triple its entry-level hiring in 2026 (Fast Company, 2026). And Klarna, once the industry's poster child for AI-driven displacement, backed off after its own CEO admitted the strategy had backfired. "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality," Sebastian Siemiatkowski told Bloomberg, adding that "really investing in the quality of the human support is the way of the future for us" (Siemiatkowski, quoted in Bloomberg, 2025).

Writing in Fast Company, future-of-work researcher Dan Schawbel names the pattern behind all of it, one he's tracked for fifteen years through automation, offshoring, and now AI: "Cut fast, overpromise the technology, underestimate the human capability you just removed, and spend the next several years trying to recover it" (Schawbel, 2026). The numbers back him up. According to Careerminds' industry survey, roughly two-thirds of companies that laid off workers citing AI are now rehiring some of those same people, and nearly a third have reopened the exact roles they eliminated (Careerminds, 2026; Robert Half, 2026). CNBC reports that about half of companies that made the swap are now caught in a boomerang effect, spending more on restaffing than they ever saved on the original cuts (CNBC, 2026).

This isn't a tech-sector story. It's happening in manufacturing, financial services, customer operations, retail — anywhere a spreadsheet made the case for AI adoption look cleaner than the reality turned out to be. As a venture investor who has spent the last several years underwriting AI companies and sitting close to how organizations actually adopt this technology, I've watched the same mistake repeat itself with remarkable consistency: leadership optimizes for opex, and forgets to ask what else is riding on the people they're cutting.

The Cost That Doesn't Show Up in the Model

The pitch for AI adoption is almost always framed as a cost equation — fewer headcount, faster cycle times, lower burden per unit of output. That math is real. But it's incomplete. What it leaves out is harder to quantify and shows up late: the customer who could tell something was off the moment a human stopped answering the phone, the brand voice that got flatter once the product marketer was replaced by a prompt, the institutional judgment that walked out the door with the "redundant" specialist who was the only person who remembered why a process existed in the first place.

By the time that cost is visible, it's already been paid. Employer surveys reported by CNBC put the regret rate at 55% among companies that made AI-driven cuts (CNBC, 2026) — not because AI failed as a technology, but because the decision to deploy it skipped the step where someone asks what, specifically, is being replaced, and whether the organization actually understands what that role was doing.

From Earned Caution to Unearned Confidence

I've seen this swing before, from the inside. Earlier in my career, I worked in marketing leadership in the medical device industry — an environment built around protecting patient health information and personally identifiable data. The caution around cloud adoption in that world wasn't bureaucratic timidity; it was earned. Every migration decision ran through one question: what happens if this data boundary doesn't hold? That instinct — treating a technology's blast radius as seriously as its upside — is exactly what's missing from how many organizations are approaching AI today.

Despite the guardrails vendors promise, AI systems still bleed boundaries they shouldn't and hallucinate with the same confidence whether they're right or wrong. In healthcare and other regulated industries, that should be sobering — a hallucinated boundary isn't a minor bug, it's a compliance and patient-trust failure. Yet the posture across most industries has flipped from cloud-era caution to AI-era overconfidence, often for functions carrying comparable sensitivity. The postmortems keep landing on the same root cause: the AI wasn't finished being built, wasn't trained on the right data, or was deployed a stage earlier than its actual capability warranted — the same mismatch Klarna's own leadership later admitted to publicly (Siemiatkowski, quoted in Bloomberg, 2025).

The Learner's Permit Problem

The analogy I keep coming back to is a teenager with a driver's license. Passing the test doesn't mean they're ready to drive other people's kids to school. They need supervised hours, a parent in the passenger seat, graduated trust — not because we don't want them to become capable, but because we do, and proficiency is built in stages, not granted on day one.

AI adoption deserves the same discipline. Human-in-the-loop isn't a permanent tether or an admission that the technology doesn't work — it's the supervised-driving phase every capability has to go through before it earns full autonomy. The goal isn't to keep AI in the passenger seat forever. It's to get it genuinely road-ready, so that the human can move on to work that only a human can do — which is a very different outcome than firing the driving instructor the day the teenager gets their permit.

The counter-example is instructive, and it isn't hypothetical. When Ingka Group, which operates the majority of IKEA stores, deployed an AI chatbot able to handle 47% of customer service inquiries, it chose not to eliminate the roughly 8,500 workers those calls used to require. It retrained them into design consultant roles instead — a bet that helped drive €1.3 billion in revenue in 2024, a channel projected to reach 10% of total revenue by 2028 (Fast Company, 2026). The workers weren't treated as the cost AI was meant to remove. They were the asset that made the AI worth deploying at all.

What Governance Actually Looks Like

This is where I think the conversation needs to move from caution to structure, because caution alone doesn't scale and structure does. Three levels, in my view:

Inside the organization, AI adoption decisions need the same governance rigor as capital allocation — an internal board or committee that asks what institutional knowledge is being displaced, what the failure modes look like, and who owns the consequences when the model is wrong. In practice, that means program managers building structured checks into every AI integration the way they would for any other high-risk system change: pre-deployment failure-mode reviews, defined rollback plans, and ongoing monitoring for where a model's actual boundaries sit in production, not just in the vendor's demo. Too many AI rollouts today are owned by whichever team moved fastest, not by a process built to catch what that team missed.

Across the industry, we need consortiums — peer companies willing to share what's actually working and what isn't, so every organization isn't relearning the same expensive lessons independently. The Ford and IBM examples above are public because they were big and visible. Most companies quietly eating the cost of a bad AI layoff never say so out loud, which means the pattern keeps repeating.

And globally, this can't be a conversation any single country runs alone. AI development and governance standards are already becoming geopolitical fault lines, but the underlying challenge — how do humans and AI systems share responsibility for outcomes — is universal. A global consortium approach matters not out of idealism but out of self-interest: in an interconnected economy, limiting good governance practice to one region only limits how far the benefits of AI can compound for everyone, including the countries that got there first.

The Investor's Stake in This

I don't think this is a side conversation for venture capital — I think it's becoming core to how we should be evaluating AI companies and their customers. The startups worth backing aren't the ones promising to eliminate the most headcount the fastest. They're the ones building for the supervised-driving phase — tools designed to make human judgment more effective, not tools designed to make humans optional before they've proven they can be. That's a better underwriting thesis, and I'd argue it's also the better bet on durability.

The organizations that get AI adoption right in the next few years won't be the ones that moved first. They'll be the ones that built the governance muscle to move well — treating this the way any good parent treats a new driver: full confidence in where they're headed, real supervision on the way there.