The tool was never the talent: AI is nothing without the education to guide it

A group turns on an AI voice agent across nine rooftops. Vendor onboarding takes two hours. The person who owns it is a marketing coordinator with eighteen months in the industry, a strong work ethic, and no service-drive experience: she has never listened to two hundred recorded advisor calls, does not know what a declined recommended service is worth in gross per visit, and has never had to explain to a service manager why a booking slot that looks empty on a screen is not actually empty on a Tuesday at 9am.

Six weeks later the agent is doing three things well: answering the phone, capturing a name, and sounding confident. It is doing three things badly. It books work the drive cannot honor, so advisors spend the morning apologizing. It quotes maintenance intervals that do not match the state the customer's own dashboard is showing, because that data lives in a system nobody gave it access to. And it answers "what will a brake job cost" with a range so wide that the caller hangs up, because a range is what you produce when you do not know the vehicle, the pad wear or the rotor condition.

The dashboard says the agent handles 71 percent of calls. The CSI score says something else entirely. Nobody is lying. The instrument is measuring containment, and containment was never the goal.

That is not a story about a bad model. It is a story about a good model with nobody in front of it who knew what right looked like.

The models are not the bottleneck, and the research is unambiguous about it

The MIT NANDA report on the state of AI in business, published in 2025, is the one to keep on the desk. It covered 150 leader interviews, a 350-employee survey and 300 public deployments, and found that roughly 5 percent of enterprise generative AI pilots produced rapid revenue acceleration while the rest delivered little or no measurable impact on the P&L. The authors were explicit about the cause: not model quality, not regulation, but a learning gap in both the tools and the organizations running them. Generic tools work beautifully for individuals, who adapt their own behavior around them. They stall in enterprises, where nobody's behavior changes.

Gartner's forecast is the same finding wearing a different coat. More than 40 percent of agentic AI projects will be canceled by the end of 2027, with escalating costs, unclear business value and inadequate risk controls given as the reasons. Read that list again. Every item on it is a human management failure with a technical surface: costs nobody scoped, value nobody defined before the build, controls nobody specified because nobody in the room knew where the risk was.

Neither finding is news to anyone who has run a project. Both are invisible to the organization that has already decided the technology is the answer.

The frontier is jagged, and only experience can see the edge

The most useful phrase in this whole literature came out of a field experiment with 758 consultants at Boston Consulting Group, since published formally in Organization Science: the jagged frontier. The idea is simple and it changes how you hire. AI capability is not a smooth improvement across all work. It is excellent on some tasks, mediocre on others and actively harmful on a third group, and the boundaries between those groups do not follow any pattern you can see from the outside.

On tasks inside the frontier, consultants with AI completed 12.2 percent more tasks, 25 percent faster, at roughly 40 percent higher quality. On a task deliberately chosen to sit outside the frontier, the same consultants using AI were 19 percentage points less likely to produce the correct answer than the ones working without it. Same people, same tool, opposite outcome. The only thing that changed was whether the task happened to fall on the good side of an invisible line.

A model cannot tell you which side of that line you are standing on. It answers with the same fluency either way. That is the entire reason a person with fifteen years of category experience is not made obsolete by this technology: they are the only part of the system that can feel the ground give way. They know that a 40 percent higher-quality output on the wrong recommendation is just a faster route to the wrong answer.

And confidence is not a proxy for calibration. METR ran a randomised controlled trial with experienced open-source developers and found that access to early-2025 AI tools made them 19 percent slower on their own repositories, with a confidence interval from a 2 percent slowdown to 39 percent. The developers believed they were 20 percent faster. The people best equipped to judge their own performance were wrong about it by nearly forty points. If that is true of senior engineers working on code they helped write, it is true of a marketing manager judging a vendor's output.

Three ways to use the machine, only one of which looks like skill

A follow-up study tracked 244 junior associates and consultants at BCG through a real memo task, with every interaction with a custom GPT-4 platform time-stamped. Three modes emerged, and the labels have stuck because they are accurate.

Centaurs, 14 percent, kept control of the questions. They knew what they wanted to know, asked narrow questions, used the model as a tool and held their own domain judgment intact. They had the highest accuracy of the three groups. Cyborgs, 60 percent, fused with the tool, iterating back and forth at every stage, sometimes following it and sometimes pushing back. They were roughly equal to centaurs on persuasiveness and learned a great deal about working with AI, less about the domain. Self-automators, 27 percent, handed the task over. Their output was quick and polished and lacked depth, and they lagged on persuasiveness.

The study's authors noted that 27 percent is probably an underestimate for the wider workforce, because their sample was a diligent population being watched. In a dealership on a Friday afternoon, the real number is higher.

Notice what the self-automators had in common. It was not laziness. It was that offloading is the natural behavior of someone who cannot yet evaluate the output. If you do not know what a good service-lane script sounds like, generating ten of them feels like progress. If you do, generating ten feels like a waste of a Tuesday, because you can already see that the ninth one tells a customer to do something the maintenance schedule does not support.

What disappears from the bottom of the ladder

There is a second-order cost that the productivity spreadsheets never carry. BCG's 2026 research on what they call distributed de-skilling found that almost 90 percent of surveyed leaders point to overreliance on AI output without stress-testing or challenge, and 53 percent point to the slower development of junior talent. One leader described training days that are "no longer used to learn but to generate content with AI tools." Another observed that the traditional entry-level employee is now expected to perform work that used to take five to ten years to earn.

That is the mechanism, stated plainly. The repetitions that build judgment in a junior person, the research, the drafting, the debugging, the fifty rejected versions of a headline, are exactly the repetitions the tool now absorbs. And those repetitions were not busywork. They were how you learned to recognize a good answer before you were ever responsible for one.

So an organization can now do something genuinely new and genuinely stupid: buy the tool, remove the apprenticeship that produced the person capable of supervising the tool, and then wonder, three years later, why nobody on staff can tell whether the output is any good. The purchase did not create capability. It removed the conditions under which capability used to form.

Automotive is running this experiment at scale right now

Cox Automotive's 2026 dealer tracking put AI use at 82 percent of dealers, concentrated in customer follow-up, content generation and task automation. Nearly half of those dealers say they feel behind their peers regardless, and 69 percent of those using AI expected it to drive sales and revenue growth. The most interesting line in the study is quieter than the headline: dealers working with external AI partners were significantly more likely to report optimal use of the tool, greater confidence in the outputs and stronger business outcomes, and many were leaning on internal AI champions in sales and marketing to drive adoption and education across the store.

Read that as a statement about knowledge, not about software. The dealers getting results had someone, inside or outside the building, who could tell good output from bad and who had the standing to reject it.

Spyne's 2026 sentiment report, drawn from nearly 1,200 dealership leaders, found 76 percent planning to increase AI budgets, with AI voice agents the top target at 74 percent, merchandising and inspection automation at 68 percent, pricing and analytics at 62. Notice the ordering. The categories where a bad output is hardest to spot, a phone call nobody hears or a price recommendation nobody tests, are at the top of the list. The categories where the mistake is immediately visible, a photo of the wrong vehicle or a mis-keyed trim, come later. The market is buying first where supervision is hardest and cheapest to skip.

Now set that against the same report's most useful caution, that adoption is uneven and expectations are running ahead of results. Both things are true at once, and they are true for one reason: the tool arrives in weeks, and the judgment it requires takes years.

The content version of the same mistake

The same pattern shows up wherever output is words on a page, and Google has been unusually direct about it. The scaled content abuse policy, last updated in August 2026, defines the violation without once mentioning automation: many pages generated for the primary purpose of manipulating search rankings and not helping users, whether automation, humans or a combination is involved. The company's own guidance on generative AI content, updated in October 2026, says it plainly: generative outputs may contain inaccuracies, and it is critical to manually fact-check and review all AI-generated content before publishing.

There is no configuration of this where you skip the reading. In automotive, the reading is the product. A hallucinated towing capacity, a fitment that does not exist for that trim, a maintenance interval that contradicts the manufacturer's own schedule: those are not typos, they are liability with a publication date. The editor is not overhead on top of the AI workflow. The editor is the reason the workflow is allowed to exist.

What the words actually mean

Two words get used as if they were soft. They are the hard part.

Experience buys four specific things, and none of them are downloadable. It buys the ability to recognize a good answer on sight, which is faster than any evaluation rubric. It buys a working sense of what being wrong costs, in gross, in CSI, in a customer who does not come back. It buys knowledge of which of your data sources lie, and how often. And it buys the standing to overrule the machine in a room where someone has already promised a board that the machine is the strategy.

What it does not buy, and this is where the industry has been lazy in both directions, is automatic authority over the tool. Senior people are the most likely to use AI badly and be certain they did not, which the METR trial demonstrates with unusual cruelty. Education is not a substitute for experience, and experience is not a substitute for education. You need both, and the second one has to be taught.

One more thing worth saying out loud: prompt engineering courses are not education. A prompt library is a snapshot of one model's behavior on one afternoon in one version. It rots in weeks. What does not rot is knowing what the output should look like before you read it.

A training plan that survives contact with a budget

Most dealer AI programs spend the training budget on a vendor webinar and a shared document of prompts. Here is the version that holds up.

The cost of that is honest and it is not zero: somewhere between a quarter and a half of one FTE's year, plus a partner if you do not have the reviewer internally. Set it against a single failed voice agent deployment across nine rooftops: the advisors' lost mornings, the rebooked appointments, the reviews, the six months before anyone trusts the next tool. The training is the cheap line item. It just is not the line item anyone gets excited about in a board deck.

The uncomfortable accounting

The tool is the cheap part now. Inference costs are falling, the interfaces are converging, and every vendor in the category can demo something impressive in eleven minutes. What is scarce is a person who knows what the output should say, can prove it, and will stop it when it is wrong.

Buy the tool without that person and you have not bought capability. You have bought a machine that produces confident work you cannot check, at a volume you cannot review, in a business where being confidently wrong ends in a warranty claim. The education is not the implementation cost of AI. It is the thing that decides whether the AI was ever worth buying.

Every group is currently being told it is behind on AI. Almost none of them are behind on tools. They are behind on the years of judgment that make a tool useful, and those years cannot be purchased in a quarter, only started.

For what that looks like inside a single store group, read A department of one.