Knowing, Doing, Deciding, Owning.

Humans Matter More, Not Less.

For months, most of what I've written about AI and the trades has circled one idea: AI is making knowledge cheap. A technician who once needed years to accumulate the failure modes, part numbers, and service histories that separated an expert from a beginner can now access most of that in seconds. That's true, and it matters. I've come to think it isn't the actual point.

The actual point is what happens after the answers show up.

AI doesn't just make information available. It makes it abundant, fast, and cheap enough to generate on demand, in volumes nobody asked for. Ask it a diagnostic question and it won't give you one answer. It will give you several, all technically defensible, and leave you to sort out which one is right for the situation actually in front of you. That sorting is judgment. The narrower skill of telling which of several plausible answers is actually true, wise, or appropriate here is discernment. Both get more valuable, not less, as machine intelligence gets more capable. Not because humans are special. Because someone has to decide what the abundance is for.

In this article, we’ll explore these main ideas: 

KNOWING → AI makes access to knowledge cheaper.
DOING → Physical experience and craftsmanship remain important.
DECIDING → Someone has to determine what actually matters.
OWNING → Someone has to stand behind the decision.

Information is getting cheap. Options are getting cheap. Good human judgment isn't.

Knowing

Start with knowing, since that's where the AI conversation usually starts and stops.

A veteran HVAC technician's expertise has always been partly a knowledge problem: thousands of failure modes, the sound a specific compressor makes right before it fails, which brand had a bad run of parts in 2014, what a service history that doesn't quite add up usually means. That knowledge took years to build because it was expensive to acquire. You had to see the failure, or know someone who had.

AI collapses that cost. A 2025 study published in the Quarterly Journal of Economics (Brynjolfsson, Li, and Raymond) gave real customer-service agents access to a generative AI assistant and tracked what happened to their output. The least experienced agents improved dramatically, handling about 35% more inquiries per hour. The most experienced agents barely moved, and in some cases ignored correct suggestions the AI gave them. New hires reached typical eight-month proficiency in about two months.

That's a study about call centers, and I'd resist the temptation to claim it proves what happens to a plumber or an electrician. But it's suggestive of something worth taking seriously: AI may raise the floor much faster than it raises the ceiling. It closes the gap between someone who knows very little and someone who knows a moderate amount. It does much less for someone who was already excellent.

A version of the same pattern already happened to bookkeeping. Federal labor projections show clerical bookkeeping work, the procedural layer of accounting, declining over the next decade as software absorbs it, while accountants and auditors, the layer that requires judgment about what a number actually means, are projected to keep growing. The knowledge didn't stop mattering. Simply possessing it stopped being enough to build a career on.

Which raises the real question underneath the knowing conversation. If a technician with two years of experience and a good AI assistant can access nearly the same diagnostic information as someone with twenty years in the trade, what does the twenty years actually buy? Knowing more stops being the differentiator. Something else has to be.

Doing

The something else starts to show up in the doing.

AI can tell you how a joint should be sweated. It can't sweat it. It can tell you how conduit should be routed through a wall. It can't get it through an eighty-year-old house where nothing is square, the stud bay is full of someone else's abandoned wiring, and the manual's diagram assumes conditions that don't exist in this particular structure. Physical execution hasn't gotten any cheaper, and there's a real reason to expect it won't for a while: skilled trades have always resisted the kind of productivity gains manufacturing got, because the work refuses to standardize. Every house is a slightly different problem. If anything, as the knowledge layer gets commoditized, the physical layer becomes a larger share of what separates a good contractor from a mediocre one, not a smaller one.

Doing serves a second, less obvious purpose too. It's where judgment gets built. Diagrams are clean. Houses aren't. A technician who has spent years working in real basements and real attics has absorbed thousands of small lessons about when the textbook answer doesn't fit what's actually there: a previous contractor's strange workaround, equipment that behaves differently than the spec sheet claims, a customer's unstated constraint that changes the whole calculation. That accumulated friction with reality is what turns information into judgment. It's hard to get any other way, and it's exactly the part AI can't shortcut, because AI hasn't touched the house.

Deciding

Which brings us to deciding, and to the scene that actually matters here.

A technician is standing in a basement. The system is fourteen years old. AI has already pulled the model, the service history, the fault codes, the manufacturer's recommendations, current repair and replacement costs, and typical energy savings from an upgrade. It has produced five defensible recommendations, ranging from a cheap patch to a full replacement, in about the time it took the technician to open the panel.

Now what.

The homeowner expects to move in three years. Money is tight this year specifically, for reasons that have nothing to do with HVAC. Their daughter has asthma, so the air quality argument matters more here than the energy savings argument does. The ductwork is questionable enough that even a good repair might not solve the actual comfort complaint. There may not be one objectively correct answer sitting in that list of five. There's a right answer for this family, this house, this year, and figuring out which of the five fits is not a task the AI can finish on its own, because it doesn't know most of what just changed the calculation.

That's the distinction worth being precise about. Judgment is making the call. Discernment is recognizing which facts in the room actually matter enough to inform it. AI is very good at the first half of this scene, laying out the options, and not particularly good at the second half, weighing them against a specific family's specific circumstances. The technician who used to be valuable mainly for generating the list is now competing with something that generates the list instantly and for free. The technician who's valuable for knowing which item on the list is right, for this house, and can say so plainly, is not.

I want to push back on my own argument here, because it's tempting to treat human judgment as automatically good simply because it's human. It isn't. People aren't reliably well-calibrated judges just by virtue of having judgment at all. A study by Logg, Minson, and Moore found that domain experts, in their case national-security professionals, discounted correct algorithmic advice more than laypeople did, and were measurably less accurate as a result. Experience can produce wisdom. It can also produce a stubborn confidence that isn't earned. The scarce, valuable thing isn't human judgment in general. It's good human judgment: the kind that knows when to trust the five options AI just generated, and when to notice all five are missing something the room is telling you. That's a narrower, harder thing to build than simply being human and having an opinion.

Owning

Eventually someone has to take responsibility for the decision, and this is where a sentence like "if this were my house, here's what I'd do" carries weight an AI-generated recommendation doesn't.

That sentence isn't just information. It's a person telling you which facts they weighed, what they'd choose under uncertainty, and that they're willing to have their name and their reputation attached to being wrong about it. A recommendation with nobody's name on it is an option. A recommendation someone stands behind is a decision. Homeowners have always paid, at least in part, for that difference, whether or not they'd describe it that way.

It would be too easy to claim this stays permanently and exclusively human. It might not. Warranties, platforms, and brands can absorb a real share of accountability over time. An extended warranty is already a kind of institutional promise that doesn't depend on trusting the specific person in your basement, and it's reasonable to expect more of that as AI-assisted diagnostics get better at predicting outcomes. Some of what currently requires a person standing behind their word could migrate to a guarantee instead. That's a real possibility, not a footnote, and it means owning doesn't automatically stay a human premium just because it's currently a human one. What seems more durable is the specific, local, repeated version of it: a person a homeowner can call back, whose name is attached to a decision they can still find in three years if it turns out to be wrong. That's harder for an institution to replicate than a warranty is.

The Paradox Worth Testing

When answers are scarce, knowing one is valuable. When answers are abundant, and AI is rapidly making them abundant, knowing which one to trust becomes the valuable thing instead. The better AI gets at generating defensible options, the more the actual bottleneck moves to deciding among them. More machine intelligence doesn't shrink the space where human judgment operates. It's more likely to expand it, because more capability generates more decisions that need to get made, not fewer.

That doesn't hold unconditionally, and it's worth being precise about where it doesn't. AI should eliminate the need for human judgment in plenty of routine decisions, and that's a feature, not a failure. Nobody needs a technician's judgment to answer what are your hours, and a scheduling decision with no real tradeoff hiding inside it doesn't need one either. The interesting question is what happens to the decisions that remain. They're disproportionately likely to involve ambiguity, competing priorities, incomplete information, unusual physical conditions, or values that can't be optimized against a single objective, which is a fair description of the basement scene above. AI may leave humans with fewer decisions to make. But those decisions may matter more.

That's the sense in which judgment becomes more valuable. Not because every choice requires a person, but because once machines handle the obvious choices, human attention can move toward the ones where there's no obvious answer. Put plainly: AI may reduce the number of decisions humans need to make while increasing the importance of the decisions that remain.

A related idea is worth testing too: that AI and judgment are complements, not competitors. Take a technician who already has twenty years of pattern recognition, genuinely excellent hands, and the humility to know what they don't know, and hand them an AI that gives instant recall of failure histories across thousands of jobs, plus something that can challenge their first instinct before they commit to it. That person doesn't get replaced by the tool. They get sharper, because the tool is doing retrieval work that used to eat into the time and attention they could spend on the actual decision. AI may raise the floor for the inexperienced. For someone who already has excellent judgment, it may raise the ceiling too, not by deciding for them, but by giving their judgment better material to work with.

The Owner's Version of the Same Problem

The same shift is coming for the person who owns the truck, not just the person driving it.

AI can already produce a marketing plan, a pricing model, a financial analysis, a hiring shortlist, a demand forecast, an operating procedure, competitive research, a growth strategy. Soon an owner may have access to more sophisticated business advice before breakfast than an earlier generation of owners received in a year of paying a consultant. That sounds like it should make the owner's job smaller. It doesn't. It changes what the job is.

A plan being cheap to generate doesn't make it right for this company. An owner still has to decide which recommendation actually fits their crew, their market, their balance sheet, and their appetite for risk, and which one is generic advice that happens to sound confident. They still have to decide what to prioritize this quarter, what to automate, where to deliberately keep a human in the loop even when the AI version would be faster, and what kind of company they're actually trying to build, which isn't a factual question any tool can answer for them. The work of generating a strategy is getting cheap. The responsibility for choosing one, and living with what it costs when it's wrong, isn't.

What This Means for How We Teach the Trades

Underneath all of this is something I think Raise the Trades should take seriously as we figure out what we're actually for. The instinct, understandably, is to teach contractors and technicians to use AI well: which tools, which prompts, which workflows. That's necessary. It isn't sufficient, and I'd argue it isn't even the harder half of the problem.

The harder half is strengthening the things that determine whether someone uses AI well in the first place: the ability to think critically about a recommendation instead of accepting it, to notice what a diagnostic tool didn't ask about, to question a confident-sounding answer, to explain a tradeoff to a scared or skeptical homeowner in plain language, to recognize genuine uncertainty instead of papering over it, to change your mind when new evidence shows up instead of defending your first guess, to understand what a specific person actually needs instead of what the average customer needs, and to take responsibility for a call once you've made it. Those get filed under soft skills, usually by people who haven't had to make a real decision under pressure with incomplete information and their name on the outcome. They're about to become some of the hardest, and most economically valuable, skills in the trades.

Humans Matter More, Not Less

That line is not a comforting slogan. It's a bet about where economic value goes once intelligence stops being scarce.

Information is getting cheap. Options are getting cheap. Generating a plausible answer is getting cheap. Standing in the basement, weighing what actually matters for this family, and being willing to say this one, and here's why, and I'll stand behind it, isn't getting cheap at all. If anything, it's becoming the most valuable thing in the room.