"When a measure becomes a target, it ceases to be a good measure." That is Goodhart's Law. It was coined in 1975 about monetary policy. It applies with uncomfortable precision to how most dealerships operate today.
The automotive industry is in a familiar cycle. Q1 every year, vendors arrive at NADA with a new class of technology promising transformation. This year, the technology is AI. The pitch decks are polished, the demos are compelling, and the founders are persuasive. By Q2, early adopters will be live. By Q3, half of them will be wondering why the numbers aren't moving.
The problem will not be the AI.
The problem will be the same one that has been silently degrading these operations for years: the metrics being optimized do not reflect the outcomes they claim to represent. The measurement layer is already broken. AI does not fix that. AI accelerates it.
The Metric Becomes the Mission
Goodhart's Law is not abstract. It operates in every dealership in the country, every month, in dynamics so normalized they no longer register as failure modes.
Consider speed to lead. The industry adopted a reasonable insight (faster responses tend to correlate with better outcomes) and compressed it into a commandment. Sub-five-minute response became "table stakes." Vendors built automation around it. BDC managers began tracking it as a primary KPI.
The problem: a two-minute response to an internet lead is often not enough time to read the lead. Customers submit forms with vehicle preferences, trade details, availability questions, financing concerns. A lead that takes thirty seconds to read and another sixty to respond to thoughtfully gets a rushed, generic template, because the clock is running and the metric demands speed over substance.
The result is predictable. Reps blast template responses, hit their speed targets, and the dashboard turns green. The customer who asked a specific question about towing capacity receives "Thanks for your interest! When can you come in?" and moves on to the next dealer who read what they wrote.
No one reports the second-order effect. The speed metric improves. The quality of initial engagement degrades. The connection between the two becomes invisible because no one is measuring what the speed metric was originally supposed to represent: meaningful, informed contact.
When Closing Percentage Lies to You
Speed to lead is a relatively clean example. Closing percentage is murkier and more dangerous.
On the surface, closing percentage appears to be a reasonable proxy for sales effectiveness. It is not. Closing percentage is downstream of everything else in the store, and the number of ways it can be distorted by upstream dysfunction is significant.
If your store uses addendums, market adjustments, or fees that surface late in the process, your "closing percentage" includes deals that closed because the customer was too far in to walk away. That is not sales skill. It is strategic information asymmetry. The metric does not distinguish between the two.
Then there is the handoff problem. When the process breaks between BDC and sales floor, between sales and F&I, between online pricing and desk numbers, every failure costs deals. The metric attributes those losses to the closer, not the process. The salesperson absorbs accountability for systemic dysfunction they did not create and cannot control.
Once closing percentage becomes a target, the behavioral distortion follows. Salespeople begin avoiding deals that would lower it: credit-challenged customers, out-of-state buyers, cash deals with no F&I upside. The metric improves. Store health declines. Selection bias masquerades as performance improvement.
Incentive Misalignment at the Pay Plan Level
The most instructive examples of Goodhart's Law in dealerships are not digital. They are structural, embedded in compensation design.
I have watched this dynamic play out firsthand more than once. The pattern is consistent: management identifies a real operational problem, designs a pay plan adjustment to solve it, and inadvertently creates an incentive for salespeople to optimize their personal compensation in ways that damage the store.
Here is a version that is common enough to be instructive. A high-volume store prices new cars aggressively, losing front-end gross by design. Salespeople are compensated on trade snap: the difference between the appraised ACV and the value given to the customer. This is, effectively, the only lever for front-end gross on a new car deal.
Salespeople raise a legitimate concern: the desk is appraising trades too low, and customers are walking in with KBB Instant Cash Offers two to three thousand dollars higher. Deals are being lost. Management responds by allowing salespeople to match the ICO value as the ACV floor. The logic is straightforward. Match the market. Keep the customer. Close the deal.
Here is what actually happens. A salesperson works a deal and closes the customer on the desk's ACV of $25,000. Instead of writing the deal, the salesperson walks the customer through a KBB Instant Cash Offer. The offer comes back at $27,500. The salesperson applies that as the new ACV, captures the $2,500 delta as trade snap commission, and submits the deal.
The dealership loses the trade to KBB's fulfillment network. The used car department loses a retail unit it would have reconditioned and sold at margin. The store pays out higher commission on a deal that simultaneously drained its most profitable pipeline. Management begins asking why trades are leaving the building at an accelerating rate. The answer is in their own pay plan.
No one acted irrationally. Every person in the chain optimized for exactly what the structure incentivized. The damage was entirely predictable to anyone who modeled past the first-order effect. Few did.
This is not an edge case. This is how most pay plan failures operate. The behavior is rational. The outcome is corrosive. And the gap between the two is where Goodhart's Law lives.
Attribution Theater
Now layer vendor accountability on top of this.
The attribution problem in automotive is not limited to one channel. It is structural across nearly every vendor relationship in the store. The common thread: the entity being evaluated controls the measurement, and the measurement is designed to be favorable.
Database marketing vendors provide a clean illustration. Most default to a "multi-touch" attribution model, which in practice means: if the customer existed in your database and received the campaign piece, and that customer later purchased a vehicle within the match window, the vendor counts it as an attributed sale.
The logic sounds defensible until you examine what it captures.
Employee deals get counted. The sales manager who bought a truck was in the database. He received the mailer. He purchased within the window. Attributed. The vendor had zero influence on that transaction, but it appears in the ROI report as a sale their campaign generated.
Customers who never engaged with the piece get counted. A direct mail campaign drops 20,000 pieces. A percentage of those recipients were already actively shopping, already had appointments, or already had deals in progress. They bought cars. They happened to share a name and address with someone on the mail list. Attributed. The report does not distinguish between a customer who was influenced by the campaign and a customer whose purchase was already inevitable.
Customers who received the piece after they purchased get counted. Match windows are generous by design, and timing validation is rarely enforced. If the mailer lands three days after the customer signed paperwork, but the sale date falls within the attribution window, it appears in the report.
But the problem extends well beyond database marketing. Vendors across every category in the store operate with the same structural incentive: define success in terms that are easy to claim and difficult to disprove. An SEM vendor who cannot demonstrate conversions will pivot to impressions and visibility. A social media vendor will point to engagement metrics that have no proven link to purchase behavior. An inventory merchandising tool will claim influence on time-to-sale without controlling for price, market conditions, or seasonality.
None of this is necessarily dishonest. It is the predictable result of vendor relationships where the success criteria are not defined with precision at the outset, and where the vendor is permitted to select the metrics by which they are judged. Once a vendor is allowed to grade their own homework, the grade will always be passing.
The responsibility sits with the dealer. Before signing a vendor contract, the deliverable needs to be defined in concrete, verifiable terms. Not impressions. Not attributed sales using the vendor's own matching logic. A specific, measurable outcome that both parties agree represents actual performance, tracked against data the dealer controls or can independently verify.
Most GMs do not speak the language of attribution methodology well enough to challenge vendor reports on technical grounds. They should not have to. But in the absence of pre-defined success criteria and independent measurement, marketing budget decisions across the industry are being made on data that conflates correlation with causation by design. This is Goodhart's Law applied to the vendor relationship itself: the metric the vendor is evaluated on has been constructed to always look good.
This is the measurement environment into which dealers are now deploying AI tools. Attribution is already unreliable. Success criteria are already vendor-defined. The vendors selling AI-powered marketing will report on their own effectiveness using these same permissive models.
AI Does Not Fix Measurement Failure. It Compounds It.
This is the part the NADA stage is not telling you.
Every AI vendor at the show had a compelling pitch. Conversational AI for your BDC. Predictive analytics for inventory. Automated follow-up that never sleeps. The technology is often genuinely capable. The demos are not fabricated. The problem is what happens when capable technology is deployed into an environment where the objectives are wrong, the metrics are gamed, and the feedback loops are broken.
AI is an optimization engine. It will optimize whatever objective it is given. Point it at speed to lead and it will achieve sub-second response times while your customers receive replies so generic they disengage before a human ever enters the conversation. Point it at appointment volume and it will set more appointments, including appointments with customers who are not ready, not qualified, or were never going to show, because the system learned that the metric rewards volume, not quality.
This is Goodhart's Law at machine speed. The same distortion that takes months to materialize with human reps takes weeks with AI. The difference is that AI will not question the objective. A human rep might notice that the speed target is producing bad conversations. An AI system will execute the flawed objective with perfect consistency and zero friction.
The better vendors will acknowledge, if pressed, that implementation requires "process alignment" and "clean data" as prerequisites. What they will not say, because it would compromise the sale, is that most dealerships have neither, and achieving them is a harder problem than deploying the technology itself.
If You Are Serious About AI
The pattern across every example in this piece is identical. A reasonable insight gets compressed into a metric. The metric becomes a target. People optimize the target. The original insight disappears. The dashboard says things are improving.
Speed to lead. Closing percentage. Compensation structures. Vendor attribution. BDC appointment counts. In every case, the map diverged from the territory, and the store kept navigating by the map.
AI does not change this dynamic. AI is the most powerful map-following tool ever built. Point it at the right map and it will outperform any human team. Point it at the wrong map and it will drive you off a cliff with a clean dashboard the entire way down.
The stores that will extract real value from AI in the next two years will not be the fastest adopters. They will be the ones that audit their metrics against actual outcomes before they automate anything. The ones that separate vendor reporting from vendor evaluation. The ones that trace their pay plans to second-order behavioral effects. The ones that fix the data layer connecting their CRM, DMS, and marketing stack before they build on top of it.
That is not a compelling NADA keynote. It does not fit on a vendor's feature slide. But it is the difference between deploying AI as a force multiplier and deploying it as an accelerant on a fire you have not yet identified.
AI will optimize whatever you tell it to optimize. The only question that matters is whether what you are measuring reflects what is actually happening in your store. For most dealers, right now, it does not.
First published on LinkedIn, February 26, 2026.