The service-to-sales conversion gap is not a training problem. It is a mechanism design problem, and the work that matters is proving a fix before a dealer stakes a compensation plan on it.
The signal is already there
Service departments generate one of the most valuable sales signals a dealership has. Predictive maintenance systems can now flag which individual customers are approaching a major repair event in the next 30 to 90 days, the precise window in which a customer makes a buy-or-keep decision. A driver staring at a $4,000 transmission estimate is already deciding whether to repair or replace, and a lease payment presented at that moment is a fundamentally different offer than the same one six months earlier.
The signal is captured, enriched, and delivered to the advisor. Yet conversion from service interaction to vehicle sale sits far below what the propensity models predict. The value has already been manufactured. It is not being collected.
The conventional fix is aimed at the wrong target
The standard response is to fix the advisor: more training, sharper referral scripts, tighter cultural alignment between service and sales. The short-term gains are real, and the long-term gains evaporate, usually within 12 to 18 months of the initiative ending.
That pattern is not a coincidence, and it is not an execution failure. It is the signature of a problem that training was never built to touch. This is not a training problem. It is a mechanism design problem.
What the advisor is actually being paid to do
Look at how a service advisor is compensated. Repair order volume is measured, attributed, and rewarded. CSI scores are measured and reflected in pay. Referrals to the sales floor for a high-repair-cost customer are not measured, not attributed, and not rewarded. Worse, a referral that converts often strips the repair order out of the advisor's commission base entirely.
Under any rational set of preferences, the advisor's optimal move is to keep the repair order. The advisor is not making a mistake. Given the rules of the game, they are playing it correctly.
The handoff is a stable equilibrium
Economists have a precise name for this. It is a multitask principal-agent problem, described in Holmström and Milgrom (1991), in which an agent's effort is split across activities and the compensation structure rewards only some of them. The handoff is also a Nash equilibrium: no advisor can improve their payoff by unilaterally referring more, because the sales side has no contractual mechanism to share credit back. Both sides stay in the lower-payoff cell.
This is why the diagnosis matters more than the symptom. You do not coach your way out of an equilibrium. The only thing that moves it is changing the payoff structure of the game itself, which is exactly what training programs never do.
The mechanism: paying for contribution, not position
The mechanism design literature points to a candidate. Shapley value allocation, a cooperative game theory framework introduced by Lloyd Shapley in 1953 and now widely used in modern AI attribution, pays each contributor according to their causal contribution to an outcome, not their presence, seniority, or place on the org chart. Applied to the service-to-sales chain, it allocates a share of deal margin across the predictive model, the service advisor, and the sales representative based on each one's marginal contribution to the sale.
That is the candidate. The harder and more important question is the one most firms skip: how do you know it works before you rewrite a compensation plan that touches thousands of advisors and billions in annual transactions, and how do you know it will not trigger the gaming behavior that has undermined every prior attempt at outcome-linked pay in adjacent industries?
Proving the mechanism before it touches a paycheck
Answering that question is where the real work lives, and it is what multi-agent simulation is built to do.
The framework models the dealership as a population of independent decision-making agents (service advisors, sales representatives, and customers), each governed by statistical distributions for risk tolerance, cooperativeness, referral thresholds, purchase propensity, and price sensitivity. Large language models generate a behavioral profile for each agent on top of those distributions, so the quantitative outputs can be validated against decisions a real advisor or customer might plausibly make.
Then the mechanism is put under load: tens of thousands of synthetic customer interactions per compensation design, Monte Carlo across hundreds of runs, measuring individual earnings, departmental revenue, dealership-level gross profit, and the stability of the cooperative equilibrium over time. The result is evidence that can be falsified, produced before anyone touches a live paycheck.
A framework, not a one-off study
The architecture is built to extend. It is modular: every agent type, scenario variable, and mechanism is an independent component, reconfigurable without rebuilding the system. It is API-orchestrated, communicating through documented interfaces that plug into an existing data stack without dictating its shape. And it is cloud-native and headless, running wherever the compute lives.
That design is the point. The same framework that models one OEM's service-lane decision reconfigures to a parts-and-service compensation question, an F&I incentive question, or a manufacturer-to-distributor pricing question. The simulation is the asset. Any single question is just its first use.
The takeaway
Across industries, incentive misalignment gets treated as a problem of training, culture, and exhortation. Those interventions fail predictably, because the problem underneath is mathematical, not behavioral. Solving it requires redesigning the payoff structure of the game and proving the redesign works before deploying it at scale.
Multi-agent simulation does both. It plugs into any architecture, produces evidence you can falsify, and gives an OEM, a dealer group, or a marketing leader enough grounded confidence to act, because the mechanism has already been tested against tens of thousands of synthetic interactions before a single real one is at stake.