Why the next five years won't be about which DMS wins

Thesis

Two years ago we shipped our first prototype to four pilot stores. We thought we were building an F&I co-pilot. What we were actually building, it turned out, was the missing layer between the dealership and a legacy DMS that nobody (not the dealer, not the OEM, not the lender) had the appetite to replace.

This post is a synthesis of what we have learned. It is opinionated. It comes from twenty months of being on the floor at thirty-something rooftops, plus access to enough internal IT roadmaps to know how the DMS conversation actually plays inside multi-rooftop groups.

The thesis is simple: the next five years of automotive retail will not be defined by which DMS wins. They will be defined by what gets built around the DMS.

Modernization fails because it is priced as if the only failure mode is doing nothing. With computer-use AI, doing nothing has a new and very different price tag.

Three lessons from failed modernizations

Every dealer principal we have spoken to has at least one battle story about a modernization initiative that consumed money, calendar, and goodwill, and then quietly stalled. The patterns are remarkably consistent.

1. The migration is the project. The new system is the gift.

Most modernization budgets pencil out the new platform’s cost and feature set. Almost none honestly account for the migration. We saw one Tier-1 group spend $4.8M migrating fifteen rooftops to a new DMS over thirty months. The platform itself was $1.6M of that. Migration was the rest: re-keying, re-mapping, re-training, parallel running. By month thirty, six of the rooftops had partially rolled back.

2. The OEM is not on your side.

OEM-mandated systems exist because the OEM wants visibility into your business at a granularity you would not voluntarily provide. “Modernization” almost always means more telemetry to the OEM, not less. Many GMs we have spoken to have made peace with the trade-off, but they did not enter into it with eyes fully open.

3. The vendor’s incentive flips on day one.

DMS vendors price implementation aggressively. They make their margin on integration fees, on support, and on the inertia that comes from being entrenched. The day after go-live, your vendor’s incentive flips: now they want you to stay in their walled garden and pay for every integration. The savings projected at signing rarely survive contact with this reality.

Why agents change the calculus

Computer-use AI changes one specific assumption underlying every previous modernization argument: that the only way to get a modern interface, modern workflows, and modern data quality out of your DMS is to replace the DMS.

If a sufficiently capable agent can sit on top of the existing DMS UI (clicking, typing, validating, and writing back), then the cost of “modernization” is no longer the migration cost. It is the cost of the agent layer plus a much smaller integration tax. We are not the first to make this argument. We are, as of writing, one of a small number of teams who have shipped it to production rooftops.

This is not a claim that the DMS will never get replaced. It is a claim that replacement stops being the only path forward. For the first time in twenty years, dealer groups can buy themselves time and a modern surface in the same purchase order.

Anatomy of a wrap-and-replace

Here is the architecture we have arrived at, at a high enough level that we can publish it. (The interesting parts are below this layer; we will not be writing those up.)

1. Read layer. Computer-use agents log in to the DMS as a service account, scoped per rooftop, with a strictly auditable session. They navigate the UI, scrape structured data, and emit normalized records to our orchestration layer.
2. Normalization. We map every DMS-specific representation into a canonical model (deal jacket, customer, vehicle, transaction) and tag it with provenance. The agent that produced it. The session it ran in. The time it ran.
3. Orchestration. Workflows live above the canonical model, not above the DMS. F&I Co-Pilot, CarOne, and the dealer-facing dashboards all consume the canonical model.
4. Write layer. Where the workflow needs to write back into the DMS, we run a different class of agent: one that requires explicit human approval for any action with consumer-protection or fair-lending exposure.

The most important part of this architecture is the boundary between layers two and three. Below the boundary lives DMS-specific logic, which is messy, regression-prone, and constantly changing. Above the boundary lives our product, which can move at startup speed because it does not care which DMS it is talking to.

What to measure

If you are evaluating an agent layer, here are the metrics we have come to trust. They are unglamorous, which is the point.

• Read-side coverage: percentage of DMS workflows the agent reads from cleanly, week over week.
• Drift incidents: how often DMS UI changes break agent runs, and how quickly they recover.
• Time-to-rollback: how fast the agent can be cleanly disabled per rooftop, in case of an outage downstream.
• Human-review queue latency: how long an action sits awaiting human approval before it gets attended to.
• Auditable actions: percentage of agent actions for which a complete audit log can be reproduced from immutable storage.

Notice what is not on this list: “hours saved.” Hours saved is the outcome you want, not the metric you should measure. Read-side coverage and drift incidents predict hours saved, two and a half quarters out, much better than hours saved predicts itself.

A 90-day playbook

If you are a dealer principal or a CIO at a multi-rooftop group, and you have read this far, here is the playbook we recommend.

Days 1 to 30: Pick the wedge workflow.

Pick one workflow, on one DMS, at one rooftop. We always recommend starting with stipulation pre-flight on the F&I desk. It has clear ROI, it has consequences only the F&I director cares about (so the sales team won’t be politicized), and the data is clean enough to evaluate honestly.

Days 31 to 60: Run side-by-side.

Have the agent run alongside the F&I director, reading and recommending but not writing. Compare the agent’s recommendations to what the human eventually does. The gap is your training signal.

Days 61 to 90: Hand it the steering wheel, within scope.

Where the gap is small and the actions are reversible, let the agent take the keyboard. Keep human-in-the-loop for any action with consumer-protection exposure. Measure the metrics in the previous section. Decide, with data, whether to expand the wedge.

Objections, answered

We are usually asked the same five questions in the same order. Here are our short, honest answers.

What if the DMS UI changes?

It does. Continuously. Our drift-detection layer flags broken selectors within minutes, and we have a regression cohort that catches most changes before they reach a customer. We treat DMS UI changes the way a search engine treats web changes: a fact of life, automated to handle.

What if the DMS vendor cuts us off?

We operate as a service account the dealer provisions, on the dealer’s behalf. Cutting us off is, in practice, cutting the dealer off from their own data. Vendors are aware of how that plays politically.

What about security?

We have a SOC 2 Type II report, GLBA-aligned controls, per-tenant encryption keys, and a tamper-evident audit log. Our security page covers the rest. The standard is the standard.

What about hallucinations?

For consumer-facing or fair-lending workflows, the model never has the last word. A human approves. For internal workflows, we have an evaluation suite that catches the long tail before deploys.

What if the DMS vendor builds this themselves?

They might. We hope they do. The DMS vendors that succeed in the next decade will be the ones that lean into agents the way Stripe leaned into developers. The ones that don’t will continue to be the ones their customers complain about.

What this means for 2026

We expect three things to happen in the next twelve months.

1. Most modernization initiatives at large dealer groups will be quietly deprioritized in favor of agent layers, not because agents are better in every way, but because they ship in months, not years.
2. OEMs will publish their own agent strategies, mostly defensive. Some will be useful. Most will be telemetry plays in agent clothing.
3. The dealer groups that move first will compound a quiet operational advantage that will be very hard for late movers to close, in the same way that early CRM adopters compounded sales-process advantages a decade ago.

We are biased, we have a horse in this race, but we have also been on the floor with enough dealers to be confident in the read. Modernization is dead. Long live what comes next.

If you would like to talk about how this plays at your group, we’d love to hear from you. There is a very low-friction form on our contact page.