Cross-category · 2026

What dealership AI actually does

Every vendor in dealer tech either claims AI or soon will, which makes the claim useless on its own. So this guide does not ask who has AI. It asks what the AI does — eight concrete things, counted across every category, with the state of the evidence printed next to each number.

100% model-detected, not yet verified. Nothing here has been confirmed by a demo or by a dealer yet, and the guide says so on every figure rather than in a footnote.

336
vendors we've mapped
207
carry a named AI capability
6
capability domains with depth
0
verified by demo or dealer

01 · The frame

AI is a capability, not a category

A voice agent that books service appointments and a model that prices your used inventory have nothing in common except a word in the marketing. Grouping them together destroys the only thing a dealer wants from a comparison. So we grade AI the way we grade everything else: by what it does.

Eight domains cover what dealership AI concretely does today. They came out of the catalog rather than off a whiteboard — the vision-and-inspection domain exists because a cluster of vehicle condition and damage tools fit none of the original seven.

Every count on this page is against 336 vendors we have mapped in capability detail, not against the whole catalog. We have not examined the rest closely enough to say what they do or do not do, and counting them as absent would be a claim we cannot back.


02 · The eight domains

What the AI does, and how common each one is

Ordered by how many of the 336 vendors we've mapped carry each capability. A vendor counts once per domain no matter how many products it sells into it.

Prediction and scoring

131 of the 336 vendors we've mapped

Models that rank, price, or forecast — the output is a number that flows into an operational decision.

In practice: Inventory pricing, lead scoring, equity mining, vehicle valuation, lender routing, churn prediction.

Analytics and insights

54 of the 336 vendors we've mapped

Natural-language questions, summarization, and anomaly detection over your own business data.

In practice: "Ask your DMS" interfaces, automated reporting narratives, call-outcome analysis.

Personalization and targeting

48 of the 336 vendors we've mapped

Deciding who sees what — audience construction and offer selection, done by a model rather than a rule.

In practice: Audience building, offer selection, send-time optimization, dynamic website merchandising.

Conversation

46 of the 336 vendors we've mapped

Voice, chat, and SMS dialog with a customer, held by software rather than a person.

In practice: BDC voice agents, website chat, service scheduling by phone, after-hours answering.

Content generation

31 of the 336 vendors we've mapped

Creating text, image, or video output that a person would otherwise have written or produced.

In practice: Ad creative, VDP descriptions, email copy, review responses, SEO content.

Vision and inspection

22 of the 336 vendors we've mapped

Computer vision over vehicles, lots, and facilities.

In practice: Damage detection, condition reports, appraisal imaging, lot and camera analytics.

Two domains we are not yet willing to describe

Document intelligence and Workflow agents each rest on a single capability in the entire vocabulary. That is too thin to write a section about, so we have not written one. The thinness is itself the finding, and we would rather print it than pad it out.

  • Document intelligence Extracting and structuring what is inside a document, as opposed to storing or signing it. 3 of the 336 vendors we've mapped carry it.
  • Workflow agents Software that plans and executes multi-step work across systems on its own initiative. No vendor we've mapped carries a capability here that we can classify with confidence.

03 · The autonomy ladder

The question underneath every AI pitch

Knowing a product does conversation AI tells you nothing about whether it drafts a reply for a person to send or holds the conversation itself. That difference is the entire purchase. The ladder makes it a checkable question instead of an adjective.

L0
no AI — rules, templates, or human-backed service
L1
assistive — suggests, a human executes
L2
supervised action — acts, a human approves or can intervene
L3
delegated action — acts unsupervised inside a bounded remit
Not published from capability data. Requires verification.
L4
autonomous — sets its own sub-goals within an objective
Not published from capability data. Requires verification.

We will not put a product above L2 on the strength of what its feature list is called. A capability named “autonomous agent automation” is a marketing decision, not an observation of the software running unsupervised in a store.

Two consequences a dealer should know about. First, AI-assisted humans grade L1 — a managed chat service where a person types the replies is a staffing purchase, whichever way it is sold. Second, everything we can currently publish sits at the same rung, because capability data alone cannot separate the rest. Today the number is effectively a yes-or-no on whether there is AI we can name, not a rating along a scale, and we would rather say that than dress it up as one.

The three questions that settle the ladder in a demo: what does it decide without a person, what happens when it is wrong, and can you show me it doing that on a live store today.


04 · By category

Where AI has landed, and where it has not

Share of the vendors we've mapped in each category that carry at least one named AI capability. Read it as a map of coverage and vendor emphasis, not as a quality score — a high share means the field talks about AI, not that the AI is good.

CategoryWith named AIOf mappedDomains present
Marketing72895
Business Intelligence72856
Finance and Insurance48717
Inventory52696
Fixed Ops43666
Digital Retailing41545
Vehicle Media34446
BDC38426
CRM30385
Messaging31346
Facilities6323
DMS15275
Logistics11224
Website18215
Accounting9203

HR, Consulting, EV Support, Auction, Reputation, Parts are left out of this table on purpose. We have mapped fewer than 20 vendors in each, and a percentage built on a dozen vendors reads far more confident than it deserves to. Those categories get their AI section back when coverage catches up.


05 · How to read this

What we checked, and what we did not

  • Every denominator is the mapped set. 336 of 705 tracked vendors have their capabilities mapped in detail. The rest are not counted as lacking anything — we have not looked closely enough to say.
  • 100% model-detected, not yet verified. These figures come from reading capability data, not from watching software run. Vendor documentation, a recorded demo, and a dealer confirming it in production are three further rungs we have not climbed yet.
  • We do not name a best. This guide lists what capabilities exist and how common they are. It does not rank vendors, and no vendor can pay to appear or to move.
  • We do not call anyone’s AI fake. That judgement needs evidence about the product, and a feature list is evidence about the marketing. When a claim and the software disagree, it takes a person to say so.
  • Capability data refreshes weekly; this cut does not move with it yet. Generated 2026-07-20.
Back to the field guide library