AI that works where the work is
Every dealership has now seen a chatbot demo.
Far fewer have seen an AI answer a real workflow question with the manual page it came from — or turn dead parts stock into a campaign list a manager can sign off.
This month: what changed when we required every AI output in our prototypes to be grounded in data the dealer already owns — and confirmed by a person before it acts.
An assistant that cites the manual
This summer we are developing five Nebula AI use cases for a multi-brand dealer group in the Adriatic region — all five already at working-prototype state, running on real data.
The most visible one is an assistant that lives as a widget inside VEGA, the CRM salespeople already have open. Ask it »zakaj mi ne pusti shranit ponudbe?« (why won’t it let me save the offer?) and it answers from the full library of official manuals — with a citation on every sentence and the screenshots from the exact pages it cited.
Users phrase problems in everyday words and manuals answer in official register, so the question is first rewritten into the manual’s own language before hybrid keyword-plus-semantic retrieval runs. When the manuals don’t cover something, the assistant says so — inventing a button name would cost more trust than it saves time.
The same widget learns each employee’s role from the application’s own click log — nobody tells it »I sell new cars« — and serves micro-learning and quizzes from the manual passages for that role. Embedding into the live application is a single script tag, isolated so thoroughly that a before/after comparison of the host UI finds not one element moved.
Grounded, or there is no answer: every sentence carries the manual page it came from.
Selling from stock in plain language
The second prototype makes the entire current new-vehicle stock searchable the way a customer talks: »a family SUV under €30,000 with an automatic« — in Slovenian, Croatian or English.
An LLM compiles the sentence into a hard, auditable filter over the real factory equipment of every version, taken directly from the vehicles system, while a second, semantic channel catches what no option code can say: »a car for camping« finds the versions whose descriptions mention roof rails and a washable interior. A concierge model orders the final shortlist and gives one honest reason per car.
The design rule throughout: the LLM chooses from closed vocabularies and exact data, never from its imagination. If a filter matches nothing, the tool says so rather than improvising.
Pipelines that prepare, people who confirm
The least glamorous prototypes may be the most valuable: three pipelines over the dealer’s own data. Spare-parts activation ages the warehouse into slow, dead and obsolete buckets, checks which customer vehicles the parts actually fit — reconstructed from the dealer’s work-order history — and drafts the campaign. Golden Moment scores which customers are ready for their next car and matches them against own stock. New-vehicle push ranks the cars that most need to move and finds the right customers for each.
All three end the same way: a ranked list with reasons, behind a GDPR consent gate and a shared do-not-contact store, waiting for a manager to confirm. The real asset is not compute but the internal data quietly accumulated in the DMS and CRM for years.
The AI prepares; the person decides.
Nebula Perspective
Three industry developments that matter — and what they mean for automotive business systems.
Hallucinations are contained by system design, not by model choice. A July research paper argues that »zero hallucination« is a property of the system rather than the model: grounded generation over approved content, evidence-based verification of every claim, calibrated abstention and full traceability. That matches the recipe our assistant prototype follows — citations beat eloquence, because in a dealership a wrong answer is a mis-sold car or a compliance breach.
The EU gave AI more time — not a reason to wait. The AI Omnibus was published in the Official Journal on 24 July 2026: obligations for high-risk AI systems move to December 2027 (Annex III) and August 2028 (Annex I), while transparency duties arrive sooner. Our view from the July issue stands: deterministic, auditable flows with human confirmation are the cheapest compliance strategy an automotive business can buy — build them now, comply calmly later.
Dealers use AI; value comes from embedding it. Cox Automotive’s new AI in Auto Retail Tracker finds that 82% of dealers already use AI, yet only 22% of users report actual revenue growth — and dealers working with specialised partners do markedly better. Consistent with what we see: returns appear when AI is embedded into the CRM and back-office processes the team already lives in, with outcomes a manager can measure and confirm.
Cox Automotive: AI in Auto Retail Tracker (11 August 2026) →
A closing thought
The dealership does not need another tool.
It needs the tools it already has to start thinking along.
Grounded in the dealer’s own data, confirmed by the dealer’s own people — that is AI that works where the work is.