Some agentic AI is genuinely ready for a dealership today, and the operators taking it will be ahead. But almost every vendor argument rests on the work happening inside one platform that holds the record — and virtually no group operates that business. What's ready, what isn't, and why the estate rather than the technology will set the pace.
There is a category of dealership work that agentic AI handles well today. It has a recognisable shape: high volume, clean boundaries, little judgement, and a failure mode you can reverse.
A meaningful share of every support desk's traffic is the same handful of how-to questions, asked by someone who needs the answer in the next four minutes because there's a customer at the desk. Today that becomes a ticket, and the reply arrives when it arrives. An agent sitting on the documentation answers immediately, decides what it can't answer, and hands the rest to a human — carrying the context across, so the person doesn't start the conversation again from the beginning.
That last part is the whole design. An escalation that forces the user to re-explain has moved the delay, not removed it.
The caveat is one every operator will recognise. The person at the desk doesn't know which supplier owns their problem. They know the transaction won't go through. Whether that's the DMS, the finance platform, the manufacturer's portal or the interface between two of them is precisely what they can't tell, and it's the question they're really asking. An agent that knows one supplier's documentation answers a third of the traffic well and routes the rest with confidence it hasn't earned.
Managers do not lack data. They lack the twenty minutes and the report-writing skill to get an answer out of it, so they run on the report someone built two years ago for a different question. Being able to ask why used gross is down in one branch and get an answer in the moment, without an analyst in the chain, changes what gets asked — and therefore what gets caught.
It also raises a question the industry has never settled: which number wins. An agent answering fluently from one source doesn't resolve the discrepancy between the DMS and the manufacturer's dashboard. It industrialises it.
The enquiry that arrives outside hours currently waits, and a proportion of it goes elsewhere. An agent that qualifies it, answers the sensible questions, prices the part-exchange and books the appointment is doing work that wasn't being done at all. The failure mode is a booking you have to unpick, which is annoying and recoverable.
Note what that single journey crosses, though. A website built by one supplier, a lead management system that may be specified by the manufacturer, and stock and pricing that live somewhere else again. Three vendors before anything is confirmed.
Almost every vendor argument for agentic AI rests, quietly, on the work happening inside one platform that holds the record. It's a good argument. It just describes a business that very few groups actually operate.
The real estate looks like this. More than one DMS, because acquisitions came with their own and the migration was deferred. Manufacturer-mandated systems for lead management, warranty, parts ordering and service booking, specified per marque and not up for discussion. A separate website provider, a separate valuation tool, a separate finance proposal route, a separate service plan administrator, a separate video platform for the workshop. A multi-franchise site running the same process two different ways because two brands instructed it differently, and a service advisor who switches between both before ten in the morning.
Nobody is going to consolidate that in time to matter. So the useful question isn't which platform to standardise on. It's this: for each action the agent takes, which system is the system of record — and does the agent write back to it?
That question travels across any estate, and it does something else useful. It separates the tools that act on your business from inside it from the ones that act on a copy and reconcile later. The first is auditable. The second is a reconciliation problem you'll discover in three months, which is roughly how long it takes for the two records to drift far enough for someone to notice.
Automate the documented process and you get the documented process. What you lose is the judgement — the part that was actually doing the work.Ground Truth Advisory
A process on paper and a process in practice are different objects. The gap between them is held together by people, and the holding is invisible because it works.
Take any point in a dealership where demand arrives faster than it can be handled — reception, the parts window, service reception, the sales desk on a Saturday. The person standing there has no authority to prioritise. They have no view of what the workshop can absorb, no visibility of which customer has already been let down twice, no mandate to tell anyone no. And yet the day works, because that person makes a hundred small judgements an hour about what goes first and what waits, they make them well, and nobody has ever written any of it down.
The same is true of the compensations. The service advisor who reorders the morning because they know which technician is slow on that job. The sales administrator who catches the deal that doesn't add up before it reaches audit. The controller who knows which supplier's next-day is really next-day. None of it is in the system. All of it is load-bearing.
If a process is sound and slow, automation is a gift. If a process is unsound, automation is an amplifier. The failures don't shrink. They get faster, more consistent, and harder to notice, because the thing that used to surface a bad process — a person visibly struggling with it — has been removed.
A two percent error rate on forty transactions a day is something a manager sees. The same rate on four hundred is a pattern in the accounts three months later.
So the first readiness question isn't about the technology. It's whether you know what the process actually is. Not the version in the induction pack — the version that runs on a Tuesday when someone's off. If the honest answer is no, the work in front of you is process work, and it always was. Nobody has automated their way out of not knowing how their own business runs.
There's a framework worth borrowing here — the 4Ds, from Anthropic's AI Fluency work: Delegation, Description, Discernment, Diligence. Decide what to hand over. Brief it clearly. Judge what comes back. Stay accountable for the result.
I teach it and it holds up. But it was built for a working pattern where a human briefs a task and reads the output, and that pattern is exactly what agentic work removes. The point of an agent is that it takes many steps without you reading each one. Discernment assumed you'd see the output. Diligence assumed you knew what had been done in your name.
So the question stops being how do I check the output and becomes where do I put the checkpoint.
The answer isn't to check more. It's to check at consequence. Reversible and cheap — a draft written, a record read, a suggestion queued — let it run. Money, the customer, or the audit trail — a price quoted, a booking confirmed, a part ordered, a document issued — that's where a human sits, with enough context to make a real decision rather than clicking through.
A checkpoint everyone clicks through is not a control. It's a liability with a timestamp. That design is made once, deliberately, by someone who understands the operation — and in a mixed estate it has to be made per boundary, because the point where an agent hands off between two suppliers is the point where nobody is watching.
When an agent contacts, quotes or handles a query, someone has to have decided whether that's disclosed and how. There's a defensible position either way for routine work. There is none for we never thought about it — particularly in a complaint, and particularly where regulated activity is anywhere near the transaction.
Something will go wrong. You'll need to show what was done, in what order, on what instruction, with what data, and who approved the parts that were approved. Human error leaves a trail of people who remember. Agent error leaves logs, or it leaves nothing.
In a single-platform business that's a straightforward requirement. In a real one it's the hard part, because the trail crosses suppliers and each holds its own fragment in its own format for its own retention period. Reconstructing an incident then means assembling evidence from three companies, at least one of which has no contractual obligation to help you quickly. That is worth knowing before you need it, not during.
Which is the point underneath all of it. Delegating the work does not delegate the responsibility. It fails against your name, your FCA position, your brand standard and your customer. The agent has no professional standing to lose, and neither does the integration.
If you can't answer the first three, the rest is premature.
And let me speak to the operator. Not the pilot. Not the reference the account manager chose. Someone running it now, twelve months in, who'll tell me what broke.
The correct answer is that it stops and asks, and hands over without losing the thread. If the answer is that it makes a best effort, you've bought a system that will confidently do the wrong thing at scale, and you'll hear about it from a customer.
If reconstructing an incident depends on the vendor's goodwill and availability, you don't have an audit trail. You have a support ticket.
Every operator has lived the version of this where both parties are certain it's the other one, and the dealership carries the customer in the meantime. Agentic work multiplies those boundaries.
Get the answer in writing, or accept that the answer is you.
Agentic AI is landing in retail and some of it is already earning its place. The ready-now work — triage, querying, out-of-hours enquiry handling — is worth taking now, and it's the safest possible way to learn what these systems do before pointing them at anything that matters.
But the pace will be set by the estate rather than the technology. Groups running one platform end to end will move quickly. Everyone else — which is almost everyone — will move at the speed of their integrations, their manufacturer mandates and their willingness to name an owner for a process that has never had one. That isn't a failure of ambition. It's the actual constraint, and any plan that doesn't price it in is a plan that slips.
The failures, when they come, won't be technology failures. They'll happen because the process wasn't owned, the checkpoint was in the wrong place, the record was in a system nobody wrote back to, and nobody could say afterwards what had happened. That's the same failure that has followed every system implementation in this industry for thirty years, and it will arrive again wearing better clothes.
Written from the retail side — thirty years in the trade, every layer from the workshop and used forecourt to Head of Business, and five agency models landed across five marques and more than 40 sites. Eleven Anthropic AI certifications, including the agent and subagent material. The 4Ds framework — Delegation, Description, Discernment, Diligence — comes from Anthropic's AI Fluency work and is taught through the free SMART AI guides. Ground Truth Advisory works with OEMs, suppliers and investors on structural change in automotive retail. The work is confidential and unattributable.
I write when there is a piece worth reading, not to a schedule. Leave your name and I'll send the next one when it goes out.