Four thin days add up to one pattern: the groups that own their records and capacity are pulling ahead of the ones renting agents.
Ben Sporn, chief marketing officer of VIP Auto Group (10 rooftops, 15 brands, Long Island plus a New Jersey Chevrolet store), says "We were spending half a million dollars easily on software that we've replaced internally" (WardsAuto). He lists an AI drafting tool he says tripled CSR productivity with staff reviewing drafts, service-record PDFs that save the service manager two hours a day, and automated replies to five-star reviews; the article gives no dollar figure for the revenue growth he calls "explosive" (WardsAuto). For scale, Cox Automotive's tracker says 82% of dealers use AI, only 22% have seen revenue or sales growth against 69% who expected it, and about one-third are not measuring impact or lack a clear method (Cox Automotive).
The constraint on a build like this is not the model, it is who can maintain it when the builder leaves. Review replies and record PDFs are low-risk jobs where a group owning the code and the logs is a real advantage; replacing a system that writes to the DMS is a different bet. Cox's own data cuts the other way too: dealers with external AI partners report better odds of beating expectations (+14 points), so the lesson is to know which jobs to build and which to rent, not to build everything (Cox Automotive). The gross lines touched are parts and service retention, and the software saving is a cost line, not gross.
Ask each vendor line over a set amount what it does that your own team could build and own, and name who maintains a build when the builder leaves.
Pick one low-risk back-office job (declined-service follow-up drafts, review replies) and run it as draft-for-approval for two weeks with a counted result.
For any internal build, require version control, a named second maintainer, and a log you can export; a key-person dependency is the main risk.
Book the software saving only after the vendor contract has ended, and record before-and-after labor by role so the claim survives a buyer's diligence.
Alaska's attorney general says Payless Used Car Sales will pay over $500,000 in restitution, about $100 each to more than 5,000 customers, over a fee of generally $499 per vehicle charged on top of advertised prices since November 4, 2018 (Alaska Department of Law). Payless has 280 days from court approval to contact customers and issue checks, and the acting attorney general says she accepted less than full restitution to preserve jobs (Alaska Department of Law). The same office's archive lists a Swickard settlement on March 24, 2026 and a Lithia Motors settlement titled "Fees Not Included in Advertised Price" on December 29, 2025 (AG archive); Swickard paid $800,000 now with $200,000 suspended for three years (Alaska Department of Law). California's CARS Act took effect October 1 with Total Price required in the first written communication that references a vehicle (California DMV), and the FTC's March 13 letters told 97 dealer groups that advertised prices must be the total price including mandatory fees (FTC). None of these sources mention AI.
The per-customer number is small and the exposure is the time span: the violation period runs back to 2018, because a subpoena dispute paused the clock. Whether a bot's reply counts as an advertised price is my inference, not a stated rule, but the wording of each source turns on the price a customer sees, not who typed it. The gross at stake is doc-fee income booked against new and used vehicle gross, and the cost of getting it wrong is a restitution order measured in years of sales.
Treat advertised price as a control with an owner. Decide whether you run one fee-in-price standard in every state rather than carving out California and Alaska.
Compare the advertised price with the buyer's order out-the-door total on ten recent deals per store, by channel, and fix any gap before a regulator finds it.
Confirm that listing feeds, website pricing, chat and AI replies all read the same fee schedule, and that the fee table has an owner and a change log.
Estimate the restitution exposure by multiplying any fee that is not in advertised price by units sold per year; the Payless period was about eight years.
Impel's Voice AI answers inbound service calls, identifies the caller through the DMS before the call connects, books appointments and escalates the rest with context; Impel says about 40% of inbound service calls are resolved without reaching an advisor in pilots, and names no customer or price (Impel). DriveCentric's Acquire Agent engages customers right after a trade valuation request, runs inside its own CRM, does not calculate valuations, and has early access on October 15 and general availability on December 1, with no customers, pricing or metrics disclosed (CBT News).
Both are workflow products where the vendor holds the conversation, which is the part a dealer should ask to own. Impel's 40% measures calls kept away from an advisor, not appointments kept or repair orders written, and the figure comes from the vendor. A booked appointment only becomes fixed-ops gross if there is a technician hour to sell (story 6). The trade-lead agent touches used vehicle gross through acquisition, but with no result published there is nothing yet to hold the vendor to.
Make the next renewal conversation about a named pilot result with a baseline, and about who owns the call and message record if you leave.
Test the phone agent at one store against open technician hours, and count shown appointments and repair orders written, not calls handled.
Confirm the DMS lookup is read-only, the phone-system integration works with your carrier, and every call, transcript and text is exportable.
Ask for pricing per shown appointment or per acquired unit rather than per conversation, and hold the vendor's 40% claim as an assumption until your own data confirms it.
In a Stephens and Surveyplanet Q3 dealer survey as reported by Car Dealership Guy, F&I sentiment is +34 net positive, parts and service +22, new-vehicle gross -29 and traffic -38, and 46% of dealers called September their worst recent month for new-vehicle gross (Car Dealership Guy); the survey itself was not retrieved. The same piece describes Kia of Port Charlotte (Morgan Auto Group) running a 3 p.m. to 11 p.m. used-car recon shift so daytime technicians do customer-pay work, a service-to-sales manager with salespeople on the service drive, and advisors who are not penalized when a customer trades instead of repairing, and being on pace for over 50 service-sourced and incremental vehicles a month (Car Dealership Guy). Edmunds' Q3 data show the average new-car amount financed at $44,664, the payment at $787 and 25.5% of buyers at 84 months or longer, with APR flat at 7.0% (Auto Remarketing); Experian says H1 subprime originations rose 24% year over year while 60-day delinquencies dipped each month (Auto Remarketing).
The money is moving to the back of the deal and to the shop, and both run on lenders and technicians, not on lead volume. Longer terms and more subprime volume keep F&I busy and raise the stakes for any tool that touches a credit or product decision. The Port Charlotte detail matters more than the sentiment numbers: the advisor pay-plan rule is a decision, not a purchase, and it is what makes a service-to-sales program work. The store's results are its own, reported secondhand.
Report fixed ops and F&I per copy ahead of new-vehicle gross in the scorecard, and fix the advisor credit rule before buying any service-to-sales tool.
Move used recon off the daytime customer-pay bays with a second shift if you have the volume, and write the advisor credit rule down.
Make sure F&I and credit-decision tools keep a record of inputs and the human decision; longer terms and subprime growth raise the cost of a gap.
Track F&I per copy and effective labor rate against new-vehicle gross monthly, and stress-test the plan against a further decline in new-vehicle gross.
PayNearMe announced its AI Servicing and Collections Agent on October 1; it authenticates customers, takes payments, captures promises to pay and manages opt-outs across voice, text and web (Auto Remarketing). At Indiana Finance Company, the one named pilot, the agent placed as many calls in an hour as about 28 support staff, and 90% of targeted accounts got a voicemail, a live conversation or a transfer to staff (Auto Remarketing). The figures are vendor-reported, they measure call volume rather than dollars collected or complaints, and the vendor's own release was not retrieved.
A collections call is a regulated act, so the number to ask for is not calls placed but dollars collected per complaint. This is mainly a lender and buy-here pay-here story; for a franchised store it matters because the contracts you sell end up in collections, and the dealer's name travels with them. The F&I line is the one touched, through contract performance and chargeback exposure. What the agent says when a customer disputes a debt, and who keeps the transcript, is the whole control question.
If you hold paper in-house, ask who is accountable for what the agent says to a customer about a debt, and whether you or the vendor keeps the record.
Not your lane unless you run in-house finance; if you do, require a human handoff for any dispute, hardship request or complaint.
Confirm consent state, call recording and opt-out handling are in the audit record, and that transcripts can be exported on request.
Measure dollars collected per complaint and per 100 accounts worked against your current collections baseline, not dial volume.
Speaking at Ford's ACCELERATE forum on Wednesday, September 30, Jim Farley cited an Alliance for America's Skilled Trades report finding that training programs produce 55 skilled-trades workers for every 100 needed, with nearly one-quarter of the skilled workforce 55 or older and only half of trades enrollees completing training (CBT News). The same article puts Ford pickup dealer inventory at roughly 40 days after the F-150 supplier disruption, against a 50 to 60 day target (CBT News). The report covers skilled trades broadly, not automotive technicians alone.
Every booking and recapture tool in this edition adds demand to a fixed number of technician hours. If the shop is already full, a phone agent that books more calls moves work around and changes the customer's wait, not the gross. The store with a second recon shift in story 4 is doing the unglamorous thing that raises the ceiling. I am inferring the technician link from a cross-trades report, so treat the 55-per-100 figure as direction, not as an automotive-specific measure.
Count technician capacity as a strategic asset: pay-plan, training and second-shift decisions come before another demand tool.
Pull open flat-rate hours for the next ten days at each store and compare them with booked hours before turning on any booking agent.
Feed live technician capacity into the scheduler so an agent cannot book against hours that do not exist.
Track effective labor rate and hours per repair order alongside bookings, so a rise in appointments is not mistaken for a rise in gross.
AutoNation named Ravi Simhambhatla chief technology and AI officer effective October 1; he previously held digital and technology roles at Avis Budget Group, Google Cloud and Tesla, and CEO Mike Manley says technology, data science and AI let the group "get significantly more from that foundation" (Auto Remarketing). Set that beside VIP Auto Group, where a marketing executive builds the tools directly (WardsAuto). Both move decision rights over AI up and in: the group, not the rooftop or the vendor, decides what gets built and who holds the log. Ford is moving the other way with commercial certified used vehicles: Ford Pro Advantage launches October 6 through more than 700 Commercial Vehicle Centers, with sourcing set to prioritize vehicles in higher demand (WardsAuto). That is an OEM-defined used program run through dealers, a fulfillment role unless the dealer keeps the customer data.
None of the three agent launches in this edition disclose a price: Impel Voice AI names none (Impel), DriveCentric's Acquire Agent has early access on October 15 and general availability on December 1 with no pricing (CBT News), and PayNearMe's agent is available to clients with no pricing stated (Auto Remarketing). When a price is missing, the unit is the negotiation: ask to pay per shown appointment, per acquired unit or per dollar collected, not per conversation. The honest counterweight is that dealers with outside AI partners report better outcomes than those without, so the answer is not to refuse vendors (Cox Automotive). Put an exportable action log and the right to take your conversation history with you into any renewal.
Governance, audit and security, over the data layer
For a service-phone agent like Impel's, a read-only DMS caller lookup, a scheduler that shows real open technician hours, a phone-system integration your carrier supports, consent for text follow-up, and an export of every call, transcript and text with timestamps (Impel). A price table with one owner that every channel reads, so an agent cannot quote a number your fee schedule would not.
An agent books against hours that do not exist, or states a price your out-the-door total does not match, and the first sign is a complaint.
Can the vendor export a full month of calls and messages, with timestamps, without a support ticket?
The Alaska Department of Law's September 22 release (Alaska Department of Law), California's CARS Act effective October 1 with two years of records for ads, initial communications, pricing disclosures, add-on consents and cancellation notices (California DMV), and the FTC's March 13 warning letters to 97 dealer groups (FTC) all turn on the price the customer sees. The DMV page was fetched today and does not mention AI or automated messages; whether a bot's reply is covered is for counsel. Separately, the Connected Vehicle Security Act cleared Senate Commerce unanimously on July 22 and still needs a floor vote and House action before December 31, according to Senator Moreno's comments as reported (CBT News); status not re-verified at the primary source.
The fee schedule version, the price shown in each channel, each outbound first message with timestamp, and the out-the-door total on the buyer's order.
Does an automated reply count as an advertisement or initial written communication in each state you sell in, and who signs off the fee schedule?
Find declined customer-pay work on past repair orders by age, safety relevance and estimated value, then offer the customer a return visit with a specific reason and an open slot. VIP Auto Group says its service-record tool tracks declined services and recalls and saves the service manager two hours a day (WardsAuto).
The service manager approves the list and the offer wording; the advisor makes the contact. Prohibited: scare language on safety items, repeat contact after an opt-out, and any slot not shown as open.
Illustration — assumption set (not a finding). At 1.5 labor hours each (assumed) that is 24 technician hours, which must exist (story 6). The most sensitive input is the return rate.
Leading: contacts made and appointments booked per 100 eligible lines. Lagging: jobs completed and gross against a matched prior period at the same store.
Andreessen Horowitz published "Who Gets Paid When AI Does the Shopping?" on September 29; it argues that platforms defending advertising revenue will resist AI agents while those earning transaction fees will welcome them, citing Amazon's 2025 advertising revenue of $69 billion against $34 billion in operating income excluding AWS (Andreessen Horowitz). It does not mention automotive. An older Andreessen Horowitz piece, August 10 and outside this window, puts computer-use agents at roughly $6 to $8 an hour of inference against about $30 to $45 for fully loaded US back-office labor (Andreessen Horowitz).
Dealer read-through: My inference, not Andreessen Horowitz's: the lead vendors and marketplaces you pay differ in whether they earn from ads or from transactions, and that decides whether an AI shopper ever reaches your inventory. Ask each one which it is. The labor figure, if it holds, is the economics behind builds like VIP Auto Group's.
Before another booking tool goes live, pull open flat-rate technician hours for the next ten days at each store and compare them with booked hours.
The owner is the fixed-ops director at each store. Set the agent's booking cap to the open hours it can see, and count shown appointments and repair orders written, not calls handled. Stop condition: if booked hours exceed open hours two weeks running, hold agent volume where it is until you add a shift or a technician.
I am dating this call: I expect the phone agents to look good in pilots and then stall at the shop, and I am wrong if a named dealer publishes a baselined repair-order gain from one in the next 90 days.
Cox says 22% of dealers have seen revenue or sales growth from AI while 69% expected it (Cox Automotive). If your store is in the 78% that has not, which line of your P&L would you point to as evidence, and who signed off the baseline it was measured against?