Do You Know Which AI Model Is Answering Your Phones?

Most dealership AI conversations stop at the vendor's name. A report published by CNBC on July 7, 2026 is a good reason to ask a second question: what is actually running underneath it?

CNBC found that the share of tokens used by U.S. companies on Chinese-built AI models through OpenRouter - a platform that routes developer traffic across many different models - has stayed above 30% every week since February 8, 2026, peaking at 46%. The average across the previous 12 months was 11%. In the first half of 2025, it was 4.5%.

Dealers do not buy tokens. They buy an AI BDC, an AI voice agent, a service scheduler. But each of those products is a layer of software wrapped around someone else's model - and that model is effectively a subcontractor you never interviewed, working under a contract you never signed, who can be swapped out in an afternoon without anyone at the store being told.

Myth vs. data

• The myth: which model a vendor runs is an IT detail, not a management issue.
• The data: Chinese-built models went from 4.5% of U.S. enterprise token usage on OpenRouter in the first half of 2025 to a weekly peak of 46% in 2026 (CNBC, July 7, 2026). The stack underneath your AI vendor can change faster than your ad budget.

Why your vendor's cost curve became your problem

The reason for the shift is not ideology. It is price.

Open-source Chinese models can be "60% to 90% cheaper" than the leading Anthropic and OpenAI models, Justin Summerville of OpenRouter told CNBC. The capability gap that used to justify the premium has narrowed: Brookings fellow Kyle Chan estimates the top Chinese models are running "six to nine months" behind the top U.S. frontier systems, and CNBC reports that Z.ai's GLM 5.2 landed within a percentage point of Anthropic's Opus 4.8 on one closely watched agentic benchmark at roughly a fifth of the cost.

When a task does not need the best model, engineering teams route it to the cheapest one that is good enough. That is exactly what is happening, and it is happening fast.

Now put that next to the pressure your vendors are under. The Q2 2026 Cox Automotive Dealer Sentiment Index, based on roughly 958 dealer responses, put the dealer cost index at 74 - its highest level in more than a year - while the profit index sat at 36. Dealers are squeezing every line item, including software. Vendors selling into that market compete on price, and the fastest way to cut price is to change the model.

None of this is a conspiracy. It is procurement. But procurement decisions made three layers away from your store now shape what your customers hear on the phone.

What actually changes when the model changes

Swapping a model is not like swapping a hosting provider. The model is not the plumbing - it is closer to the employee. Change it and several things move at once:

• Instruction-following on disclosures. Recording notices, consent language, "I am an AI assistant" statements. A model that follows your script 99% of the time and one that follows it 96% of the time look identical in a demo and very different across 4,000 calls a month.
• Escalation behavior. When does the agent stop and hand a live customer to a person? That threshold is a function of the model's judgment, not just your prompt.
• Hallucination profile. Quoting a payment, confirming a rebate, promising a VIN is on the lot. The failure modes differ model to model.
• Conversational mechanics. Latency, interruption handling, how it behaves when a caller talks over it. On a phone call, a 400ms difference is the difference between natural and robotic.
• Tone. Verbosity and warmth drift. Your BDC has a voice; a new model may not keep it.

The important nuance: this is not a story about Chinese models being unsafe. A vendor moving from one U.S. model to a cheaper U.S. model creates exactly the same drift. And an open-weight model designed abroad can be downloaded and run entirely on U.S. infrastructure under your vendor's control, which is a very different risk profile from an API call to an overseas endpoint. Both get described the same way in a sales meeting. The nationality is the headline. The lack of disclosure is the actual problem.

The rule you already live under

Dealers who arrange financing are financial institutions under the Gramm-Leach-Bliley Act, which puts them under the FTC's Safeguards Rule. Section 314.4(f) is the part that matters here: dealers must take reasonable steps to select service providers capable of maintaining appropriate safeguards for customer information, require those safeguards by contract, and periodically assess those providers based on the risk they present.

An AI phone agent touches customer names, phone numbers, addresses, vehicle and trade details, and buying intent. If you cannot name the model handling that conversation, or say what country it runs in, "periodically assess" is not a box you can honestly check.

This is not a reason to rip out AI. It is a reason to write the questions into the contract - the same way you already do with your DMS and CRM providers.

The oversight layer is arriving from the regulators first

The clearest signal so far is not American. In the UK, the FCA's Mills Review sets out how AI could reshape retail financial services. Reading it for the motor finance market, iVendi CEO James Tew told Motor Trader on July 9, 2026 that his interpretation is that a second layer of AI - ideally built on completely separate technology - will be required to supervise AI agents used in any regulated activity, with a real-time kill switch that hands the consumer to a human, independent scoring of every interaction, and alerts to a senior manager when something goes wrong.

That is a vendor's reading of a UK review, not U.S. law, and it should be weighed as such. But the direction of travel is worth noting, because the operating principle translates cleanly: record everything, score everything, escalate fast, and be able to prove what your agent said. U.S. dealers already carry their own version of that burden through the TCPA and state consumer-protection law - a subject we covered in TCPA and AI calling.

Eight questions worth asking your AI vendor this month

1. Which model or models power this product today? Name them, with versions.
2. Do you route across multiple models? On what basis - cost, latency, task type?
3. Where does inference run: your infrastructure, a U.S. cloud, or a third-party API - and in which country?
4. Will you notify us before you change the underlying model, and how many days ahead?
5. Is our customer data - transcripts, recordings, CRM records - used to train any model? Is that in the contract?
6. What is retained, for how long, and who can access it?
7. What triggers a handoff to a human, and how fast does it happen?
8. Can you show us scored transcripts of our own calls, plus a before-and-after benchmark the next time you change models?

If a vendor cannot answer the first question in one sentence, that is information. If they can, and the answer is a model you have never heard of, that is not automatically bad - ask where it runs and what changed in your call scores when they moved to it.

The upside nobody puts in the sales deck

Falling inference costs are, on balance, good for dealers. The same price collapse that is making vendors switch models should eventually reach your invoice as lower subscription costs, more included minutes, or capabilities that were priced out of reach a year ago. If your vendor's underlying costs fell by 60% to 90% and your monthly bill has not moved, that is not a scandal - but it is a conversation worth having at renewal.

The bottom line

Per Kerrigan Advisors' 2025 Dealer Survey of more than 525 dealers, 43% are already deploying AI in their operations and another 47% plan to, leaving just 10% with no plans at all. The industry has largely settled the question of whether to use AI. The unsettled question - the one this month's data drags into the open - is which AI, run by whom, on what terms, and with what notice when it changes.

You do not need to become an expert in model architectures. You need to be able to name what is answering your phones, and to know before it changes. That is a vendor-management question, and dealers have been good at vendor management for a long time.

If you want to see how we answer those eight questions - including which models we run, where, and what happens when we change one - that conversation is open at carbuki.com.

Sources

• CNBC, "Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge," July 7, 2026: cnbc.com
• Cox Automotive, "Q2 2026 Cox Automotive Dealer Sentiment Index," May 26, 2026 (approximately 958 dealer responses, surveyed April 21 - May 4, 2026): coxautoinc.com
• Kerrigan Advisors, "The 2025 Kerrigan Dealer Survey" (525+ franchised dealer responses, collected June - November 2025): kerriganadvisors.com
• Federal Trade Commission, "Automobile Dealers and the FTC's Safeguards Rule: Frequently Asked Questions" (16 CFR 314.4(f), service provider oversight): ftc.gov
• Motor Trader, "Dealers could embrace AI agent compliance for motor finance," July 9, 2026 (iVendi's interpretation of the FCA's Mills Review): motortrader.com