How to work with AI, October 28. Five methods for people who sell cars

This week I collected tips from people who build AI and work with it every day. I take only what you can repeat today, and I explain it in plain words.

This time there are five methods. One is a HubSpot guide to Claude in the CRM, and one is a memo from Amazon leadership to engineers. Two tips come from people at Anthropic, and one is a story from a 404 Media reporter. For each one I added how I would use it in car sales.

Connect AI to your CRM, but check price emails and deal stage changes yourself#

On October 28, HubSpot updated its guide on how salespeople can work with Claude inside the CRM. You connect it once, and it works on any HubSpot plan, even the free one. You need a paid Claude plan, and it works in the browser and the desktop app, not on mobile. The sales plugin has eight ready skills, including call prep and a scheduled daily brief. In the HubSpot example, call prep goes from 15 minutes to five.

HubSpot also lists what a person must check. Emails with prices or contract terms, moving a deal to a new stage based on a call summary, and scoring large accounts. This guide comes from a company that sells its own CRM.

In simple words. Let AI prepare summaries and drafts from the CRM, and let a person confirm anything about money or the customer's status.

My opinion. This is the right split. AI prepares, a person decides and is responsible.

In a dealership I would start with a morning brief for each salesperson. "Show me customers who have waited over a day for a loan answer, and those with a test drive today". Before a call I would ask this. "Give me five lines on this customer". The lines say which car he looked at, what we promised and where we stopped.

The salesperson sends an email with a price and a payment only after checking the numbers.

Source HubSpot

Write down one task where AI helped and one where it did not, and show both to the team#

Terk Terkowitz, VP and Distinguished Engineer for Amazon's worldwide stores, sent a memo to Amazon engineers about working with AI. Leaders must use AI on real work themselves. Each of them should name a task where AI helped and a task where it did not. "If your work does not benefit from current AI tooling, that is a finding, not a failing," he writes. He says to write down the reasons, discuss them with peers and the manager, and look for a fix.

He also warns that the number of AI requests and the amount of output do not show results. Look at what got done.

In simple words. An honest list of "AI helps here, and not here" is worth more than a report on how often the team opened AI.

My opinion. In a sales team I would make this a weekly meeting topic. Each salesperson names one task where AI helped, for example an answer to the "too expensive" objection. And one where it did not, for example a trade-in value without the appraiser's inspection.

After a month, the manager has a map of where to put an agent and where not to. I would count test drive bookings and visits, not how often a salesperson opened an AI chat.

Source Amazon

Do not give an AI agent your passwords, keep access separate from it#

Anthropic's heads of engineering and product, Katelyn Lesse and Angela Jiang, talked on the Latent Space podcast on October 27 about how Claude agents work. One rule goes like this. The agent is never in the same place as the credentials. The agent sees only a placeholder instead of the password. When it needs to reach a system, a separate proxy adds the right credential on its way out.

In simple words. The agent does the work, but it does not hold the keys, someone else keeps them and hands them out.

My opinion. A salesperson does not build agents, but the rule carries over one to one. Do not paste the password for your CRM, bank portal or listing site into an AI chat. Do not send a customer's full passport there.

If a store connects an agent to the CRM or a messaging app, I would give it a separate account with narrow rights. The agent can create a task and write a draft. It cannot delete deals or change prices. Then an agent mistake costs one draft, not the customer database.

Source Latent Space

Ask AI about your store, because buyers already do#

On October 26, Jason Koebler of 404 Media described how ChatGPT keeps sending people with tech support problems to his newsroom. A man in Russia lost access to an Instagram account with 13 years of photos. ChatGPT told him to contact journalists, named 404 Media, gave Koebler's email address and wrote the email for him. For two weeks the outlet has asked people how they found the address. Everyone who answered said a chatbot told them.

In simple words. AI decides on its own where to send a person, and it can send them to the wrong place.

My opinion. Car buyers also ask AI where to go. Once a month I would ask ChatGPT and Gemini a question as a customer would. "Where in my city can I buy a Tiguan on finance, and who do I call if the dealer is silent". I check if my store is there and if the phone number is correct.

The second lesson. AI writes more and more customer emails, so they look smooth. I would teach salespeople to answer the person's question, not the polite style of the email. I would also add "recommended by AI" as a lead source in the CRM. No record of every inquiry means no sales team.

Source 404 Media

Bring your manager a working sample built over a weekend, not a list of ideas#

Felix Rieseberg hires engineers at Anthropic. In a post on X that Latent Space covered on October 28, he said he stopped using classic coding interviews. Claude now does better on them than candidates.

He gave five tips to junior engineers. Two of them are these. Build a working sample over a weekend instead of a long round of requirements and approvals. And stay with one strong idea, because several ideas in one sample hide the good one.

In simple words. Show how it works with one example, and do not try to fit everything into it.

My opinion. In dealerships, the idea "let AI answer night leads" often lives in meetings for months. I would do the opposite. I would take 10 real questions that customers sent at night, for example about a test drive booking or a monthly payment. Over a weekend I would build an assistant in ChatGPT or Claude with the price list and stock, and run the 10 questions through it.

On Monday the manager sees 10 answers, not slides. And one idea only, night leads. Trade-ins and service come later, as separate projects.

Source Latent Space

What to watch#

• OpenAI. Getting a Head Start with Codex, Mitchell Hashimoto, how to give agents research work in the evening and start the morning with results.
• OpenAI. Why Merge CEO Shensi Ding stays hands-on with Codex, why a leader uses AI personally and sets the level for the team.
• OpenAI. From legal bottleneck to business partner, Legora, lawyers gave standard contract checks to AI and kept the decisions.
• Latent Space. How Anthropic ships the #1 LLM in the World, how Anthropic builds agents and checks their work.
• Lenny's Podcast. Your cold email isn't a sales pitch, Rob Snyder on why a cold email should not try to sell.