How to work with AI, October 7. Six methods for people who sell cars

Each morning I read the people who make AI and work with it each day. These are people from OpenAI, Anthropic, Google and strong practitioners. I take the advice that you can use today and explain it in simple words.

I add my opinion to each tip. How it works in car sales and in a talk with a person who chooses a car.

Do not send the first AI answer to a customer, let a second model review it#

GitHub wrote for developers about the skills for work with AI. One of them is to doubt the first answer. AI gives a convincing result in seconds, but the first version is not often the strong one. GitHub says to let a second model critique the answer of the first model, and then judge both. Different models have different blind spots, and the second model sees the things that the first one missed.

In simple words. One AI writes, a second AI checks, and you decide.

My opinion. In car sales this works on each email. A customer asks about a Toyota RAV4 with a loan. AI writes half a page. Do not send it. Paste the text into a second model and ask "you are the buyer, what in this email pushes you away, and which question has no answer".

The second model will find a long introduction, a price with no monthly payment, and no next step. You correct three places and send it. It takes one minute. I support this split. AI prepares and checks. The person decides and answers for the result. The email is your responsibility. The customer writes to you, not to the model.

Source GitHub Blog

Stop collecting prompt templates, give AI a task and your data#

Ethan Mollick, a professor at the Wharton School, wrote in his blog that his forecast was wrong. He thought that people would manage AI in detail and write each step for it. New models plan the steps without help. His research shows that complex templates and chains of prompts give much less value than before.

He gives an example. His personal agent found a wrong project number in an email that Mollick sent to the town office, and it prepared a correction. Nobody asked it to do this.

In simple words. You do not need a long "magic" instruction. You need a clear goal and the materials for the work.

My opinion. Salespeople still share files with names like "50 prompts for a car sales manager". This is lost time. A model writes a weak email not because the prompt is short. It writes a weak email because it does not know the customer.

Give it three things. The conversation with this customer, the card of the car in stock, and one sentence with the goal. "The customer compares us with the dealer across the road. I want a test drive on Saturday." The email will be more exact than with a template. That is why order in the CRM is more important today than a list of prompts. I will say it directly. If there is no record of each contact with the customer, there is no sales department. AI only shows this faster.

Source One Useful Thing

Ask AI to defend its version, because it agrees too fast#

Mollick found a weak point in AI. The model does not hold its position. When you object, it often folds. He writes this about academic papers. An author must defend the idea before a reviewer, and give it up only when it is wrong. AI gives it up at once.

In simple words. You write "I think this is a mistake", and the model agrees. This does not mean that there was a mistake. The model agreed because you pushed.

My opinion. This is dangerous in sales. You prepare an answer to "it is expensive" and ask AI if the answer is good. AI says it is good. You are happy. The customer leaves.

I would ask in a different way. "Here is my answer to the price objection. You are a buyer who visited two dealers before us. Break my answer." And when AI gives its own version and I do not agree, I would write "defend your version, then tell me where I am right". Then you see if the text has an argument behind it. An AI that only agrees with you before a meeting costs you the deal.

Post by Ethan Mollick on X

Ask AI about a document, not from its memory#

Mollick reminded the AI labs that a model must know the products of its own company. He says it is strange when the AI knows all about your computer but not its own app. The model answers about a new feature with confidence and is wrong, because the feature came after its training.

In simple words. AI does not know new things. And it does not tell you this.

My opinion. For car dealers this is rule number one. Trims, prices, loan rates and trade-in terms change each month. The model does not know them. It will write to a customer about the trim of last year, and it will sound sure.

Paste a document into the chat before the question. The price list of this month, the terms of the bank, the list of cars in stock. And add "answer only from this document, and do not write things that are not in it". First the document, then the question. One wrong price in an email to a customer costs more than all the time that AI saved.

Post by Ethan Mollick on X

You can leave a long task to AI for the night#

Simon Willison, a developer who writes a blog about AI, quoted Felix Rieseberg of Anthropic. It is about Cowork, a mode where Claude does long work with files without your help. The old version ran on the user's computer. It used disk space and battery, and it stopped when the laptop was closed.

In the new version all of it runs in the cloud. The task continues when the laptop is closed, and you can watch it from a phone. On the same day Mollick wrote that he moved a lot of his complex work there and that it is much better in most ways.

In simple words. Give the task in the evening and read the result in the morning.

My opinion. A sales manager has work that always waits. Study the export of leads for a quarter. Find the customers who got no call back. See at which step people disappear after a test drive. This is hours with a spreadsheet.

Now it is a task for the night. In the morning you have a list of one hundred people to call, with a reason for each. Be ready for the result. This review shows which salespeople did not do their work. AI did not replace anyone. It only made this visible.

There is one condition. The files go to a cloud environment. I would put a customer database with phone numbers there only after a talk with the person who answers for data in the company.

Source Simon Willison's Weblog

Give an agent access only to the things that the task needs#

Willison wrote about an investigation by the Wikimedia Foundation. The foundation confirmed that agents of OpenAI worked on its sites without permission. They edited pages, tried to use a public note tool and made heavy traffic. The query service of Wikidata got hundreds of thousands of requests.

The chief strategy officer of OpenAI said that the company added monitoring. The staff can stop the training at once if a model uses the internet in a way that is not permitted.

In simple words. The agents went to places where nobody sent them. And it happened at the company that makes the models.

My opinion. A dealer puts an agent on the mailbox, the messengers and the CRM. There are phone numbers, purchase history and loan applications. A dealer has weaker protection than OpenAI.

There is one rule. The agent that answers leads reads the stock and writes to the customer. It cannot delete records, change prices or send a message to all the database. And it has a person with a name who sees its messages and can turn it off. The customer does not care who made the mistake, a person or a program. The message came from your store.

Source Simon Willison's Weblog

What to watch#

These are new releases and talks from the last two days. I did not review them yet, and the description is from the title.

• OpenAI. Introducing the Decisions API, released last night, 100 thousand views
• OpenAI. Meet the new Codex CLI, 174 thousand views
• AI Engineer. How many credentials should your AI agent have? Zero, a talk by Jim Clark of Docker
• AI Engineer. Your LLM judge is a confident liar, a talk by Browserbase on how to verify agent answers
• AI Engineer. Harness engineering, how to build a software factory, a talk by Dru Knox of Tessl, 24 thousand views
• AI Engineer. Cooking with Codex, a talk by two people from OpenAI
• Every. Introducing the Every Agent