How to work with AI, October 23. Four tips for people who sell cars

Every week I read people who build AI and use it every day. I pick the tips you can repeat today and explain them in simple words.

This time there are four. Two come from developers that GitHub interviewed, and two come from Ethan Mollick, a professor at the Wharton School. For each one I add how I would use it when selling cars.

Ask the AI to question you before it starts work#

Patrick Baumgartner is one of three Java developers in a GitHub article. He stopped giving an agent a task in one sentence. He wants to understand the problem first. So he often asks the agent to interview him and find hidden assumptions. Only then does the work begin.

In simple words. Let the AI ask what it does not know before it writes anything.

My opinion. AI usually writes a weak email to a customer because it does not know what the person has in mind. Not because the model is weak. You ask it to "reply to the customer about a Kia Sportage on credit", and it guesses.

I would write this. "A customer asked about a Sportage on credit. Before you write the reply, ask me five questions. Ask only what you need to avoid a mistake."

It will ask about the down payment and the trade-in. It will ask about the term and whether the customer already called. You answer in one minute, and the email fits one real person.

A sales manager can use the same step when reviewing calls. First the model asks what your team calls a good call.

Source GitHub Blog

Do not accept "done" from the AI, name the proof before the work starts#

Baumgartner is strict about this. "Done" should not mean the model said it is done. For each task he sets a check that produces evidence, such as passing tests or matching images. He also runs linters and quality checks and lets the agent fix things until nothing is left.

In simple words. Before the work starts, agree how you will know it is correct.

My opinion. Models like to say everything is finished. For a customer email, the proof is a short list that you give with the task.

The prompt is this. "Write a reply to the customer. It is ready only if it has the price from this month's price list, the monthly payment, and one sentence that offers a test drive. Check each point before you answer and show me." Now you see what is missing without rereading the whole text.

I will say it plainly. AI prepares and checks against the list. A person decides to send the email and answers for it.

Source GitHub Blog

Hand over the whole task and ask for several options at once#

Ethan Mollick described in his One Useful Thing blog how he made a post full of charts. He did not lead the AI step by step. He handed over the whole job. The agent started its own helpers, and they did the research and the charts. They also did the layout and the citations.

Handing work over is cheap, so he asked for several versions and chose between them. He also writes that the AI caught his math mistake.

In simple words. Give the whole task, get three versions, and pick the best one yourself.

My opinion. Managers often ask AI for one text and then rewrite it by hand. That takes longer than choosing from three.

Take a customer who says "too expensive". I would ask this. "Give me three replies to the price objection. One about the monthly payment, one comparing us with a competitor, and one that asks the customer a question. For each, say when it will not work." You read three short texts, pick one, and adjust it for the customer.

Check the result yourself. Mollick says AI is still a weak writer of long texts. Short replies to customers are easier to check.

Source One Useful Thing

Tell the AI the exact file or document where the answer is#

Frank Delporte, the third developer in the same GitHub article, gives the agent a couple of paths to the right files. He says this removes a lot of searching in a large project. He also asks for a test for the change, because it shows a wrong diagnosis. If he does not understand a fix, he asks for an explanation.

In simple words. Say where to look. Then the AI does not invent things or waste time.

My opinion. For a dealer, the "large project" is the price list, the bank terms, and the stock. Paste the right file into the chat and write "answer only from this file". A customer asks about a trim level that is not in stock. From memory, the model will name last year's price.

Here is a phrase for work. "This is today's stock export. Pick three cars under 40,000 dollars for this customer from the list. If none fit, say so." The last sentence protects you from an invented answer.

Source GitHub Blog

What to watch#

These are fresh talks and videos from the last two days. I have not reviewed them yet. The descriptions follow the titles.

• AI Engineer. How to test AI agents when no benchmark exists, a talk by Nathanael Weill from Air Canada
• AI Engineer. Why AI writing still reads like slop, a talk by Louis-François Bouchard from Towards AI, 8 thousand views
• AI Engineer. 1,368 AI failures and what they teach, a talk from New York Life
• OpenAI. Uber rolls out ChatGPT Work to 30,000 employees, a talk between Greg Brockman and the Uber CTO