How to work with AI, October 14. Four tips: practice with an AI copy of the customer

Today I have four tips. You will learn how to practice a hard talk with AI and when to ignore its review notes. I also cover the new Google model and how to watch an agent.

They come from a writer at Every, Anthropic engineers, Google itself and the security chief of Uber. You do not need to be a programmer. For each tip, I added my opinion and an example from a dealership.

Practice a hard talk with an AI copy of the person, but let the real person decide#

Mike Taylor leads model evaluation at Every. On October 9, he described an experiment. He built AI copies of colleagues from their public messages in the work chat and their public talks. He used the copies to test drafts and to practice meetings before he talked to the real person.

He reported the results honestly. The copy rated his email to his boss 6 out of 10 at first and 9 by round 16. It never gave 10 in 73 rounds, and the email became stiff and formal. Of 12 answers from a copy of an executive, 2 matched well, 4 matched partly and 6 missed. The real person questioned the question itself in 66% of cases, and the copy in 32%.

Taylor's conclusion is this. The copy learns the person's rules, but it does not know when the person will break them. "The meeting is for finding out how she'll surprise you."

In simple words. AI can play a person well for practice. But a real person will say things the copy cannot predict.

My opinion. For a salesperson, this is a ready trainer for objections. I would give the AI a profile of a typical customer and ask it to play this customer.

"You are a buyer, 40 years old, with a family and a budget of $35,000. You compare our crossover with a Chinese brand. Answer 'too expensive' to my price and do not agree the first time. After ten lines, rate me from 1 to 10 and tell me where I pushed too hard".

Five rounds before a shift are more useful than a lecture. But the copy argues like a textbook. A real customer says "my wife is against it" or "I am only looking". AI does not replace the manager. It shows who did not prepare for the talk.

Source Every

When AI looks for weak points, you decide which notes to accept#

On October 13, engineers at Anthropic explained how they changed their code review process. In May, the company merged about 2,500 code changes a day. In September, it merged more than 30,000. Another copy of Claude reviews each change and looks for errors.

The company writes that this review works well but has a narrow view. Engineers had to steer Claude away from extra complexity that it added for problems that did not exist. Simon Willison, who writes a blog about practical work with AI, shared this post. He sees the same thing when he asks models to review each other's work.

In simple words. Ask AI to find problems, and it will always find them. Half of them will be problems that do not exist.

My opinion. A sales manager asks AI to check a call script or an email to a customer. The manager gets twenty notes and fixes all of them. A week later, the script is three pages long, and no salesperson reads it.

I would ask in a different way. "Find the three most serious problems in this email to a customer. For each one, tell me what the customer will understand wrongly. If there are no serious problems, say so."

From the three notes, I would take the one I agree with. AI prepares. A person decides and answers for the result.

Source Simon Willison's Weblog

Test Gemini 4 Argon on your own work task while the price is half#

On October 13, Google released the Gemini 4 Argon model. Google AI Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Search. Developers can use it through Google AI Studio. Google says the model follows style instructions better than earlier Gemini models and answers more directly.

The starting price is $2 per million input tokens and $10 per million output tokens. After November 30, the price doubles. Simon Willison says you should not compare models by price per token. He says what matters is how many tokens a model needs to finish a task.

In simple words. The new Google model is already in the normal app. Compare it on your own task, not by the advertising.

My opinion. I would take one task that I do every week and give it to two models. For example, a review of ten chats with customers. "For each chat, write whether we gave the customer a price, offered a test drive and agreed on a next step. Answer briefly, without disclaimers."

Then I would compare the answers with the chats. Where did the model make mistakes, where did it invent things, and where was it short and correct? If Argon does better, it makes sense to switch before November 30. If it does worse, I saved time on arguments about new models.

Source Google and Simon Willison's Weblog

Keep a log of what an AI agent does and review it like the work of a new employee#

Philip Martin is the chief information security officer at Uber. On October 13, he explained on the Claude blog how Uber watches its agents. His main idea is that AI agents need the same controls as employees. Companies already have these controls for people.

Uber records what its agents do in real time. The company decides which tools an agent may use and where it may connect. All AI use inside the company goes through one gateway. Martin turns every problem he finds into a new rule and keeps watching, because "constraints create new behaviors."

In simple words. If an agent has access, you must see its work. You need a log of its actions, like a report from a new salesperson.

My opinion. A dealer sets up an agent to answer leads at night. In the morning, nobody knows what the agent wrote, what price it gave and what time it booked. Then a customer arrives at a closed door.

I would add a morning review. The agent writes every message and every booking into the CRM. Each morning, the sales manager asks AI for a summary.

"Show all night replies that mention a price, a discount or a visit time. Mark the ones where the price is different from the price list".

No record of every inquiry means no sales department, and with an agent this is even more true.

Source Claude

What to watch#

These are new videos from the last two days. I have not reviewed them yet. The notes come from the title and description.

• OpenAI. Work with your team in ChatGPT Space, a new feature for team work in ChatGPT
• AI Engineer. Decoupling the Brain from the Hands. Claude Managed Agents, Part 1, Anthropic engineers on how managed agents work
• AI Engineer. I Believe Everything My Agents Tell Me. Three Levels of Review, Corey Gallon on how to check the work of agents
• Every. We Tested Gemini 4 Argon For A Week, the Every team used the new Google model for a week
• Latent Space. How to kill your SaaS. Replacing a $43K/yr tool with AI agents, a team replaced a paid tool with its own agents
• AI Engineer. The Agent Works. Does Anyone Trust It?, Michele Grant on trust in agents at work