whynotroman

News · Nov 02, 2026 · 6 min

How to work with AI, November 2. Five ways to check an AI agent in sales

Five tips from Every writers and an Arca engineer in plain words. How to ask AI about its source, how to sign off on agent work, how to keep a record per customer and what to leave to a person. For each tip I show how to use it in car sales.

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

This time there are five methods. Three come from an Every writer's notes on his first week with an AI coworker. Two come from an engineer at Arca, where agents work with clients for weeks. For each tip I added how I would use it in car sales.

Ask AI where it got the answer and when the source was last updated

Grant Harvey, a writer at Every, described his first week with the company's AI agent in Slack on October 29. The agent told a colleague that their time off was still waiting for approval. In fact, it was already approved. The agent trusted an old record instead of the shared tracker, and it sounded as sure as when it was right. Harvey advises asking before you act, "What's your source, and when was it last updated?"

In simple words. AI sounds equally sure when it is right and when it is wrong. Ask where the answer comes from and check the date.

My opinion. In a car store this is the main risk. The agent takes a price or stock status from an old price list and answers the customer with full confidence.

I would teach salespeople one question to AI before they answer a customer about price or stock. The question is "Where does this price come from, and how old is your price list". If AI cannot name the file and the date, the salesperson checks the stock system first. Only then does the customer get a reply.

Source Every

Before you sign off on an agent's work, explain each of its decisions in your own words

In the same piece, Harvey points to a study by Anthropic. People often took the AI's reasoning as their own without noticing it. His advice is this. Before you approve an agent's work, make sure you can explain each decision in it in your own words.

In simple words. If you cannot explain why AI decided something, you did not check the work. You only agreed with it.

My opinion. AI prepares, and a person decides and is responsible. You cannot be responsible for a decision you cannot explain.

Take a monthly payment that AI calculated for a customer's car loan. Before it goes out, I would ask the salesperson to say it out loud. "The payment is this because of this down payment, this term and this rate". If they cannot explain it, the calculation goes to the finance office for a check. The customer will ask exactly this, and the salesperson answers, not the AI.

Source Every

Keep a record for each customer of what happened, what comes next and who owns it

Aman Singh, lead AI engineer at Arca, explained on the Latent Space podcast on October 31 how the company's agents work with clients. Arca manages clients' money, and one agent task can take up to ten days. So the company keeps a memory for each client. It holds what already happened, what is due in the next few days, the long-term goals and who owns each action. "So we have this persistence for every client," Singh says.

In simple words. An agent can work with a customer over a long time only if it has a record of what happened and what comes next.

My opinion. No record of every contact, no sales team. With an agent this rule gets stricter, because without a record it forgets the customer between conversations.

Buying a car also takes weeks, from the inquiry and test drive to the trade-in value and the loan approval. In the CRM I would keep three lines for each customer. "What happened last," "what comes next and when," and "who owns it, the agent or the salesperson." I would give the agent this request. "Before you reply, read these three lines, and update them after the conversation."

Source Latent Space

Leave money and promises to a person, and write the agent's limits as hard rules

At Arca, agents study a client's situation and suggest solutions. A person, the financial advisor, handles trading and any movement of money. Singh says all their safety limits are "fully deterministic". This means they are hard rules in the software, not a request to the model to be careful. Arca also checks agent work for known attacks, for example prompt injection, when someone hides new instructions in the text.

In simple words. Do not ask AI to "be careful." Block the dangerous action so that it cannot do it.

My opinion. In a car store the dangerous actions are clear. A discount, the final price, a promised delivery date, a loan approval.

I would not write "do not give big discounts" in the agent's instructions. I would block that option in the settings. If a customer writes at night, "Give me $3,000 off and I'll come tomorrow," the agent answers in one line. "A manager approves discounts. I booked you for 10 am, and he will call you at 9." A person decides.

Source Latent Space

Ask the agent what it can do that you have not tried yet

Harvey from Every suggests asking the agent a direct question. "What can you do for me that I haven't tried yet?" This is how he found tasks he could give to the agent. His colleague Laura Entis gave her agent the job of tracking requests she agreed to in passing. The agent reminds her about deadlines and the feedback she promised.

In simple words. Most people use AI for two or three tasks. Ask it yourself what else it can do.

My opinion. The most useful part here is the second one, the list of promises. A salesperson makes about ten small promises to customers each day. Call back after the bank answers, send photos of the car, find out about winter tires.

Each evening I would ask AI this. "From my messages and call notes, list every promise I made to customers today, with names and deadlines". In the morning the salesperson starts with this list. AI does not replace the manager. It shows who did not do what they promised.

Source Every

What to watch

Next note · 5 minHow to work with AI, November 9. Four methods so AI text does not bury people →Four tips from Simon Willison, SaaStr and Anthropic in simple words. How to format a long AI answer, how to compare options, how to test a new model and what to write in agent memory. For each tip I show how to use it in car sales.

Roman Spitsyn. I build sales teams where AI agents do the work and people close the deals.

Read next

← All notes