How to work with AI, October 12. Five tips: ask AI about yourself, put bans in the system

Today I have five tips on how to keep AI under control and how to stay present in its answers. They come from the founder of SaaStr, the CTO of Workday, the CEO of Every and Google. You do not need to be a programmer to use them.

I added my opinion to each tip. It shows how the tip works in a showroom and in a talk with a person who is choosing a car.

Every two weeks, ask AI who it recommends in your category#

Jason Lemkin, the founder of SaaStr, reviewed the latest episode of the 20VC podcast. On the show, he asked an AI agent where to host an app. The agent ranked the service Render first and called it "my pick". The well-known Vercel was fifth, marked as not suitable for this use case.

Lemkin writes that Render received a qualified lead from him with no sales rep and no ad. Vercel's own numbers show that the share of deployments started by agents went from under 3% to more than 50%.

His advice is this. Every two weeks, ask 10 people you trust, who run agents in production, to ask about your category. Track your ranking. Lemkin says he knows of no product that does this check.

In simple words. The customer does not ask you. The customer asks their AI. Learn what it answers before your rival does.

My opinion. Today's news issue has a measurement of US dealers. AI names three to six stores to a buyer and ignores the review count. The dealer usually does not know about this list.

I would start a table and ask the same question in ChatGPT and Gemini every two weeks. "Where in Leeds can I buy a Kia on finance and trade in my old car? Name three dealers and explain your choice." The table has the date, whether AI named us, our place and what it said about us. I would fix a wrong phone number or address the same day.

Source SaaStr

Put a ban that must hold into the system, not into the prompt#

Gabe Monroy, the CTO of Workday, wrote a guest post for VentureBeat on October 11. His point is this. An agent can complete each action with no error and still break company rules. The agent decides which tool to call and when.

Here is his example. An agent onboards a contractor from another country and skips the checks of local labor law and taxes. Each step worked, and the task failed. Monroy sees a long list of rules in the prompt as weak protection. He writes, "A prompt is a request, not a guarantee."

His answer has three parts. The agent sees only the information it has a right to see. Each action passes the same check as a human action. The agent acts with the authority of one specific person, with no master key. Monroy advises asking the vendor where the guardrails run.

In simple words. It is not enough to write "do not do this" to AI. You must make it unable to do it.

My opinion. A showroom has rules that cost money when broken. These are a discount above the limit, a trade-in price before inspection and a promised delivery date. The line "do not give more than 3% off" in the agent's prompt will fail one day.

I would close this inside the CRM. The agent cannot record a discount above the limit, and only the sales manager can open that field. I would ask the vendor Monroy's question. "Where does the ban work, in the prompt text or in the system?" AI prepares, and a person decides and answers for the result.

Source VentureBeat

When you correct AI, ask it to save the correction as a rule#

Dan Shipper, the CEO of Every, wrote on October 6 about the shared agent his company has used for almost six months. One of the practices is about mistakes. When the agent is wrong, people correct it in the thread and ask it to remember the correction.

The article gives an example. The agent prepared an email where the paragraphs collapsed into one. Brandon Gell, an employee, asked it to fix the draft and to update the skill itself. Now the agent checks how the email looks for the reader before it finishes.

Good work is saved the same way. People ask the agent to save it as a skill, and anyone on the team can run it with all the corrections.

In simple words. Do not fix the same mistake twice. When you fix it, turn the fix into a rule at once.

My opinion. A salesperson asks the assistant to write a reply to a lead. The assistant names the price and does not say until what date it is valid. The salesperson adds it by hand today, tomorrow and next week.

I would say it once. "Fix it and remember the rule." The rule is that every reply with a price gives its end date and what the price includes. Where the assistant has no memory, I would keep a file called "Rules for customer replies" and paste it at the start of a chat. The whole department shares one file.

Source Every

Work with AI in the team's shared chat, not in a private one#

In the same article, Shipper explains why Every dropped personal agents. From March, each employee had their own. Then the company replaced them with one shared agent that works in public threads of the work chat.

On October 11, Every's newsletter added details. Willie Williams, an employee, says the personal bots mostly died off. They were hard to maintain, and "no one could remember which one did what." Shipper noticed another thing. People learned the most when they watched a colleague's agent work.

In a shared chat, everyone sees both the request and the corrections.

In simple words. One person sees a good prompt in a private chat. In a shared chat, the whole department sees it and repeats it.

My opinion. In a sales department, one salesperson has already found how to ask AI for a reply to "too expensive". The other five do not know about it. The method lives in that person's phone.

I would open a shared department chat with the assistant. The one rule is that we write work requests to AI there and paste no customer phone numbers or documents. In a month, the sales manager sees who prepares for calls and who does not. AI does not replace the salesperson. It shows who was not doing the work.

Source Every

Give the agent its own account and open only the files it needs#

On October 8, Google showed a universal Gemini agent for work at its Gemini at Work event. VentureBeat reports this. The agent takes a goal, splits it into steps and can work for several days. For now, it is in a private preview for selected customers.

The description shows how Google limits the agent. It can have its own account with email and Drive storage. It can access only the files that people share with it. Every action goes into a log. A project spending limit pauses the agent when the budget runs out.

Google has not confirmed that every agent will get these services. There are no outside tests of reliability on long tasks either.

In simple words. An agent is a new employee. You give it its own pass and a key to one room, not to the whole building.

My opinion. People often do the opposite. They let the agent into a salesperson's mailbox, and it writes to customers under that person's name. A month later, nobody can tell what the person wrote and what the software wrote.

I would give the agent its own address and its own signature. I would open one folder for it, with price lists, offers and finance terms. I would read its action log once a week. No record of every contact, no sales department, and this applies to the agent as much as to a salesperson.

Source VentureBeat

What to watch#

These are new episodes and talks from the last two days. I have not reviewed them yet, so each line follows the title.

• Lenny's Podcast. Why every company now needs to think and operate like a lab team, a talk with Josh Woodward, VP of Google Labs
• AI Engineer. Containment Is Lying to You. Measuring AI Agents Like Employees, a Druid AI talk on how to measure agents
• AI Engineer. How Secrets Leak Through AI Agents (and How to Stop It), a Venice AI talk on passwords and keys
• AI Engineer. Hill-Climbing Skills. Improve Agents Without Changing the Model, a talk by Shubhankar Srivastava of Browserbase
• AI Engineer. Building Self-Learning Loops for Your Agent, a talk by Fuad Ali of Arize AI
• AI Engineer. Turning My Obsidian Vault Into a Local AI Engineer, a talk by Filip Makraduli of Superlinked