How to work with AI, October 19. Four tips: give every agent an owner

Today I have four tips. Who answers for the work of an AI agent, and how to unload a strong model. How to treat numbers from second-hand reports, and how to teach a team without lectures.

You do not need to be a programmer. I added my opinion and an example from a dealership to each tip.

Give every agent an owner with a name and read its edits like a colleague's edits#

Mike Cannon-Brookes, CEO of Atlassian, told Fortune on October 19 that you cannot blame AI for a mistake. When a person works in ChatGPT, he said, "That's Mike using ChatGPT. It's not ChatGPT doing it." He added, "The accountability is with Mike. It has to be with the end user, ultimately."

On October 6 Atlassian launched a protocol where agents appear as team members. Each has its own profile and a named owner. The edits of an agent show in the document history, like the edits of a colleague. "You have to be able to see who did what when," Cannon-Brookes says.

He thinks personal AI tools are fine for personal work, but not for shared work. In shared work a chart built from internal documents can leak if someone shows it carelessly.

In simple words. The person who started an agent answers for it. So every agent needs the name of a person and a trace of its edits.

My opinion. An agent tells a customer in chat that the trade-in is worth 900,000. A day later the customer arrives with that number. If the agent has no owner, "the system" gets the blame, and the customer stands in front of a salesperson.

I would make a one-page table. It lists the agent, what it may do, who owns it, and who reads its messages on Fridays.

In the first line of the agent's instructions I would write this. "You are the assistant of salesperson Ivan. Mark any answer with a price, a payment or a date as 'Ivan's draft'. Do not send it until he approves."

AI prepares, a person decides and answers. If nobody answers for an agent, switch it off.

Source Fortune

Give simple sorting of leads to a cheap model and keep the strong model for hard answers#

Jason Lemkin, founder of SaaStr, reviewed this week's AI software news. He writes that the Claude model Haiku 5.5 costs 90% less than Haiku 4.5 on prompts under 100,000 tokens. For a large part of his tasks, he says, it is almost as good as Sonnet. Lemkin notes that his AI costs fell for the first time in two years.

He also describes a "decision model". It returns a decision and a confidence score, not text. Lemkin uses it in production and says it gets 70% to 80% of carefully described decisions right. He estimates that such tasks are about 10% of all AI use. They are sorting, routing and yes-or-no filters.

In simple words. Not every task needs an expensive AI. A cheap one can sort letters into folders, and a strong model should write the text to the customer.

My opinion. A night lead arrives as one block of text. "Hello, how much is a Kia Sportage on credit, and I also have service on Saturday." It needs a label, not a nice text.

I would give the cheap model this prompt. "Read the message and choose one word from this list, buy, price, service, spam, unclear. If you are not sure, write unclear." The strong model writes the reply, and a salesperson reads it before sending.

I would compare the first 20 leads with how I would sort them myself. If there are more than three mistakes, rewrite the prompt.

No record of every inquiry, no sales department. The label of each lead and the time it was set must sit in the customer card.

Source SaaStr

Treat a number from unnamed sources as provisional and find the original#

Simon Willison, a programmer who writes one of the most read blogs about AI, reviewed a Financial Times report on Anthropic revenue on October 15. The paper writes that the company told investors its third-quarter revenue grew almost 50% to $16.8 billion. Willison points out that the figures come from people with knowledge of the matter, not from official filings. The phrase "told investors" appears four times in the article.

He calls his own estimate of what rivals spent on copying Anthropic's models a napkin calculation, not a reported fact.

In simple words. If a number arrived through "sources", it is not checked yet. Until you find the original, treat it as provisional.

My opinion. A sales manager reads in a newsletter that Chinese brands took a quarter of the market and brings it to the team meeting. The number may come from a vendor's press release. Then he explains it to the owner as a fact.

I would ask AI to mark the numbers in any text before I bring them to people. The prompt is this. "For every number in the text, write who stated it and whether there is a link to the original. Mark numbers without a link as 'provisional'." I would open the links myself, because AI also makes mistakes with addresses.

Source Simon Willison's Weblog

Show your team one working example instead of talking about AI#

The editors of Every, a publication about working with AI, gave teams a tip on October 18. Do not explain AI in general terms. Show colleagues one working method in the team chat. One good example spreads through a team faster than a lecture.

The same issue has a live example. Becky Isjwara at Every uses the company agent as a buddy for new hires. The article lists the prompts she advises a new employee to try on day one.

In simple words. People learn from a ready example that worked for a colleague, not from a talk about how AI works.

My opinion. A sales manager gathers everyone for a meeting about AI and talks for an hour about possibilities. A week later two people use it. The salespeople saw slides, not their own work.

I would pick one salesperson who gave AI a post-test-drive email today. And I would ask them to post three things in the team chat. What they asked, what they got, what they fixed.

The phrase is this. "Here is my prompt, here is the AI answer, here is what I fixed. Take the prompt and put in your customer." On day one I would give a new hire three prompts from that chat.

AI does not replace a manager, it shows who did not do the work. If nothing changes for a salesperson after the demo, the talk is no longer about AI.

Source Every

What to watch#

These are fresh episodes from the last two days. I have not analysed them yet, and I describe them from the title and the announcement.

• OpenAI. The new ChatGPT Voice, powered by GPT Live, released October 19. The page that describes the feature did not open, so there is no analysis
• AI Engineer. Stop telling agents how, the OpenProse team on writing tasks for agents as a contract in plain text
• AI Engineer. Evals for high-stakes AI agents, Langfuse with a banking assistant as the example
• AI Engineer. Why AI agents still will not pay for anything, Jan Curn of Apify on agent payments
• AI Engineer. Backtest your agent against yesterday's web, Ryan Ahern of Exa on checking an agent's decisions after the fact
• Lenny's Podcast. Fall in love with the problem, not the product, Josh Woodward of Google Labs on the signal he uses to judge early prototypes