THE NUMBER
Bots passed us in June.
Most of the internet's traffic is no longer human: 57.5% of webpage requests now come from machines, a crossover Cloudflare's CEO had predicted would not arrive until late 2027. Meanwhile about 80% of your customers still want to order from a person. The gap between those two facts is where the money is.
3 THINGS HAPPENING RIGHT NOW
McDonald's and Wendy's are automating the drive-through anyway
McDonald's is rolling out AI ordering chatbots called ArchIQ. Wendy's has run FreshAI since 2023. Both push ahead while the research says customers want a person at that moment. "There's a significant gap between how fast companies are moving on AI and how ready consumers are," said Layne Haaksma of the advisory firm Metrigy.
Agents are a customer segment you have never designed for
Traffic from agents that actually do things, clicking links and filling forms, grew 7,851% in a year. Seventy percent of commands reaching the payments company Stripe now come from agents. Your website has an audience you never designed for.
Argentina tried to write humans out of the company, and couldn't
Argentina submitted a bill in May creating the "non-human corporation," run by AI agents that can sign contracts and hold assets, human shareholders optional. But the people are still in it, in three required roles. The most aggressive attempt yet to remove humans only reduced them.
THE DEEP DIVE
The most expensive lesson in AI is being learned at the drive-through
McDonald's and Wendy's are running the largest public experiment in customer-facing AI that exists, spending enormous sums to put an agent between the customer and the order. About 80% of customers still prefer a human there, a number that moved just over four percentage points last quarter.
Jerry Jacobs, a sociologist at Wharton, proposes a split: let AI take the repetitive part of the ordering, and move the employee toward hospitality, greeting people and fixing what went wrong.
I would sort it differently, and the 80% is my reason. Repetitive is not the test. Witnessed is. Automate where being wrong is cheap and nobody is watching; keep a person where being wrong has an audience. Taking an order is repetitive, which is why it looks like the obvious thing to automate, and it is also the one moment the customer is standing there watching. Most owners get this backwards, cutting their biggest labor cost first. A bungled invoice costs you an hour. A bungled order costs you the customer.
ONE THING TO TRY THIS WEEK
Give an agent a bad instruction on purpose, somewhere it cannot hurt you.
Hand a task to a person and they ask a question. Hand it to an agent and it does not. It picks an answer and hands you something fluent. Ten minutes to watch that happen safely.
On paper, write one real instruction from your week, as you would say it out loud. "Write the follow-up note for anyone who didn't get back to me." No customers yet? Use your own work: "Put together this month's expenses."
Underneath, write your prediction: what will it get wrong? Be specific, before you run it. This step is the whole exercise. Skip it and you are just reading output again.
Open Claude Code, or claude.ai in your browser if you have never used a terminal. Say: "Draft only. Don't send anything or contact anyone." Then paste your instruction, exactly as written, and nothing else.
Compare what came back against your prediction. The gaps you did not predict are the expensive ones: decisions you never knew you were leaving open.
Rewrite the instruction with every guess spelled out, and start a file called instructions.md wherever you keep work notes. In a month it is your set of instructions known safe to hand over.
The lesson is not that agents are unreliable. It is that fluency is not evidence. An agent that misunderstood you sounds identical to one that understood you perfectly. You have to go looking for the difference.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
Next issue: how to read what your agent actually did, not just what it handed you. The trail it leaves is more honest than the summary it writes, and almost nobody looks at it.
Manu