Key Takeaways
- →Most AI implementation problems are human problems. We forgive a $100,000 hire and quit on a $20 agent after one mistake.
- →An agent arrives with knowledge but no wisdom. A written skill file hands it your process every time.
- →Own the audience, own the process, own the taste.
On October 30, Reddit grew revenue 53% and still had the worst day of its life as a public company. The stock fell 29%. The company had told investors its Google search referrals were down by more than a million a week.
Reddit didn't lose its product. It lost the road people used to find it.
I've spent the last month watching companies learn the same lesson with AI. Everybody wants to know why AI agents fail. Most of the time, the answer looks a lot like Reddit's: they built on something they never owned.
Sponsored by PFS Global. PFS helps independent insurance agencies with accounting, taxes, payroll, and commission reconciliations. Get an assessment: https://pfs.ryanhanley.com
Find the work you're built for. Get your free Edge Finder assessment: https://easymodebook.com/assessment
Reddit Rented Its Audience From Google
On the earnings call, Reddit said October search traffic dropped by more than 1 million referrals a week. It warned that daily users could fall by a low double-digit percentage in the fourth quarter.
Reddit CEO Steve Huffman didn't dress it up on the earnings call: "This means Google has really shifted in their relationship with the rest of the Internet."
That's the bill for a rented audience. I've made the owned versus rented argument to owners for years.
LinkedIn, Instagram, YouTube, and Google are where your customers spend their time. Be there. Add value fast. Then move people somewhere you control.
The one asset nobody can take from me is my email list. If my email platform changes the rules tomorrow, I export the list, upload it somewhere else, and I'm right back in business.
Single-channel demand is a landmine. And none of us knows yet what AI is going to do to the roads we use to find customers.
Why AI Agents Fail Is Mostly a Human Problem
The pullback is real. KPMG surveyed 2,145 senior leaders at larger companies, and 49% said they scaled back AI agent deployments because the operating costs outweighed the benefits.
Cloudera found 89% of enterprises scaled back or redesigned AI plans over token costs this year. Fewer than a quarter said AI generates more value than it costs.
Meta scaled back plans to replace staff with AI agents after employees spent more and more time cleaning up technical and security errors, according to Reuters reporting summarized by The Guardian.
Those numbers describe a deployment problem. Here's the pattern I see over and over. A process breaks. Somebody says, throw an agent at it. The agent gets a two-sentence prompt and an expectation of perfection.
Then we judge it on a standard we would never apply to a person. Hire a middle manager at $100,000 a year, and when they make a mistake, we say that happens, you're learning.
Buy an agent for 20 bucks a month, and one mistake gets a verdict. AI sucks. It doesn't work. We can't use this.
You have to train the AI like you train an employee. Same scrutiny, same grace. The issues AI implementations are having are human issues, not AI issues.
READ NEXT: The Airplane Didn't Work Either (an AI Story)
AI Is Born With Knowledge, Not Wisdom
When the AI is born, it has knowledge, but not wisdom.
Ryan Hanley, on Finding Peak
A new agent has read more than anyone in your building. It has no idea how your company does anything. It doesn't know where your newsletter lives, what the sections are, or which sources you trust.
Our job as users is to give it the wisdom. The best way I've found is context and skills. A skill is a written workflow the agent follows the same way every time, so you get consistent results.
Here's what that looks like in my world. In July, I started Badass Breaks, a sports card business, with my 12-year-old son. I could not have started it without AI. I don't have the hours.
The business runs on a small crew of agents, and the newsletter is the clean example.
Without a skill, I say, "Hey, I want to work on this week's newsletter." The agent asks where it's housed, what it looks like, and what company we're talking about.
With the skill, it reads the file and already knows. It sends through Beehiiv. It's for the Finding Peak brand. It knows the four sections, the sponsor graphics, the tone and voice, and the sources I recommend.
If your AI output is inconsistent, this is usually why. It's relearning your process from scratch every time you ask. Memory in ChatGPT, Grok, or Gemini is not the same as a skill file.
The loop is simple. Dial in a repeatable task. Have the AI write the skill. Reference it every time. Update it as the process improves. I call skill files an absolute superpower, and I mean it.
Good companies already do this for people. They keep Loom directories and Google Docs that show a new hire how the work gets done. A skill file is the same training manual, written for a teammate that reads it every single time.
My read on the pullback: the companies hitting the brakes don't need more agents. They need written processes.
The Filter: Bot or Skill?
Before you add another bot, run the task through three questions.
Can you explain it on one page? If you can't write the process down, the agent can't follow it. Neither can your next hire.
Does it repeat? Skills pay off on work you do every week. One-off judgment calls stay with you.
Will someone own the update? A skill goes stale the day the process changes. Whoever owns the process owns the file.
Three yeses, write the skill before you buy the bot. Any no, fix the process first.
Production Can Be Outsourced. Taste Can't.
Fanta wanted to own Halloween the way Coca-Cola owns Christmas. It built a Haunted Universe ad with GRAiL, an AI-powered entertainment studio. Creative Bloq's headline called it "a real horror show."
The consumer data is more interesting than the pile-on. In Zappi's test, the ad beat the norm on love, 35% against 27%. It also beat the norm on dislike, 7% against 2%. Polarizing, in other words.
Fanta's global VP of marketing, Ibrahim Khan, pushed back on the idea that AI made it easy. He said the work took a year and a half.
The lesson I take from Fanta is about taste. AI can tell you, tactically or technically, that something works. It can't tell you if it's worth posting.
Use AI for production. Use it for idea generation. Use it for the first draft. When you hit publish, your taste is what's on the line, and you have to like it. Never outsource your taste.
Even the week's political story fits the pattern. Trump's push to ban the term "AI" in favor of "super intelligence" is a rebrand, and renaming an idea is how you own the narrative around it.
Own the audience. Own the process. Own the taste.
Get Back Up
The best lesson of my week came from my fifth grader, Colton. He had never played football. This year he asked to play, and his team went 8-0 and won the league.
What I told him all season was simple. In football, you figure out what you're made of the first time you get knocked down. The guys who win get back up. You might get knocked down again on the next play. Get back up.
When we recapped the season, he said it himself.
Dad, I just kept getting back up.
Colton, my son, after his first football season
That's the AI lesson too. Your first agent will make mistakes. So will your fifth. Give it grace, write the skill, and get back up.
It's easy in today's world to skirt hard for comfort. Find something hard, something with an element of danger. It will wake your brain up in ways comfort never will.
I wrote more about why to do hard things on purpose after my conversation with Chris Paradiso.
Go find something hard.
Sources from this episode
Follow Ryan
This is the way.
Hanley.
Disclosure: I currently work for PFS Global, the company behind this episode's sponsor read.
