I Had Two AI Agents Debate the Future of Work

I’m kind of obsessed with AI and how it’s going to change business, so I spun up two Claude Code subagents and gave them personas: "Mike," is the owner of a $5m accounting firm, and "Sarah," is the owner of a $3M marketing agency.

They are both skeptical, busy, and tired of buzzwords.

Then I tried to explain something called a "semantic layer,” and why it matters for their businesses.

They pushed back pretty hard.

What emerged was one of the best explanations I’ve seen of how AI will change knowledge based businesses.

I told Mike and Sarah that:

Then I introduced the "semantic layer" concept. It’s basically means documenting how your business works in a way AI can actually use.

Mike's response, "I've tried documentation before. Year one, someone writes a 400-page manual. Year two, it's outdated. Year three, nobody's looked at it.”

“So how do new employees figure out what to do,” I ask.

“Usually they ask Janet. She’s my senior CPA and has 20 years of experience. Even if I document stuff, it’s impossible to know everything Janet’s learned. You can't just write that down."

Sarah's response, "We already have brand voice docs we write. We’ve tried to use ChatGPT to develop them but it just produces generic garbage so we’re still writing them ourselves for the most part."

Fair points.

“The course was the most beneficial of the AI courses I’ve taken due to its practical, real-world application and I’m using the tools I built every day in my business,"

-Chris C, SMB owner from our Fall Cohort.

AI is quickly moving from an experimental toy to a mission critical tool. I’m teaching owners the key skills they need to start taking advantage. In six knowledge-packed sessions you’ll learn what you need to make AI your strategic weapon.

Here's what I told them, “The difference isn't the documentation. It's what's using it.

Old world: You document → Humans read → Humans apply judgment

New world: You document → AI reads → AI applies it consistently, every time

Sarah's brand voice docs sit in Notion. When someone uses ChatGPT, they copy-paste a snippet. That's not AI that knows her client; it's a human translating for a dumb tool. Mike's senior CPA, Janet, explains the same things to junior accountants 50 times every tax season.

What if Janet explained it once, and that became instructions an AI could follow?

It wouldn’t replace Janet. It would scale Janet.”

For Mike, I made it concrete, “Janet knows that for restaurant clients, you always check tip reporting, food waste deductions, equipment depreciation, Section 179 opportunities, FICA tip credit, and state-specific liquor license treatment.

Today the junior employee asks Janet what to do. She explains, the junior employee does it, and Janet reviews. Repeat 50 times.

With AI you can record Janet explaining this once, turn it into a document the AI can read, and feed it to an AI tool.

Now:

  - the junior employee says: "Restaurant, $800K revenue, Texas"

  - AI generates the client-specific checklist

  - the employee does the work

  - Janet reviews just the output without explain the whole process.”

Mike said, “I get that but how do I juggle the 100s of things Janet knows? It’s just a bunch of files sitting somewhere like my 400 pages of documentation. How does the AI know what to use.”

“That the cool part with LLMs like Claude. It can detect what the junior employee is working on and call the right process at the right time. She just needs to record it once and any new employee can use it. Janet stops being a bottleneck and her knowledge compounds.”

Mike got quiet.

Then he said, "Wait. I've had Janet working for me for 12 years. How many client calls has she been on? How many times has she explained something that just... vanished? All those conversations where she told a client exactly what to do. That's gone. That was my firm's value, and it walked out the door every night."

Sarah jumped in, getting it now, "It's worse than that. Every client call where Jess explains our strategy. Every internal meeting where she tells a junior employee why the headline doesn't work. Every Slack message where she gives feedback. That's our intellectual property. And we're just... letting it evaporate."

That's the moment it clicked for both of them.

“Your semantic layer isn't a nice-to-have. For a knowledge business, it is your business.

Every client call you don't record, every expert explanation you don't capture, every "let me show you how we do this" that happens over someone's shoulder. That's value slipping through your fingers.

Not future value but value you created today that you'll never get back.”

Mike started doing math out loud, "Twelve years. Say 200 client interactions a year. That's 2,400 conversations worth of Janet's expertise just... gone. If I'd been capturing that, I'd have a goldmine. Instead I have a few files in OneDrive."

Sarah was already thinking ahead, "Okay, so we start capturing everything. But here's what's bothering me, if we get really good at this, if we document all of Jess's insights... don't we just make ourselves easier to replace? Client takes the playbook and walks."

That's the fear underneath the fear. If your value is your knowledge, and you write it down, doesn't it stop being yours?

Here's what I told her, “The rules aren't the value. The updating of the rules is the value. Think about it. Jess's insight about Client X, "sound authoritative, not defensive.” That's true today. But:

  - What happens when Client X's market shifts?

  - When a competitor changes positioning?

  - When what worked 6 months ago starts sounding stale?

Who notices? Jess.

Who updates the rules? Jess.

Who knows when to update versus stay the course? Jess.

A document is static. Judgment is dynamic. The client can take your playbook. They can't take your ability to evolve it.”

Something clicked for Mike, “Janet doesn't just see one restaurant. She sees dozens. Right now, Janet handles maybe 40 restaurant clients. She notices that the ones in Texas are handling tip reporting differently than the ones in California.

She spots that three clients got flagged for the same depreciation issue this quarter. She realizes a new deduction is being underused across the board.

That pattern recognition? It's locked in her head. She might mention it at a staff meeting. Mostly it just... sits there.”

I say, “Exactly. Now imagine Janet with AI assistance. The system surfaces, ‘17 of your restaurant clients haven't claimed the energy efficiency credit. Based on their profiles, 12 likely qualify.’

Janet reviews it. Confirms it. Suddenly she's not just doing returns, she's proactively calling clients about money they're leaving on the table.”

"Wait," Mike said. "She goes from processing work to generating work. From reactive to proactive. From 'did we do this right' to 'here's what you're missing.'"

Sarah saw it too, "Jess reviews content for 15 clients. If she's capturing what works and what doesn't across all of them, she's not just better at Client X. She's seeing what's working across the whole B2B space in real time. She becomes a strategist with market intelligence, not just brand guidelines.

“Your expert's judgment isn't just preserved, it's also multiplied. They're not just updating rules for one client. They're seeing patterns across 50, 100, 200 clients that no individual engagement would reveal.

Janet with a checklist is a good accountant. Janet with AI surfacing patterns across your entire client base is a strategic weapon.

That's not replaceable. In fact it’s a real moat and one you could start building today.”

But then Mike said what was really eating at him,"Okay, I get it. Start capturing. Janet evolves the system. We build a real asset. But aren’t there legal issues? What if I clients don’t want us recording calls? What if the data leaks out?"

I reply, "You’ve got to work through the issues. Most likely you’re already using note taking apps. Clients are getting used to it. Put something in your engagement agreement about it. And sure there’s risk of data leakage but not any more than storing documents on the cloud. You already do that, right?”

Mike nods, "And the risk of doing nothing is that I lose 12 more years of Janet. And maybe the whole firm by 2030."

Sarah added, "The gap isn't 'who has AI.' It's 'who's been feeding it.' A competitor who started 6 months ago has 6 months of compounding knowledge I can't catch up to."

Mike, “ I’m scheduling 30 minutes with Janet to record her explaining the restaurant client checklist. We’ll transcribe it and test it in Claude Projects.”

Sarah, “I’m going to ask my strategist, ‘What 5 things do you always fix on Client X?’ We’ll record it and make it something the AI can read. Then we’ll rinse and repeat.”

The "semantic layer" isn't a tech project. It's answering one question: What do your best people know that you've never written down?

Just start recording and saving anything you can. That’s real intellectual property. Next you can experiment with LLMs to extract the knowledge and patterns. Then you can start using tools like Claude’s skills feature to make that knowledge useful to your other people.

In the end though the tools are a commodity and the knowledge is the moat. Every day you wait is another day of expertise evaporating and pretty soon your competitors are going to start bottling theirs.

I’ll be teaching this exact method in my next AI for Business Owners Cohort starting January 22nd. Sign up here if you are interested.