Last week, four AI agents beat Minecraft together. They split the work, shared what they learned, and kept going until they won. No human help.

Now compare that to what’s happening in your business.

A qualified lead hits your inbox. Does someone follow up? How fast? Or does everything stop until you weigh in?

For a lot of entrepreneurs I talk to, it’s a full stop.

Last week I said most businesses aren’t ready to hand work to AI. This week that gap got a lot harder to ignore. AI got better AND cheaper this month, yet most businesses still aren’t seeing the payoff.

The tools can do the work. So what’s holding up the results? Spoiler: it’s not the AI.

Here’s why most business automations keep failing, and exactly what to do instead.

What can AI agents do when they split the work?

Google dropped four Gemini 4 Argon agents into a fresh Minecraft world with one goal: beat the Ender Dragon.

No assigned roles and no detailed playbook to follow. They got a basic bot library, shared notes, and a group chat.

The agent team wrote 2,783 programs. Picked up each other’s lost items. And won in just over 27 hours. No human stepped in.

Oh… they also died 605 times.

In a game? That’s cute. In YOUR business? That’s 605 pissed-off clients and maybe a lawsuit.

But here’s what made it work:

• A clear goal
• Shared information
• Tools they could actually use
• A way to check progress

That’s your checklist. If you want AI to do more work with less babysitting, you need ALL four of those things locked in.

What two decisions should you make before handing work to AI?

Before you hand over the work, you need to make two more decisions.

1. Set limits and approval rules

As agents get more capable, big firms are drawing hard lines. Schwab is blocking outside agents from moving customer money. JPMorgan’s doing the same thing. They’ll let AI research a customer. But making promises TO that customer? No way.

Your business needs the same rules:

• What can an agent do on its own?
• When does it need approval?
• Who steps in when things go sideways?

Figure that out before you flip the switch. Not after you accidentally refund someone $10K.

2. Decide what stays human

Some work should stay with people. Some can go to AI. Some needs both. The key is knowing which is which, and when AI should hand the work BACK to a person.

My favorite framework for this is Find, Filter, Free. I recorded the full walkthrough, including how we helped a CEO get two full days back each week without hiring anyone: Before You Hire Another Employee, Do This.

And NOW is a very good time to put it to work, because AI just got cheaper. Again.

How much cheaper did high-volume AI work get?

Earlier this month, Anthropic launched Claude Haiku 5.5. For prompts up to 100,000 tokens, input and output prices are 90% lower than Haiku 4.5. Anthropic estimates average running costs dropped about 75%.

Forget the model names for a sec. Think about the WORK:

• Pulling key facts from documents
• Sorting incoming requests
• First-pass research
• Turning long notes into usable summaries

All about 75% cheaper now. Lower prices make these tasks cheaper to TEST, and running them across your whole business starts to look realistic.

Why can cheap AI output still create expensive work?

Here’s the thing: cheap output can still create expensive work.

If someone spends 20 minutes fixing a summary that used to take 10 minutes to write by hand, you didn’t save time. You added work.

Count the human checks, fixes, and the back-and-forth. Measure the cost of a USABLE result, not just the cost of an AI response.

Before any automation goes live, I run it through The 3S Test.

But here’s the million-dollar question. If AI is doing MORE and costing LESS, why are most businesses still not getting much out of it?

Why are most businesses still not getting value from AI?

Gartner surveyed 2,000+ tech leaders. Only 13% said their business had gained major or game-changing value from AI.

Bain studied 951 companies and found them realizing about 38% of AI’s expected value.

Two studies. Different yardsticks. Same finding: most businesses are getting a FRACTION of what they expected.

My take? It’s easy to get excited about what an AI tool can do. But before buying it, there’s a question worth answering: what’s ACTUALLY broken in the business?

If that’s unclear, you can spend a lot of time automating work that wasn’t the problem. Too many businesses are generating offers faster while qualified leads still wait a day for a reply. You’re summarizing every meeting while nobody owns the next steps.

You made a task faster. Did you make the business better?

How would you fix slow lead follow-up with AI?

Let’s look at one place your business is probably losing money while everyone waits on you.

Say you run a consulting business. Someone fills out your inquiry form. Your assistant reads it, asks you if they’re a fit, and waits. Eventually someone follows up. Sometimes that’s the next day. (Whoops.)

Find the constraint

Why does every inquiry come to YOU? Usually because nobody else knows your qualification rules, or your form doesn’t collect what’s needed to decide.

Write the rules down. Then measure two things:

• How long leads wait for a reply
• How much time your team spends on each one

Now you have a clearly defined problem to fix.

Filter the work

Give AI one small piece. It reads the inquiry, gathers key details, and drafts a reply based on YOUR rules. It flags missing info and suggests a next step. A person checks before anything goes out.

AI cannot invent an offer. It cannot approve a discount. It cannot promise a result.

You don’t need to script every click. You need a clear goal, the right information, and firm limits. The Minecraft agents had the first two. For your business, make the limits just as clear.

Free the people, then prove it

Run it for two weeks.

• Are leads getting useful replies sooner?
• Is your team spending less time checking and fixing them?
• Are more qualified prospects booking calls?

If yes, expand. If no, find out WHY before you add another agent. Maybe key details are missing. Maybe the rules need work. Maybe you fixed a delay that wasn’t the main problem. Find out first before you change anything else.

Your next move

Don’t try to automate everything at once.

Pick ONE process that’s killing time: the one where work keeps waiting on you as the leader. Map it. Measure it. Hand ONE piece to AI. Then see if your business actually works better. Check the time saved, the corrections needed, and the result.

Do this once a week for 90 days, and your business won’t look the same when you’re done.

This is the way forward.

Onwards, Joe

Want help putting this to work?

• Binge my best training on using AI to grow revenue without hiring more people on YouTube.
• I’ve installed AI in 400+ businesses. This video breaks down why most people get stuck and how to actually save time and drive revenue.
• Want me to install AI agents in your business and train your team to run them? Learn more.