The delegation test

"Why does the business still depend on me this much?"

"Why is AI helping with tasks but not changing how the business runs?"

They look like two different problems. They're usually the same operating problem. And there is a number for it.

The gap nobody asks about

Goldman Sachs surveyed small business owners early this year. 93% of the ones using AI said it had a positive impact on their business. 84% said it made them more efficient. Great numbers, right? Everybody shares those two and moves on.

But in that same survey, the same people, only 14% said AI is actually integrated into how their business runs.

So somewhere between "this helped me" and "this is how we work now," about eight out of ten companies fall off. And nobody's really asking what happens in that gap.

What I think the 93% actually is: people chatting with ChatGPT. And nothing wrong with that. I do it every day. But when an owner tells me "I use AI," they might mean exactly that. McKinsey found basically the same gap on the enterprise side. Only 21% of companies actually redesigned their workflows around AI, and those are companies with real R&D budgets and people whose entire job is just this.

So the first real question was never which AI tool should we buy. All of them are pretty good right now.

The test

There is a test, and you can run it this week. I call it the delegation test.

Take one customer escalation. Something real from this week or last week, not a made-up one. Open ChatGPT or Claude and ask the agent to handle it from start to finish. Don't guess the answer. Actually try it.

Then watch. Can it see the account, and does it know what's already been promised? You'd probably need connectors for that. Does it know who's allowed to approve a refund, and at what dollar amount? Here you need a standard operating procedure documented somewhere. Does it know when to stop and raise a flag instead of guessing? That's guardrails and an escalation procedure. And when it routes something to a human, does that human know AI is coming to them, and with what? So you need to name that human as well. And who is overseeing the output of all of that?

Now run it with a person

Now run the same comparison with a person. Imagine the agent is a real human hire who walks into the office tomorrow. Could they trust the data? Could they act inside your systems without breaking something? And most importantly, who's their manager, accountable for the results once whoever built it walks away?

If you can't hand the task to a brand-new hire without babysitting them for a month, an AI agent is going to hit exactly the same wall. Except a human can raise a hand and ask you for a quick call. AI would probably just guess. And the scary part is it will sound very confident doing that.

Most companies I work with, 10 to 100 million in annual revenue, honestly can't answer those questions yet. And I don't think it's a failure on anybody's part. There's just never been a role in most organizations whose job is documenting processes and keeping the data clean. Companies were always built for people by people, and people fill gaps with judgment. We improvise. Those gaps generally never hurt anybody until the automation conversation started.

How long it takes

How long does the test take? It depends. If everything the agent needs is ready to go, about five to ten minutes. If it isn't, that might take anywhere up to a week, I think. And passing it on one task is not the end of AI implementation. That's only the beginning.

Run the test on one real escalation and tell me how long it took you. I'm genuinely curious.

93% said it helped. Only 14% actually run on it. That whole distance is operational improvement.