The AI productivity paradox: when AI doesn't pay off, and how to fix it

If your team is in AI every day and you still can't point to what it changed in your numbers, you're in good company. Most companies are in the same spot. Busy just isn't the same as productive, and the difference is where the money is.

I know the busy side well. I've built about fifty tools with AI, and I probably use five or ten of them. The ones that paid off all started the same way. And when a business leader tells me their company is using AI a lot, the first thing I ask is how people are actually using it.

What is the AI productivity paradox?

The AI productivity paradox is that AI is everywhere in business, but it barely shows up in productivity numbers or profit. Economists saw the same thing with computers: in 1987, Robert Solow wrote, "You can see the computer age everywhere but in the productivity statistics."

What I find interesting is what history teaches us: computers and technology cost money before they save any. They took years to show up in the productivity numbers, and AI is in that same stretch now, but it doesn't have to stay there.

McKinsey uses Solow's line to describe AI today, and its latest survey shows the gap: far more people feel more productive than see it in earnings.

When I read those numbers, my reaction was: that's exactly what I see when talking to prospects. People say they're using AI, and they are, but to the most minimal degree, as a question-and-answer oracle. That does help them be more productive. It just isn't what AI is most powerful for: working as a tool, making tools, writing code and running automation.

Does AI actually increase productivity?

Yes, for the people and tasks it's pointed at. Whether it shows up in profit depends on what happens to the time it saves.

The task-level evidence is strong. In one well-known study, support agents with an AI assistant resolved 14% more issues per hour. Across whole companies, the effect is still small: in an NBER survey of nearly 9,000 executives, AI raised annual labor productivity growth by 0.2% over the past three years, though they expect about 0.6% a year over the next three.

Here's the part I think gets missed. If AI makes an employee faster, but that employee doesn't take on other work, how does it help the bottom line?

Say you have a bookkeeper, and AI helps them do 40 hours of work in 30. If they're used to stretching their work to fill a 40-hour week, the same 30 hours of work now stretches out to 40. You still have the same payroll, and now you also pay for AI on top of it. The only thing that changed in the numbers is a new cost.

That isn't a reason to skip AI. It's a reason to decide where the time goes. Unless that bookkeeper puts the extra time, attention and skills towards another business constraint, you haven't cut a cost or added revenue. This is where most AI rollouts fall short: operating costs, the SG&A line, are sticky. Time saved doesn't turn into money on its own without reallocation.

Are you using AI as a fancy search or to do the work?

Most teams use AI to answer questions, not to do work, and that keeps the gains small. My first question leads to two more:

1. How are people actually using it? Almost always, it's question and answer, or copy and paste: paste something in, then copy or retype what comes out.
2. Have you built any agents that make your people more efficient?
3. If they say yes: can I watch how your team uses it? How are the agents triggered or scheduled?

That usually starts a real conversation about AI as a fancy search versus the real power, which is AI actually doing work for you. So far, everybody I've talked to has said they basically use it as an oracle instead of a tool.

So far, everybody I've talked to has said they basically use it as an oracle instead of a tool.

How do you point AI at what your customers actually need?

Watch how the work actually happens, ask the people doing it, find the step that keeps costing your customer or your team time, and build for that. The tools that pay off are born from a specific need; the ones that sit unused started as ideas.

Most of my fifty tools started as ideas. One was an app that turned YouTube videos into AI skills. Skills are great to have in your AI workflow, so I thought it would be really helpful. I've rarely needed it.

Building tools like that taught me a lot, but it was a lot of time for not enough benefit.

One of the tools that paid off most started differently. CakeMiami is a bakery in Miami, and I built its operating system, CakeApp, with the owner, from what I saw her doing day to day and what she told me.

Here's how she took a phone order. She wrote it on a sticky note and stuck it on her board, saved the customer's contact on her phone, went to the computer to type up the sales order or pro forma invoice, sent it to the customer, and then tied it all together so she had a record. By my estimate, that took five to ten minutes per order.

Now it's done while she's still on the call, in a minute or less. Before the customer hangs up, they have a checkout link with their exact order and details already filled in. Once they pay, everything flows into her CRM and the rest of the process runs on its own. She told me the difference was absolutely significant: it changed her daily workflow and freed up a lot of her time. It also greatly improved her accuracy and her customers' satisfaction.

The rest of CakeApp grew the same way. She wasn't communicating consistently enough, because it was a pain, so we built a conversations module that puts all the company's communication on a desktop computer, where typing on a keyboard is much easier. And when she kept asking me how to run certain workflows, we connected her AI agents to the system, with runbooks behind them, so now her agents can teach her.

None of it started as my idea. It started as her need.

How do you measure AI ROI?

Start with the task: minutes, steps and errors, before and after. Then name the line on your P&L it should move, and check that the freed time actually went there.

Decide which one before you start. The implementation has to solve a real need and be strong enough to move one of those lines. A more efficient team frees up capacity. The payoff comes from focusing that capacity at the constraint, which will always end up being the market in the long run: finding and serving more customers.

Take the checkout link. The task went from five to ten minutes to a minute or less, the sticky notes and retyping are gone, and there are fewer mistakes.

Then comes the check most people skip: where did the time go? When I asked her what she does with the time the whole system gives back, she wrote that she spends it "making more meringues, finding customers and expanding the business." That's time pointed at growth, not absorbed by the week.

If your team is still using AI as a question-and-answer machine, our AI workshops are hands-on: you build an agent on a task you repeat in your own business, so AI starts doing the work instead of answering questions about it. If you'd rather talk through where AI could move your P&L, book a 20-minute discovery call.

You don't need fifty tools. You need one that removes real steps from real work, and a plan for the time it gives back.