17 minutes of work, 22 days of waiting: where AI time savings actually hide

Ask a team that's been using AI for a few months whether it saves them time, and most will say yes. Then look at the business. Are proposals going out sooner? Are clients getting answers faster? Is the same team handling more work?

Often, no. The time savings are real, and they still don't show up where the owner can see them. Here's why, and what to do about it.

Where does the time in a business actually go?

In 1990, Michael Hammer wrote a famous Harvard Business Review article about why companies spent heavily on computers and got so little back. In it he mentions an insurer that measured its own application process. An application spent 22 days moving through the company. The time anyone actually worked on it: 17 minutes.¹

Everything else was waiting. Sitting in a queue. Being passed from one department to the next. Waiting for a second review.

Now picture putting the best AI in the world on that process and making every step twice as fast. You'd save about 8 minutes. The application would still take 22 days.

That's the trap. AI is very good at making a piece of work faster. But in most businesses, the work is the small part.

Is there evidence that this happens with AI today?

Yes. The UK's Department for Business and Trade gave 1,000 staff Microsoft 365 Copilot from October to December 2024 and published an honest evaluation.² A few of the findings:

• People liked it. 72% of respondents were satisfied or very satisfied.
• They didn't use it much. About 72 actions per person over the whole pilot, roughly 1.14 a working day.
• Speed and quality didn't move together. In observed tasks, Copilot users made PowerPoint slides over 7 minutes faster but at worse quality and accuracy. Excel analysis came out slower and worse. Time saved on emails was "extremely small."
• The bottom line, in their words: "We did not find robust evidence to suggest that time savings are leading to improved productivity."

That isn't a story about bad software. It's what happens when a tool is handed to individuals and the way the work flows doesn't change. People enjoy it. The business runs the same.

Why doesn't speeding up one step speed up the business?

I'm an industrial engineer, and I spent years in supply chain at Tesla and KLA. One of the first things you learn in that work is that a process only moves as fast as its slowest point. Speed up any other step and the work just piles up in front of the bottleneck.

In an office, the bottleneck is rarely someone typing. It's usually one of these:

• Waiting in an inbox until the right person gets to it.
• Handoffs between people or departments, each one adding a delay and a chance for something to get lost.
• Second rounds of review because the first version had an error, or nobody trusts it yet.
• Chasing a reply from a client, a colleague or another office.
• Re-keying the same information from one system into another.

None of those get faster because someone drafts a paragraph in 30 seconds instead of 5 minutes.

How do you find where AI would actually help?

Before adding AI to a process, measure two numbers for each step:

1. Work time: how long someone actually spends on it.
2. Wait time: how long it sits before the next step starts.

Lay those side by side and the answer is usually obvious. The biggest gaps between work time and wait time are where the money is. Vas, the CEO of Varick Agents, describes a client whose new-case setup took 25 minutes of real work and anywhere from 2 days to 2 weeks to actually go live.³ The 25 minutes was never the problem.

Also write down where you're starting from: how long the whole process takes end to end, how many jobs go through it, and how often something has to be redone. A year from now, that's the only way to prove the change paid off rather than just felt good.

What changes once you know where the time goes?

The fixes look different from "give everyone AI":

• Remove handoffs. If three people each touch a job for a minute, ask whether one person, with AI doing the prep, can finish it in one pass.
• Stop the chasing. Have information arrive in a set place and format, so nobody hunts for it. Some of this isn't AI at all.
• Make review fast and trusted. When AI drafts something, show the reviewer exactly where every fact came from. At one San Diego law firm, a court document packet that took about 45 minutes now takes about 5, because the paralegal checks linked figures instead of re-deriving them.
• Put AI where it removes a wait, not just where it saves keystrokes.

This is also why I don't think of AI as something you add to a business like a coat of paint. You redesign how the work flows, and AI is one of the tools that makes the new design possible. I wrote more about that in The AI isn't what gets better. Your business is.

Common questions

Why doesn't saving time with AI make my business faster?

Because most of the time a job spends in a business isn't work. It's waiting: in an inbox, for a handoff, for a second review, for a reply. AI that speeds up the work itself only shrinks the small part.

How do I find where AI would actually help?

For each step, measure how long someone actually works on it and how long it sits before the next step starts. The biggest gaps are where the time and money are. Fix the handoffs first, then put AI where it removes a wait or a re-check.

Should we measure anything before adding AI?

Yes. Write down how long the process takes end to end, how many jobs go through it, and how often something is redone. Without that, you can't tell later whether AI paid off, only whether people liked it.