THE NUMBER
Nearly nine in ten CEOs say their company already gets a benefit from AI. Only 14% have worked out what it does to their bottom line.
That is a July survey by Boston Consulting Group, a big management consulting firm, of 152 chief executives at companies worth $500 million or more. They feel it helping; almost none have proven it paid for itself. How often an automation must run to earn that cost back is what this issue is about.
3 THINGS HAPPENING RIGHT NOW
One operator’s test: multiply minutes by how often it runs
One operator writing on dev.to reduces it to one step: multiply the minutes a task takes by how many times a month you do it. The biggest number wins. A once-a-month task gives an automation twelve chances a year and eleven months to quietly drift. Rare tasks lose this math almost every time.
The most re-cited number in AI: 95% got no measurable return
A 2025 report from MIT keeps getting quoted for one figure: about 95% of organizations that tried generative AI got no measurable return. Not that the tools failed in a lab; most deployments never reached the point where the payback showed up. The number is a year old and gets stretched in headlines, so hold it loosely. The direction is the same everywhere: for most, it never got there.
THE DEEP DIVE
The task that hurts most is usually the wrong one to automate
One operator, writing on dev.to about six automations he gave up on, kept rebuilding one for a job he hated: creating product listings. It fought him every time, because the format kept shifting underneath it. He killed it, did the listings by hand until the work got boring, then rebuilt it in a single afternoon once something stable was finally there to automate. His rule now: run the task manually until it is boring.
Here is why. An automation is a fixed cost you pay once, in setup and supervision, against a benefit you collect only when it runs. Break-even is that cost divided by what one run is worth. A rare task barely runs, so it almost never earns the cost back, however much it hurts. A dull daily task clears it easily.
The automations worth keeping are rarely the memorable painful ones. They are the frequent, dull ones nobody brags about. Automate by pain and you built the wrong half.
ONE THING TO TRY THIS WEEK
You picked what to automate by how much it annoyed you. The math cares about how often it runs. Here is how to find each automation’s break-even run count, and which of yours never reach it.
List every automation or paid AI tool you run. Beside each one, write a single number from your own records: how many times it actually ran last month. Check your calendar, your sent folder, wherever the task leaves a trace. A real count from the month you just had beats a guess.
For each one, write two more numbers. What one run is worth to you: the time it saves times what an hour of yours is worth, or the money a single run brings in. And what it costs you in a month, all in: the subscription plus a rough hour or two of your own time keeping it alive. (Did issue #25’s cost worksheet? You already have this number. If not, a rough estimate is fine.)
Open Claude Code and hand it the list. On a Mac open Terminal, on Windows open PowerShell, type
claudeand press enter. Then paste something like: “Here is a list of my automations. For each I’ll give you what one run is worth and what it costs me per month. Work out the break-even: how many times it has to run in a month to pay for itself. Put them in a table next to how many times each actually ran last month, and sort them into ‘pays back’ and ‘never will.’” Let Claude do the arithmetic and the sorting. That is the part you do not have to think about.Read the two piles. “Pays back” is safe, leave it running. “Never will” is the one to act on: kill it, drop it back to doing it by hand, or find a way to run it far more often. All three are fine. Paying to keep it as-is is the only wrong answer.
Keep the table. Point it at the next tool you are tempted to buy: ask how often you would really run it, before you pay. Stuck? Reply to this email. I’ll help.
WHAT’S COMING
Next issue: the same task priced three ways. An agent, a person, and not doing it at all. Why the option nobody bothers to cost is sometimes the right one.
Manu