THE NUMBER
Ninety-four percent.
That is how many routine HR requests IBM's assistant now resolves on its own. The interesting number is the other one. The remaining 6%, the cases calling for ethical judgment, still need a person - and IBM responded by tripling its US entry-level hiring. It automated almost everything, then hired more people.
3 THINGS HAPPENING RIGHT NOW
Nearly a third of managers who cut a job for AI have hired someone back
Robert Half surveyed nearly 2,000 US hiring managers. Roughly a third of those who cut a role because of AI have since rehired for that role or a similar one. The first estimate of what the software could carry was too high, and the bill arrived as a job posting.
Ford brought back 350 veteran engineers to fix what automation missed
Over three years Ford hired, promoted or rehired 350 experienced engineers, its "gray beards," to mentor younger staff and solve quality problems automated systems could not. "Artificial intelligence is a fantastic tool, but it's only as good as the information you use to train it," said Ford VP Charles Poon.
A founder shut down his own AI product and said why
Last November, Joe Braidwood closed Yara AI, the therapy app he built. Low thousands of users, and he turned down a fund that wanted to invest. He judged the software fine for everyday stress and dangerous for someone in crisis. "Sometimes, the most valuable thing you can learn is where to stop."
THE DEEP DIVE
The last six percent is not leftovers
The obvious way to read 94% is that the rest is mopping up, and next year's version finishes it. IBM's own account says otherwise: what remains is the cases needing ethical judgment. Not the same work in smaller quantity. Different work that was always mixed in, which the automation sorted out by failing at it.
Which leaves a problem most people miss. The routine 94% was where junior people learned. You handled a hundred easy cases badly, then adequately, then well, and somewhere in there became the person who could handle the hard ones. Automate the easy hundred and the hard six still arrive, but the pipeline that produced someone able to answer them is gone. IBM's HR chief, Nickle LaMoreaux, put it plainly: "If we don't continue to invest in entry-level hires, what happens in three to five years? There's no pipeline."
ONE THING TO TRY THIS WEEK
Turn one of them off and find out if it was doing anything.
You cannot judge a system from the inside: everything it produces looks like output, and no complaints looks like success. The only clean test is removing it. You also have a smaller version of IBM's problem - some of what you handed off was how you learned the thing.
List what runs without you. The scheduled report, the auto-reply, the morning summary, the reminder that goes out. Everything that happens whether or not you show up.
Cross off anything that is insurance. Backups, security scans, anything whose job is to matter on the day something goes wrong. Those are not idle, they are waiting, and a quiet week proves nothing about them. From what is left, pick ONE that is reversible and does not go to a customer. If everything you have touches customers, pick the one you could do by hand for a week.
Write down what you expect to happen without it, and when you would notice. "Nobody will miss the Monday summary" is a prediction. So is "I will feel it by Wednesday."
Turn it off, and do not announce it. Leave it off for one full cycle: a daily thing gets a week, a weekly thing a month, a monthly thing a quarter. If you are not willing to wait that long, pick something that runs more often - a short look at a slow process tells you nothing. If someone relied on it, you want to hear that from them, not from a warning you sent.
Compare what happened to what you wrote. Somebody complained, you had to do it by hand, or a number moved: it was earning its keep, and now you know what it is worth. Nothing happened at all: turn it off for good and stop paying for it. Not sure either way? That is a result too - it was never doing enough for you to tell.
Most of what we automate is never looked at again after the day we build it. It keeps running, and running gets mistaken for working.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
That closes four weeks on checking the machines you built. Next issue we start on the part nobody measures: what one of these things actually costs you per task, once you count your own time, and how to tell which ones have never been worth it.
Manu