AI will eliminate 300 million jobs in the next 5 years.

Yours doesn't have to be one of them.

Here's how to future-proof your career:

Start learning AI now

Two numbers from the same study that should not be able to sit next to each other.

Eighty-seven percent of people use AI at work. Thirteen percent see a significant impact from it (Glean Work AI Index 2026, via Superhuman, July 5 2026).

Most leadership teams have a version of that gap on a slide somewhere, labelled adoption versus value, followed by a plan to drive more usage. More training, more licences, a champions programme. The assumption underneath is that the value is waiting on the other side of adoption.

Glean went back and looked for the missing hours. Six thousand digital workers. AI does save time, and then a large share of that saving goes straight back into cleaning up what it produced. They gave it a name: the botsitting cycle (Glean Work AI Institute, via Superhuman, July 20 2026).

So the tool works, and then somebody spends the afternoon babysitting the output.

In a CPG marketing org that cleanup has specific names, and every one of them is a real hour on someone's real calendar. Legal sends the copy back because the model wrote a benefit nobody can substantiate. Brand sends it back because it is technically accurate and sounds like no brand in particular. Someone re-briefs the agency because the input they were handed was AI-generated and about seventy percent right, which is the worst number there is. An analyst rebuilds the retail media summary by hand after finding one figure in it she cannot trace.

None of that shows up on the time-saved slide. It comes out of the same team's evenings.

Here is what makes this a management problem rather than a technology one. Clarinet measured the other side of it and found that people who are genuinely fluent with these tools reclaim about thirty percent of their time, and raise the quality of the work while doing it. Usage is not fluency (Clarinet, via Superhuman, July 17 2026).

The tax is not fixed. Some teams pay it in full. Some barely pay it at all. What separates them is where the checking happens.

I use a plain frame for this. Inputs, processing, outputs. If you are clear about your inputs and clear about what a good output looks like, it matters much less whether a machine or a person does the bit in the middle. Most companies dropped a model into the middle and left both boundaries exactly where they were. Which means the check still sits at the very end, after the work has already been handed to someone with no way of verifying it.

Anyone who has stood on a factory line knows you do not catch a bad batch at the loading dock. You put the check where the defect gets made, and you sample the rest.

Same discipline, applied to output. Three tiers.

Ships on its own. Internal first drafts, meeting summaries, reformatting, internal translation. No gate. If it is wrong, the person reading it can see that for themselves.

Ships on a spot check. Retail media reporting, PDP variants, brief summaries. One named person samples it. Nobody reads every line.

Never ships without a named human. Anything carrying a claim, a price, a consumer promise, or a regulatory exposure. Not a committee. A name.

The version with no gate at all has a price too. A vibe-coded app leaked 1.5 million API keys because nobody reviewed the code before it went out (Superblocks, via Superhuman, July 18 2026).

The commercial stake is simple enough. Your AI business case is almost certainly written on gross hours saved. Nobody lied when they built it. The cleanup was never counted, because it lands in a different team's week than the one that booked the saving. And AI spend inside a CPG is rarely new money. It comes out of an agency retainer, a research line, a martech renewal. You are trading a cost you could see for one you cannot, and the P&L finds out before the deck does.

One thing worth doing this week. Pick a single AI-assisted workflow. Ask the three people who actually touch the output how long they spend fixing it. Subtract that from what the business case claimed. If the net comes out negative, do not go shopping for a better model. Move the gate.

The hours are real. They are just being spent twice.

-- Imteaz

If this one landed, two earlier pieces on the same thread:

This one is for the leader whose AI business case is built on hours saved that nobody has gone back and checked. If someone forwarded this to you, subscribe here.

Or just hit reply and tell me one thing: which AI-assisted output in your team eats the most cleanup time on the way out the door? I read every one.