They didn't pick the wrong tool. They deployed it the wrong way. "Give everyone an assistant" carries an assumption nobody checked: that a company's work should be divided by person in the first place.
That assumption is exactly what an AI transformation should make you question again.
01Work got bundled into "one person" only because handoffs were expensive
Roles aren't a law of nature. A string of related tasks gets bundled together for one person to own for a very practical reason: chopping work into pieces and passing it between people used to cost so much in handoffs, alignment, and waiting for replies that it wasn't worth it. Easier to give the whole bundle to one person to do from start to finish. Middle management is an extension of the same logic: assign the work, chase progress, and pull scattered information together into a report.
Once AI starts taking over execution item by item — first drafts, running numbers, meeting notes, draft replies can all be handed off — the cost premise behind "it has to be bundled" starts to loosen. A role goes from a stable package of duties to a task list you can rearrange.
So the question to ask was never "will this role be replaced." It's these three:
- Within this role, which work should AI do first?
- Which calls must a person make?
- Which results must a person check before they move on?
The order can't be reversed. Most companies answer only the first one and think they've finished their AI transformation.
02Take one role and do it on paper
Talking in the abstract doesn't help. Pick the busiest role in the company and take it apart on a sheet of paper.
Take an e-commerce operations person. Their day used to come down to roughly five things: reviewing data, writing product copy, scheduling push notifications, watching competitor prices, and writing the weekly report.
Run those five against the three questions.
Reviewing data — AI goes first; it pulls numbers faster than a person and doesn't miss any. But "conversion dropped this week: is it the redesign or a competitor slashing prices?" is a judgment call. A person decides.
Writing copy — AI goes first and produces three versions at once. But "can this line go out as our brand's voice?" A person decides.
Scheduling pushes — AI goes first. But whether the send list and timing will annoy customers, a person checks before it goes out.
Watching competitors — AI goes first, and it should run every day. A person only gets notified when a price band is broken.
Writing the weekly report — AI goes first. A person only adds one paragraph: "So here's what we change next week."
Once it's taken apart, you see two things. About sixty percent of their day went to the part AI can take over right now. And the remaining forty percent isn't scraps — it's the most expensive part of the whole role: judgment, trade-offs, and commitments to the outside world.
That's what recutting work means. Not replacing them, but lifting them out of that sixty percent so they can do the forty percent full-time. Their job title can stay exactly the same, but the work in that seat has already been redrawn.
One line is worth drawing up front: prices, contracts, commitments to customers, account permissions, and final publishing — no matter how strong AI gets, these five should keep a person's final confirmation. Not because AI can't be trusted, but because the cost of getting these five wrong is in a different league from everything else.
03The work got split up. Accountability didn't get rebuilt.
Splitting tasks is the easy part. What actually goes wrong is that after the split, nobody answers "who's responsible."
Klarna is the most often cited case. In 2024 the company announced, loudly, that its AI customer service had taken on the workload of 700 full-time agents, and it stopped hiring. A year later, CEO Sebastian Siemiatkowski publicly reversed course and started hiring human agents again, saying the cost-cutting had gone too far and service quality had paid the price.
This is often read as "see, AI still doesn't work." What it really shows is something else: the company did a one-for-one swap — people replaced with AI, costs on paper went down — but nobody answered "when a customer gets angry, who picks it up, who has the authority to make an exception, who owns that decision." The work got moved. The accountability was left hanging where it was.
Deloitte's 2026 research puts a number on it: 84% of companies haven't redesigned their roles for AI. The vast majority slot AI into the existing org chart and expect the organization to grow new lines of accountability on its own. It won't.
The check is simple: for everything AI takes over, you should be able to find a named person who answers "if it goes wrong, who notices, and who fixes it?" If you can't find that name, it isn't ready to hand off yet.
Only once you can check it can it actually work for you.
04Replace, augment, restructure — which level are you on?
Back to the company that handed out AI accounts to everyone. The place it got stuck actually has a name.
AI transformation comes in roughly three levels.
Level one is replacement: treat AI as a cheap employee, swap out a seat one-for-one, with saving money as the only goal. Klarna walked through this level, and stumbled on it.
Level two is augmentation: everyone gets an AI tool, but not a word of the workflow changes. Each person gets a little faster, yet the company as a whole doesn't get better, because the way work is handed between people hasn't changed and the jammed intersections are the same ones as before. Most of Deloitte's 84% are stuck here. So are the companies that handed out AI accounts to everyone.
Level three is restructuring: what gets redrawn isn't headcount, it's how the work flows. Which work AI does first, which calls a person makes, which results a person checks — every workflow gets questioned again, and only then do you decide how many people you need and where they go.
The gap between the levels isn't skill. It's whether you're asking the right question. Level one asks, "Can this person be swapped for AI?" Level three asks, "Who should do this piece of work?"
There's a useful historical comparison. In the decades when factories switched from steam engines to electric motors, the first plants to swap engines saw almost no gain in output. Steam-era factories were built around a single central drive shaft, so machines had to be crowded near it. Electric motors let every machine have its own power, but as long as the factory layout stayed the same, the advantage never showed. The real leap came on the day someone redrew the factory.
AI is the new engine. Handing out accounts to everyone is swapping the engine and keeping the old factory.
05Start with one role. You don't have to touch the whole org chart.
Restructuring doesn't have to wait for an organization-wide overhaul. It can start in an afternoon.
Pick the role in your company that's busiest and makes the fewest external commitments. List a week of its work, and mark each item "AI goes first / a person decides / a person checks." Once it's marked, three things usually show up: a batch of work you can hand off right away, a batch that needs its data cleaned up before it can be handed off, and two or three blanks where it turns out nobody was ever really responsible.
Those blanks are worth more than any set of tools.
Finish one role, then do the next. It's slower than buying a set of "AI employees," but it's slow in the right place: every step leaves behind a division of work you can actually write down, rather than a bill whose effect nobody can explain.
The hard part isn't the splitting. It's admitting that some work was never properly defined in the first place. Thinking this through on your own, it's easy to circle back to the org chart you already have, so it's worth having someone from outside take it apart. That's exactly what OAKOrc's AI workflow diagnostic does: pick the one or two roles that hurt most in your company, lay the work out on the table and recut it, mark the decision points and who's accountable, then tell you which box to start with.
People belong at the point where the call gets made
AI transformation is easily misread as headcount arithmetic: how many people it saved, how many seats it replaced.
But once you lay the work out and recut it, you find the real gain isn't in the headcount column. It's that the people who were always buried in busywork, with no time to make judgment calls, can finally make only judgment calls — and the company knows, for the first time, who stands behind each decision.
Roles shouldn't be cut per person anymore, and not because people don't matter. Quite the opposite: the most valuable part of a person has been buried under the least valuable work all along.
Let AI move fast. Let people hold the anchor.
Rooted in craft. Built for the new wild.