Two presidents stood on the same Scottsdale stage on 1 June and told Fortune’s COO Summit two opposite things about how to redesign work for AI. Both were right. Most of the rooms watching are auditing for neither of the ways each approach quietly fails.

The split matters because the senior operating-officer class has finally stopped pretending there is one playbook. There is not. There are two, they cut in opposite directions, and the diagnostic question is which failure mode you are already living with.

The two camps, fairly stated

Okta’s President and COO Eric Kelleher made the top-down case. His argument is that the bottleneck is not the model, it is the manager. “We have trained every manager in the world to think about one thing and that is: what’s their headcount,” he told the room, before pitching a shift from “workforce planning to work planning.” His mechanism is concrete. Push token budgets down to people managers, give AI agents names, put them in business reviews. His own team has four: Leo, Sloan, Hank, Walker. They show up alongside human staff in the operating cadence. Kelleher’s framing: “It’s really uncomfortable, but it’s very transformative.”

Wayfair’s President Jon Blotner made the opposite case on the same stage. Wayfair had a top-down AI mandate and reversed it. Every employee got access to Claude, Gemini, and ChatGPT. Teams self-redesigned. Blotner’s read: “We see people reinvent their jobs and say, okay, look, I basically automated my work. That person’s incredibly valuable.”

Both stories are real. Both are working inside their respective companies. This piece is not going to call a winner.

The Cognizant data shared around the same conversation is worth holding next to it. $4.5 trillion of US labour value is exposed to AI, 93% of jobs are affected, and that is from a reassessment of 18,000 tasks in the O*NET database. The surface area is too large for either approach to be the right one on principle. Both have to work. The question is whether yours is breaking in the way its own design predicts.

The failure mode hiding in the top-down play

Token budgets fail the same way headcount budgets failed for forty years. Managers protect them.

Once a token budget becomes a line in the operating review, it becomes a thing to defend, justify, and grow. The same manager who would have padded a hiring plan to keep an org-chart box now pads an agent budget to keep a fiefdom. The agent gets named. It gets a seat in the review. It does not necessarily do useful work, because nobody is asking the harder question Kelleher actually wants asked, which is what work needs to exist at all.

The transformative version of “work planning” requires managers to argue that work should disappear, not that more of it should be budgeted to a digital colleague. That is the muscle no operating cadence currently rewards. Ship token budgets without changing what gets promoted and you have built a more expensive version of headcount politics with better branding.

The audit question for a top-down shop is simple. In the last quarter, did any manager propose retiring work entirely? Or did every proposal route token spend toward preserving the team’s current scope?

The failure mode hiding in the bottom-up play

Universal access fails when the wins stay personal.

Blotner is right that the employee who automates their own job is incredibly valuable. The problem is that the value is locked inside that person’s head, their prompts, their custom workflow, their Claude project. When they leave, it leaves. The institutional capability never accrues. The org learns nothing it can put on an onboarding doc. Two years of distributed self-redesign produces a thousand local optimisations and zero codified ones.

This is the bottom-up version of the trap that killed corporate knowledge management in the 2000s. Individuals get better. The organisation does not. The second cohort of hires arrives, gets the same tools, and reinvents the same automations the first cohort already built, because nothing got written down or built into the system of record.

The audit question for a bottom-up shop. Of the automations your top quartile of employees built last quarter, how many are running today as the default workflow for someone who did not build them? If the answer is “we don’t track that,” that is the answer.

The actual diagnostic

Pick the failure mode that scares you less. Then instrument hard against the other one.

If you run a Kelleher shop, instrument the rate at which work is retired, not the rate at which agents are deployed. Track promotions of managers who shrank their scope by automating it out, against managers who grew their token budget. If nobody got promoted for shrinking, your top-down redesign is headcount politics in a new costume.

If you run a Blotner shop, instrument the codification rate. For every employee-built automation that demonstrably saves time, measure how long it takes to become the default workflow for the team, and then for the function. If that number is “never,” your bottom-up redesign is an individual productivity programme, not an org redesign.

Both leaders on that stage have done the hard part already, which is admitting the work itself has to change. Most of the companies reading the coverage have not. The trap is reading Kelleher and rolling out token budgets without changing the promotion criteria. Or reading Blotner and rolling out universal access without building a codification function. Either move is theatre.

If you are picking your camp this quarter, talk to us first about which failure mode you are already running.

Sources