Decide where the gains go

WHAT I’D DO

Before your next wave, ask the people whose routine work will be automated what worries them about it. Then name the new role each of them moves to.

TO MAKE THIS WORK

First: Have role-based training and protected time ready before any routine task is handed to a machine.

As you do it: Roll out one site or team at a time, with a named owner for each wave, and fix what blocked the last one.

Then: Track quality, cost removed and where every redeployed person ended up, not licences handed out or hours saved.

BY ROLE

Function heads: Name the next role for each person whose routine task an agent or camera takes over, before the switch rather than after it.

HR and L&D: Agree the retraining path for each affected role before the tool goes live, mixing classroom and on-the-job training.

Finance and risk: Review licence cost against active use regularly so unused seats move to people who will use them, and check each country's rules first.

BEHIND THE NUMBERS

Schick says the cameras catch far more defects than a person could, and that volumes and quality have both improved. Jabil's sites in Guadalajara and Chihuahua, Mexico, partnered with local universities to build what he calls "almost a Jabil University" for new hires and reskilling, a model now coming to the US. He calls fiscal 2026 the groundwork and expects a leap forward in fiscal 2027.

HOW SOLID IS THIS

An executive interview on a podcast. Schick gave no figures for quality gains or licence numbers ("I don't know" the exact count), and nothing was independently checked.

WHAT I’D DO

Pick one slow decision in your business and time it end to end this week. Write down where a person reviews it and who signs off, before anyone builds anything.

TO MAKE THIS WORK

First: Before building, check that the data and controls behind the decision are sound enough to trust an AI recommendation.

As you do it: Put a human review on any step that shapes a customer decision, and meter model consumption from day one.

Then: Run it with a small group of reviewers, track time saved and overrides, and agree with risk what proof scaling needs.

BY ROLE

Finance and risk: Set two review paths: a fast track for internal tools and a full model risk review for anything that shapes a customer outcome.

Technology and data: Publish guidance on which models fit which jobs and meter consumption. Sakthivel says developer assist tools don't need the most expensive models.

Executive sponsor: Ask what business problem each AI request solves before approving it. Sakthivel says the answer is sometimes an API, not AI.

BEHIND THE NUMBERS

Sakthivel says Manulife has publicly committed to realize a billion dollars Canadian of value from its AI initiatives by the end of 2027. He says the group is well on its way, with John Hancock contributing. He says the job is finding use cases big enough to show measurable value but small enough not to buckle under their own weight.

HOW SOLID IS THIS

Executive interview on a podcast recorded live at Ai4. The weeks-to-days result is company-reported, with no baseline or volume figures given. He says a press release exists, but neither it nor the figure has been checked independently.

WHAT I’D DO

Before any headcount decision tied to AI, put three numbers side by side for one team: what it produces, what reached customers, and hours spent fixing incidents.

TO MAKE THIS WORK

First: Give each team a named owner and a way to roll back changes, so problems stay small and get fixed fast.

As you do it: Set a limit on incident and fix time that pauses further rollout until the cause is found.

Then: Make operating reviews and incentives reward results that reach customers, and stop rewarding teams for how much AI they consume.

BY ROLE

Executive sponsor: Don't approve cuts to layers or headcount based on usage or output figures. Ask what reached customers and what it cost to fix.

Finance and risk: Count incident and cleanup time as a cost of AI-written work, next to compute and licences, when you work out the return.

Technology and data: Limit what agents can change without a human check. In April, Meta's infrastructure teams warned that unchecked agents were taking large-scale disruptive actions.

BEHIND THE NUMBERS

In scenario planning, executives explored cutting many teams by as much as 60%. A 10% cut went ahead. The incidents included service disruptions and possible data leaks. Zuckerberg told staff that when compute spending goes up, people-related spending has to go down. He also said they should not expect other company-wide layoffs this year.

HOW SOLID IS THIS

These are Meta's own internal data and posts, as reported by Reuters and read out on a podcast. They have not been independently checked.

WHAT I’D DO

Ask your sales managers which reps got higher-value work since AI arrived, split by performance quartile. If it's mostly the top, give each bottom-quartile rep one stretch assignment with a named coach.

TO MAKE THIS WORK

First: Find a few top reps who already build their own AI tools, and have managers protect their time to show others how they work.

As you do it: Teach on live deals with managers coaching in the moment, and decide in advance where the time AI saves will go.

Then: Roll the best home-built tools out to other teams in waves, fixing data and access problems before each team starts.

BY ROLE

Executive sponsor: Use AI visibly on your own work and say how you learned it, since managers named executive role modelling as a success factor.

People managers: Check whether better assignments keep going to the same reps, and give a lower performer one stretch task you coach them through.

HR and L&D: In hiring, ask candidates how they learn rather than whether they use AI, and build skills through practice, peer sharing and examples from power users.

BEHIND THE NUMBERS

Managers were 16 percentage points more likely to report quality improvements for top-quartile reps than for bottom-quartile ones. On a 1-to-5 scale, each one-point rise in AI literacy went with a 0.71-point rise in reported benefits, while simple trust in AI had little correlation with gains. At Brex, CRO Garrett Marker automated 20–40% of SDR manual work and automatically promoted SDRs who met their targets and closed a set number of full-cycle deals, which the authors say attracted stronger talent and increased internal promotions.

HOW SOLID IS THIS

Survey and interviews by two economists and the CEO of Terret, a firm that sells AI deployment help to revenue teams. The figures are self-rated and manager-reported, come from leading-edge AI adopters, are not yet released elsewhere and were not independently checked.