RETAIN Series 5: Exhaust the Ladder Before the Exit: The Foundation Comes First

In brief:

• A cost ladder needs a solid foundation – a hiring gate that makes every new role hard to justify, both in good and bad times. Meta (Nov’22) and Google (Jan’23) each cut thousands of jobs after growth bets that their own CEOs later tied to a different economy.
• Google’s hiring system, as Laszlo Bock describes it, put several independent layers of review between a hiring manager and an offer, ending with the CEO. The same design can be applied to headcount requests.
• AI does not remove the need for the foundation. It gives everyone extra capacity, and someone still has to judge the output. The record on AI-driven layoffs is already mixed.
• The rungs of the ladder follow in the next two articles.

Exhaust the Ladder Before the Exit: The Foundation Comes First

I ended Post 4 with a question about reviews built around a decision already made. This article goes back to the weeks before that, when the decision exists but nobody has made it public. What should a company try before it asks anyone to leave? The answer is a ladder of cost levers, applied in a set order, with layoffs at the very end.

A ladder is only as sound as the ground it stands on. This is the first of three articles on E, and it is dedicated to covering that ground. The rungs follow in the next two articles.

The cheapest layoff is the one you never need

Some of the biggest recent tech layoffs followed growth bets that didn’t pay off. In Nov’22, Mark Zuckerberg wrote that he had expected the pandemic surge in online activity to last, and had raised investment sharply. The surge did not last. “I got this wrong, and I take responsibility for that,” he wrote. Meta cut more than 11,000 people, ca. 13% of its staff.

In Jan’23, Sundar Pichai told Google staff that the company had “hired for a different economic reality” than the one it now faced. Google cut 12,000 jobs.

I am not claiming that hiring alone caused these rounds of cuts. Each CEO tied the two together in his own words.

The foundation is critical to breaking that pattern. The rule is very clear: hire less and hire better. Make every new role hard to justify, both in good and bad times, regardless of the mood of the stakeholders. Stick to documented justification and hiring processes no matter how tempting it is to cut corners. When you finally get the approval, hire someone better than yourself. Fewer but stronger hires leave a smaller base to cut from later.

What Google’s hiring system teaches

In “Work Rules!”, Laszlo Bock describes how Google hired back in the days when it was adding about five thousand people a year. Four things are especially important here.

A standard-keeper at the top. Bock says that the founders had the instinct to enforce an objective standard through “a single, final, central reviewer who is charged with upholding that standard.” Google later split that job across two teams of senior leaders. CEO Larry Page became the final reviewer of every single candidate. He received a weekly report with links to each full hiring packet and the recommendation from every earlier layer. Bock says the message mattered as much as the feedback: hiring is taken seriously at the highest levels.

Why the standard-keeper matters: standards dilute. Bock’s argument is that each generation of hires drifts a little below the one before, because new people do not share exactly the same picture of the standard. As the company grows, he adds, so does the temptation to hire a friend or a customer’s child as a favour. He calls those hires almost always a compromise of quality.

Deciders who do not own the outcome. Even well-meaning managers lower their standards as the search drags on, Bock says. His example is a search for an administrative assistant: a very high bar on day one, and by day ninety most managers will take anyone who answers the phone. So Google took the final call away from the hiring managers. Hiring-committee members step aside on candidates for their own teams. Each candidate also meets an interviewer from another function, chosen because they have little or no stake at all in filling the role. New managers resented this at first but came to appreciate it after about six months.

Results, as Bock reports them. He says even the Googlers who complained about the slow process agreed it produced remarkable hiring quality. At Google’s April 2013 board meeting, Intel CEO Paul Otellini praised the system as one that scales with the company and improves with each generation.

Applying the same design to headcount

Headcount requests are easier to review and compare than candidates. There are fewer role types and many of them are standardized, which helps improve the processing speed.

Most large companies already have automated systems in place to stop requesters from approving their own roles and block offers for unapproved ones. The design I propose differs in who signs, what evidence they see, and what is checked a year later. It has five parts:

1. Independent evidence. Someone outside the team supplies workload data, backlog, and what has been tried before filing the request. Otherwise the reviewer will be limited to the requester’s side of the story only.
2. Senior owners who decide. Every new role goes to the C-suite, with a counter-signature from a peer in another function. Each signer reviews the evidence and answers for the decision. Passing along the recommendation from below does not count. A peer who reviews a request from their own organization steps aside, as Google’s committee members do.
3. A standard form. The same content every time: role type, cost, business case, what had been tried before the request was filed, the expected measurable outcome, and the date it will be verified. Standard roles allow standard benchmarks. Only exceptions need longer conversations.
4. Contractors count. A contractor or vendor seat is headcount for this purpose.
5. A verification tied to pay. At 12 months, someone compares what was promised with what got achieved. The result counts in the remuneration of requesters and approvers, both short- and long-term. Incentives are a pillar on their own (the A in RETAIN), and will be covered in a later article.

What it costs

Senior review takes senior time. Until Google reached about 20,000 people, its top executives spent a full day a week on hiring. By 2013, with about 40,000 people, the average Googler spent 1.5 hours a week on it, and Bock says the time per hire fell by about 75%. Those figures are about reviewing candidates. A headcount gate should take less senior time: one decision covers every candidate for that role.

Gates that already exist

Parts of this methodology are already in use:

• In Jul’22, Google paused the majority of new offers for two weeks. It said it would use the time to review the headcount needs and set a prioritized list of staffing requests for the next three months.
• In May’23, in a leaked memo, AWS made it mandatory for hiring managers to route all potential hires through an internal system called Roster. By the end of that month, a recruiting tool was to block offers for requisitions not matched to an approved position.
• In Apr’25, Shopify’s CEO Tobi Lütke told his teams they must “demonstrate why they cannot get what they want done using AI” before asking for more headcount.

Things to watch out for – gaming the gate

Requests may get inflated or split. Work may get relabelled. Requests may come in early, to claim a slot.

Offloading the work to contractors is a common practice. Political scientist Paul Light argues that the US federal cap of about two million employees pushes agencies to move work to contractors and grantees, whom he puts at roughly nine to eleven million in recent years. In other words: when agencies cannot add employees, they add contractors instead. Google’s numbers tell a similar tale. An internal count seen by the New York Times in Mar’19 showed 121,000 temps and contractors against 102,000 full-time staff. That is why the contractor rule exists.

Peer counter-signatures can turn into mutual favours. Rotating the counter-signer and keeping a record of who signed what help surface any potential collusion. A signer who never says no is a rubber stamp. Tracking each signer’s rejections would show it.

Hiring the best, with care

The case for hiring the strongest people has strong grounding in research.

Performance is concentrated. O’Boyle and Aguinis (2012) studied 198 samples and 633,263 people, from researchers to athletes. In 94% of the samples, output was lopsided. A small group at the top produced a large share of the total, instead of most people clustering around the average, as a bell curve would predict. In their researcher data, the top 1% produced about 10% of the output and the top 5% about 26%. An even spread would give them 1% and 5%. Hunter, Schmidt and Judiesch (1990) found that the spread in output across workers grows with job complexity. For non-sales jobs it was about 19% of average output in low-complexity roles, 32% in medium and 48% in high.

Bock quotes Google’s SVP Alan Eustace saying a top engineer is worth 300 times or more than an average one. That is, however, an executive’s claim, not a measured result, so I’d take it with a pinch of salt.

Another thing to keep in mind is that stars do not always travel. Groysberg, Lee and Nanda (2008) followed star analysts for nine years. Hiring them often did not pay off for the stars themselves and the firms that hired them. The authors point to company-specific resources, such as colleagues and capabilities playing a critical role in how a star performs. Bock shares a similar warning: being a star in one setting does not make you one in another.

So, to summarize: the hiring rule has two parts. Hire better than yourself and make sure it’s a good fit for the environment you’re hiring into. It also takes patience. Bock encourages slowing down the hiring, since at best 10% of applicants are top performers and the best are often not looking for new jobs.

What AI changes, and what it doesn’t

People often bring up two arguments. One says AI lets a small, strong team do far more. The other says AI favours experts. Both feel intuitively correct but the evidence is mixed.

Early studies often show AI helping less experienced people more. In one customer-support study, the average gain was 14% – novices gained 34% while experienced workers saw minimal impact (NBER working paper). BCG consultants in the bottom performance half gained 43% and those in the top half 17%. A field experiment with 640 Kenyan small-business owners found no average effect. The strongest businesses gained 15% or more. The weakest lost about 10%. The key thing the researchers pointed out is that stronger owners were better at judging the advice AI gave them. METR found experienced developers 19% slower with AI tools in 2025, but the outcome of its 2026 follow-up study was less clear.

I will therefore not claim that AI further amplifies the performance gap between top and average performers. It is safer to agree that AI gives everyone extra capacity and that someone needs to be there to assess the output. What a company does with that capacity is a leadership choice: reinvest it in growth or bank it as cuts.

The record of AI-driven cuts is already mixed:

• Orgvue’s 2025 survey of 1,163 senior leaders found that, among those who had made AI-related redundancies (39%), 55% admitted their decisions were wrong.
• Robert Half, as reported in Jul’26, found that 32% of nearly 2,000 US hiring managers had scrapped a role mainly because of AI, then brought the work back soon afterwards.
• Careerminds, an outplacement vendor, found that 35.6% of 600 HR professionals had rehired more than half of the roles they cut, and 30.9% spent more on rehiring than they had saved.
• Commonwealth Bank of Australia cut 45 service jobs in Jul’25 and reversed the cuts in August the same year. It called the decision an error.
• Forrester forecasts that half of AI-attributed layoffs will be “quietly reversed,” often offshore or at lower pay.

One more thing is of immense importance for the R in RETAIN. Employees who fear for their jobs tend to perform somewhat worse. A 2019 meta-analysis by Sverke and colleagues pooled 53 studies (about 21,500 employees). It measured how closely job insecurity and task performance move together, on a scale from -1 to +1, where 0 means no relationship. The result was about -0.14 (-0.17 after correcting for measurement error). That is a small but consistent link: the more insecure people feel, the lower their performance tends to be. On its own it accounts for only a few percent of the differences in performance.

Two important notes. First, this is a correlation, not proof of cause. Insecurity may lower performance, but weaker performers may also feel less secure. Second, the ten studies that followed the same people over time found a link of only -0.04, which, for practical purposes, cannot be told apart from zero. The same goes for creativity (-0.10, ten studies). So the effect is modest, and the direction of cause is not settled. A clear bar helps, but a bar alone does not make a job feel safe in a downturn. The foundation adds to that safety. And then come the rungs of the ladder.

Next in this series: E, part 2: the first three rungs.

Sources

• Zuckerberg, Nov’22 letter (and Meta’s 11,000 cuts): The Drum, November 2022
• Pichai, Jan 20, 2023 (hiring, 12,000 cuts): Google blog; also Ars Technica, January 20, 2023 (mirror)
• Bock, Work Rules! (2015), chapters 3 to 5: Michal’s copy (the time figures are in chapter 5, p. 112). Page’s weekly review is also reported in Business Insider, via Yahoo, April 2015
• Google’s Jul’22 hiring pause: Android Central, July 21, 2022
• AWS Roster memo: Fortune, May 2, 2023
• Shopify memo: TechCrunch, April 7, 2025
• Paul Light on the federal workforce ceiling: Brookings, 2020
• The Mar’19 count (121,000 vs 102,000): New York Times reporting, via Inlander (NYT News Service) and MIT Technology Review
• O’Boyle and Aguinis, “The Best and the Rest,” Personnel Psychology, 2012: paper
• Hunter, Schmidt and Judiesch, 1990, output variability and job complexity: abstract
• Groysberg, Lee and Nanda, Management Science, 2008: INFORMS release, July 29, 2008
• Support-agent study: Brynjolfsson, Li and Raymond, NBER Working Paper 31161
• Kenyan entrepreneurs: Columbia Business School
• BCG study and METR trials: Digital Applied, “Who gains most from AI at work” (secondary summary)
• Job insecurity and performance: Sverke, Låstad, Hellgren, Richter and Näswall, “A Meta-Analysis of Job Insecurity and Employee Performance: Testing Temporal Aspects, Rating Source, Welfare Regime, and Union Density as Moderators,” International Journal of Environmental Research and Public Health, 2019, 16(14):2536: paper on PubMed Central
• Orgvue survey: Orgvue press release, April 29, 2025
• Robert Half survey: IBTimes UK, July 2, 2026
• Careerminds survey: Careerminds, “The cost of AI layoffs”
• Commonwealth Bank reversal: Information Age, August 2025
• Forrester forecast: Forrester, Predictions 2026