Salesforce Stopped Hiring Engineers and Credited AI. The Productivity It Bet On Was Never Measured.
Salesforce said it would hire zero new engineers this year because AI made the team more productive. A year on, Forrester finds 55% of employers regret their AI-driven cuts and many are quietly rehiring. The gap between the two is a productivity claim that was asserted at headcount scale and never actually measured. Here is what measuring it would have required.
Salesforce said it would hire zero new engineers this year. The reason given was AI. In the company's own framing, engineering had reached higher productivity, so headcount could stay flat. Marc Benioff put the broader version of it more bluntly when describing cuts elsewhere in the business: I need less heads.
That is a clean, confident story, and it is the story a lot of companies told through 2025 and into 2026. AI made the people we have more productive, so we need fewer of them.
A year later the other half of the story is arriving. Forrester reports that 55% of employers regret their AI-driven layoffs. Companies are quietly rehiring roles they cut, often offshore or at lower wages, and by one survey a third of them spent more putting people back than they saved by removing them. The productivity that justified the cut turned out to be harder to bank than it was to announce.
What Was Actually Claimed
It is worth separating the pieces, because they get blurred in the headlines.
One piece is customer support. Salesforce reduced its support organization from around 9,000 to about 5,000, crediting its Agentforce AI agents with handling enough case volume that it no longer needed to backfill those roles. That is a specific, partly measurable claim: case volume handled by agents versus humans.
The other piece, and the one that matters for anyone running an engineering org, is the productivity claim. Salesforce said it would hire no new engineers this year and framed it as higher productivity allowing flat headcount. That is not a claim about case volume. It is a claim that the existing engineers, augmented by AI, now produce what a growing team used to produce.
That second claim is the one worth examining, because it is the one thousands of other companies are quietly making when they freeze engineering hiring and point at AI. And it is the one that almost nobody has the data to back.
The Claim Underneath the Headcount Decision
Strip the language away and the decision to hold engineering headcount flat because of AI contains a precise, quantitative assertion: AI has increased output per engineer by enough to cover the work the additional hires would have done.
That is a number. It has to be, for the decision to be rational. If you are not hiring the ten engineers your growth would normally require, you are implicitly saying AI has added at least ten engineers worth of output to your existing team. You may never write it down that way, but that is the claim the headcount plan is resting on.
And here is the uncomfortable part. In almost every case that number was never measured. It was inferred from the general feeling that the team is moving faster, from vendor productivity statistics, and from the very visible fact that engineers are now shipping code with AI assistance all day. None of those are the same as measuring that output per engineer actually rose by the amount the headcount decision assumes.
Pro Tip: If your hiring plan assumes AI is covering the work of roles you are not filling, write down the number. How many engineers worth of output is AI adding, specifically? If that figure came from a feeling or a vendor slide rather than your own delivery data, you have an assumption sitting where a measurement should be.
Why This Is a Measurable Claim That Nobody Measures
Output per engineer is not a mystical quantity. It is measurable, and the measurement is not exotic. Merged PRs per engineer, cost per merged PR, throughput per dollar, lead time from commit to merge, all of these can be tracked before and after AI, against a baseline, and compared.
So why does the claim get made without the measurement? Because the claim is available immediately and the measurement takes work. The feeling that the team is faster arrives for free, the moment engineers start using the tools. The number that would confirm or deny it has to be built by joining AI spend to delivered work, and most orgs never built that. So when a headcount decision has to be made, the feeling is the only input in the room, and the feeling says yes, we are more productive, so the decision gets made on it.
This is the same trap that shows up when a team reports that AI saves each engineer two hours a day. The feeling is real and the number behind it is missing, and a randomized trial has already shown that the feeling of being faster can sit right next to the measured reality of being slower. Betting a support budget on that gap is expensive. Betting your engineering headcount plan on it is a different order of risk.
The Regret Data Is What Betting on a Feeling Looks Like at Scale
The Forrester finding is not really a story about AI being weak. It is a story about decisions made without measurement, and what happens to them when reality shows up.
Fifty-five percent regret is what you get when a large number of companies act on the same unmeasured assumption at the same time. Some of them genuinely did get the productivity gain and their cuts will hold. Others did not, and they are the ones rehiring, discovering that the work the departed people did was not actually being absorbed by AI, only assumed to be. The regret rate is high precisely because nobody had the number that would have told the two groups apart before they cut.
Klarna is the case study everyone points to, having replaced a large chunk of a function with AI, watched quality fall, and brought humans back. But Klarna is not special. It is just early and public. The quiet version is happening across the industry: a headcount decision justified by AI productivity, made without measuring AI productivity, corrected a few quarters later at a cost that often exceeds the original saving.
How Would You Actually Know AI Replaced N Engineers?
You would measure output per engineer against a baseline, the same way you would validate any other productivity claim big enough to reorganize a team around. The steps are not complicated, they are just work that has to be done before the decision, not after.
What a Measured Headcount Decision Requires
- A pre-AI baseline · Merged PRs per engineer, cost per merged PR, and lead time from a stable period before AI adoption. Without a baseline there is nothing to measure the gain against, and the claim is unfalsifiable by construction.
- The same metrics after adoption · The identical measures once AI is in real use, tracked over a period long enough to be stable rather than a good week. The delta is the actual productivity change, in the unit the headcount decision depends on.
- Quality held constant · Change failure rate and correction rate alongside throughput, so you are not counting a rise in PRs that later break as a productivity gain. Faster and worse does not replace an engineer.
- Translation into headcount · The measured output-per-engineer gain, converted into how many roles of capacity AI actually added. That is the number the hiring freeze is claiming, made explicit and checkable rather than assumed.
Run those four and the decision is defensible in either direction. Skip them and you are not making a productivity decision, you are making a cost decision wearing a productivity story, and the regret data suggests those two get told apart eventually.
Asserted Productivity Versus Measured Productivity
| Asserted | Measured | |
|---|---|---|
| Source | Feeling, vendor stats, visible AI usage | Delivery data against a baseline |
| Available | Immediately, for free | After the measurement is built |
| Unit | Vague sense of faster | Output per engineer, cost per merged PR |
| Falsifiable | No | Yes |
| Survives a bad quarter | Becomes regret | Holds or corrects on evidence |
| Safe to set headcount on | No | Yes |
Both columns describe the same belief that AI made the team more productive. Only one of them is safe to reorganize a company around.
Key Takeaways
- A hiring freeze credited to AI contains a hidden number. Not filling the roles growth would require is an implicit claim that AI added that many engineers worth of output. For the decision to be rational, that number has to exist.
- In most cases the number was never measured. It was inferred from the feeling of moving faster, vendor statistics, and visible AI usage, none of which is the same as measuring that output per engineer actually rose by the assumed amount.
- Forrester reports 55% of employers regret their AI-driven layoffs. That regret rate is what a large number of unmeasured decisions look like when reality arrives, with companies quietly rehiring and often spending more to restaff than they saved.
- Output per engineer is measurable and ordinary to measure. Merged PRs per engineer, cost per merged PR, lead time, and change failure rate against a pre-AI baseline turn the claim from a feeling into a checkable figure.
- Unmeasured, it is a cost decision wearing a productivity story. Measured, it is defensible in either direction. The difference is work done before the decision, not after the regret.
Why the Claim Gets Made Anyway
The pull toward the unmeasured version is strong, and it is worth being honest that it is not stupidity. The AI productivity story is easy to believe because part of it is visibly true. Engineers really are using the tools all day, code really is getting written faster, and the general sense of acceleration is real. When every signal you can see points to more productivity, the leap to we need fewer people feels less like a bet and more like an observation.
The pattern I keep running into is a leader who is not naive about this at all. They will say the team is clearly more productive and then, if you push a little, admit they could not actually tell you by how much. They can feel the gain. What they cannot do is size it, and headcount decisions are about size. So the decision gets made on the part they can feel, because the part they would need to measure was never built, and it holds up right until a quarter arrives where the assumed capacity was not really there.
The companies that will avoid the regret are not the ones who refused to believe AI helps. It clearly does. They are the ones who measured how much before they bet payroll on it, so that when someone asked whether AI had really replaced the roles they were not filling, there was a number on the table instead of a feeling. With that number, holding headcount flat might be exactly right. Without it, you are finding out the expensive way, one rehire at a time.
AI made your team more productive is not a conclusion. It is a hypothesis with a number attached, and the number is measurable. The only question is whether you measure it before the headcount decision or after the regret.
· Vukasin
How Yardstick Measures the Productivity Behind the Claim
Yardstick joins AI coding spend to delivered work from your Git provider and shows output per engineer, cost per merged PR, and lead time against a pre-AI baseline, updated daily. It turns AI made us more productive from a feeling into a figure you can actually put behind a headcount or budget decision.
Because the data is attributed at the engineer and PR level and tracked over time, Yardstick shows not just whether output rose but by how much, which is the number a hiring plan credited to AI is implicitly claiming. That is the difference between a defensible decision and an assumed one.
For leaders being asked to justify headcount plans with AI productivity, Yardstick provides the measured version of the claim, so the decision rests on delivery data rather than on a sense of acceleration that a bad quarter can turn into regret.
To see the platform or join the early access list, visit yardstick.fi/platform or review pricing for current plan options.
FAQ
Did Salesforce really stop hiring engineers because of AI?
As reported in 2025, Salesforce indicated it would hire no new engineers that year and framed it as AI-driven higher productivity allowing flat engineering headcount. Separately, it reduced its customer support organization from roughly 9,000 to about 5,000 roles, crediting its Agentforce AI agents, with CEO Marc Benioff summarizing the logic as needing fewer heads. The engineering claim and the support claim are different, and the engineering one is a productivity assertion rather than a case-volume one.
What did Forrester find about AI-driven layoffs?
Forrester reported that 55% of employers regret their AI-driven layoffs, with many quietly rehiring roles they had cut, often offshore or at lower wages. Separate surveys found a large share of companies rehired more than half the cut roles within months, and that a meaningful fraction spent more on restaffing than they originally saved. The pattern indicates many cuts were made on productivity assumptions that did not hold.
Why is freezing engineering hiring an unmeasured claim?
Because not filling roles that growth would normally require implies AI has added an equivalent amount of output to the existing team. That is a specific quantity, and in most cases it was inferred from the feeling of moving faster and from vendor statistics rather than measured from the company's own delivery data. The decision assumes a number that was never actually calculated.
How do you measure whether AI increased output per engineer?
Establish a pre-AI baseline of merged PRs per engineer, cost per merged PR, and lead time, then measure the same metrics after AI adoption over a stable period, holding quality constant with change failure rate and correction rate. The delta is the real productivity change, which can then be translated into how many engineers worth of capacity AI actually added.
Is this an argument against using AI or reducing headcount?
No. AI clearly increases productivity for many teams, and a headcount decision based on a measured gain can be entirely correct. The argument is against making that decision on an assumed gain. Measured, holding headcount flat may be justified. Unmeasured, it is a cost decision framed as a productivity one, and the regret and rehiring data show how often that framing fails to hold.
Recommended
Leads commercial and go-to-market. Founder of Emberwood, an AI-driven cold-email agency for B2B, with a background in outbound, lead generation, and sales.
