There may be nice nervousness and uncertainty about AI changing jobs. How can we transfer previous imprecise warnings and bombastic predictions and produce information to bear on this query? One great way is to take a look at the career the place AI capabilities are furthest alongside and adoption has been exceptionally fast: software program engineering.
On this essay, we argue that there’s sufficient proof to reject the narrative that when AI capabilities attain a sure threshold, it is going to trigger mass layoffs. Provided that that is true even in a sector with only a few regulatory boundaries, most different professions are prone to be much more cushioned.
We even have a great understanding of why that is the case. We are able to consider many varieties of information work, together with software program improvement, as a “decide-execute-deliver sandwich”. AI compresses the “execute” layer — the center of the sandwich — however the different two layers resist automation in a means that won’t be overcome by functionality enhancements alone.
We conclude on a word of cautious optimism concerning the future trajectory of demand for software program engineering. This essay is the primary in a collection, and the following one will take a look at the reason why particular person software program engineers’ careers is likely to be rocky even when total demand is wholesome. The collection relies on the revealed literature in economics and software program engineering, our personal evaluations and observations of AI brokers, and lots of software program engineers’ reflection on the current and way forward for AI impacts on their career, gleaned each from revealed writings and our interactions with the group.
Contemplate three tales that made the headlines and the way they contrasted with actuality:
In February, fintech firm Block (maker of Money App, Sq., Afterpay, and different such apps) introduced layoffs of 4,000 workers as a result of, in keeping with founder Jack Dorsey, AI is “enabling a brand new means of working” with “smaller and flatter groups”, particularly citing late-2025 enhancements in mannequin capabilities.
However subsequent reporting revealed a radically totally different image. After rising headcount greater than threefold through the pandemic, the corporate was beneath large monetary strain. An information scientist on the Money App crew, Naoko Takeda posted that Block “shoved AI down everybody’s throats” but she noticed “very restricted features in productiveness.” She refused a 75% retention increase and stop. Different workers interviewed had a sharply totally different understanding of what AI was able to at Block and whether or not Dorsey had a reliable understanding of the problems.
As Aaron Levie has identified, CEOs are uniquely susceptible to delusions about AI’s usefulness as a result of they’ll construct fast prototypes however can’t see the 90% of labor it takes to show it right into a completed product. Dorsey’s public statements about AI appear to suit precisely this sample.
In April, Snap laid off about 1,000 folks, with CEO Evan Spiegel primarily citing AI as the explanation in his layoff memo. He additionally stated that AI generated 65% of latest code. In actuality, the layoffs adopted a marketing campaign by an activist investor demanding price cuts. (Snap has posted a internet loss each full 12 months since its 2017 IPO and shares have been down over 30% in 2026). Tellingly, the character of the cuts, comparable to 150 jobs spanning varied roles within the augmented actuality division, don’t correlate with the cuts we might count on to see in the event that they have been pushed by AI (i.e. programming and different “AI-exposed” jobs throughout the board, not concentrated in any unit).
In Might, Intuit introduced 3,000 cuts, alongside offers with Anthropic and OpenAI. The press linked the 2, framing the layoffs as AI-driven restructuring. For as soon as, the CEO really pushed again on this straightforward narrative, saying that “none of it needed to do with AI” and that the cuts focused “coordination-heavy roles” and too many administration layers.
We didn’t cherry-pick these examples. In each story about AI-driven software program engineering layoffs that we examined, the identical narrative violation emerged. It seems that “AI washing” of job cuts is an economy-wide phenomenon, evidenced by many surveys:
59% of U.S. hiring managers admitted they emphasize AI when explaining hiring freezes or layoffs as a result of it performs higher with stakeholders than citing monetary constraints.
Forrester principal analyst J. P. Gownder says of corporations getting ready supposedly AI-driven layoffs: “After we ask if they’ve a mature, vetted AI app able to fill in these jobs, 9 out of 10 instances, the reply isn’t any—and so they haven’t even began.”
In a HBR survey of over 1,000 world executives, 21% had made giant headcount reductions “in anticipation of” AI, with one other 39% having made low or average anticipatory headcount reductions. In distinction, solely 2% had already made giant reductions in headcount associated to precise AI implementation. The 10x hole means that executives, like everybody else, are extremely susceptible to succumbing to the deceptive narratives about AI changing jobs.
One other fascinating information level comes from the WARN Act, which requires sure disclosures of plant closings and mass layoffs affecting over 100 staff. In March 2025, New York grew to become the primary U.S. state so as to add an AI disclosure checkbox to WARN Act filings. Within the full first 12 months, greater than 160 corporations filed WARN notices. Not a single one checked the AI field. We reached out to the NY Division of Labor who confirmed that as of late Might, just one firm, Nespresso, checked the field. If these filings are correct, solely 46 out of about 25,000 laid off staff in New York State within the related interval, or about two-tenths of a p.c, have been affected by AI.
Much more damning for the AI-driven-mass-layoffs narrative: layoffs are the unsuitable sign of AI’s potential productiveness advantages within the first place! The analysis is obvious that the impact operates via “slower hiring somewhat than elevated separations”. Firing current staff ends in the lack of exactly the tacit information and organizational capital that permits staff to function AI successfully. In addition to, it’s costly when it comes to severance, injury to morale, and rehiring threat. Given these prices, it’s largely pointless on condition that pure turnover achieves the identical end in just a few years.
So what does the information inform us once we look past layoffs to total employment tendencies? An essential paper from Federal Reserve economists compiles the proof within the U.S. context. Employment remains to be rising, however they discover that it’s rising slower post-ChatGPT in comparison with a no-AI counterfactual, by about 3 share factors per 12 months. One essential limitation of this research is that the methodology can’t seize self-employment, so it’s attainable that a number of the slowdown in progress is being absorbed by entrepreneurship as a substitute. We do have proof from different research that AI makes entrepreneurship simpler. So the true image might be even more healthy than the Federal Reserve research suggests.
Lastly, it’s value acknowledging two sorts of indirectly-AI-driven job losses in software program engineering which might be actual, however totally different from AI changing software program engineers. First, AI generally decimates demand for the product, in circumstances like Chegg (homework assist) or Stack Overflow (technical assist), each of which have laid off staff. AI doesn’t immediately do the job that these staff did, however somewhat obviates the necessity for it. The historic parallel is powerful: Among the many 270 jobs within the 1950 U.S. census, just one job was automated away — elevator operator. However many others have been rendered out of date by new know-how, just like the job of telegraph operator.
One other credible AI-driven layoffs story is amongst corporations that promote AI, somewhat than purchase it. So when corporations like IBM or SAP announce layoffs due to AI, a extra correct framing is “we reallocated headcount from legacy features to our fastest-growing product line.” That’s strange company restructuring round a income alternative, not know-how displacing staff.
Many tech leaders, just like the Snap CEO above, report the share of code written by AI alongside stories of layoffs or predictions of future job losses. This feeds into the simplistic psychological mannequin that when AI writes all of the code, there isn’t a want for coders. Happily, this psychological mannequin is unsuitable. This AI-written-code metric is nearly fully disconnected from what issues for labor displacement. Right here’s why.
First, writing code isn’t, and by no means was, the bottleneck. For instance, a 2019 paper summarized current research with the conclusion that “builders spend surprisingly little time with coding, 9% to 61% relying on the research”. This discovering was in keeping with the paper’s personal information from 6,000 builders at Microsoft. As coding brokers started to be taken up, there was an explosion of weblog posts in late 2025 stating that writing code isn’t the bottleneck, as builders realized that utilizing brokers to put in writing many of the code led to little affect on total productiveness [1, 2, 3, 4, 5, 6, 7, 8].
If writing code isn’t the bottleneck, what’s? The duty-breakdown surveys level at issues like conferences or debugging. This simply results in extra questions: what are builders doing in these conferences and why can’t it’s performed by AI? Received’t debugging get automated as capabilities enhance? To know the true bottlenecks, we’ve got to get qualitative, and dig into software program engineers’ personal understanding of what it’s they do this resists automation.
After we did this evaluation, it revealed three issues as the true bottlenecks (1) deciding and specifying what to construct, (2) verifying and being accountable for what’s delivered, and (3) the deep human understanding — of the codebase, the enterprise, and the setting — required to hold out each of those.
In different phrases, software program engineers’ work consists of a “decide-execute-deliver” sandwich (with understanding being a prerequisite for all three). AI has compressed the center of the sandwich, however has left the 2 ends largely unchanged. So long as software program improvement groups are in command of choice making and accountable for what they ship, engineers nonetheless must spend time build up a deep understanding of the system. These are the three bottlenecks.
Determine: Software program improvement consists of three layers: (1) Choice making — drawback framing, specification, planning (2) execution — design and implementation (3) supply — testing, verification, integration, upkeep, and so on. Be aware that these are conceptual layers, not temporal phases. It is not uncommon to change backwards and forwards in the midst of a undertaking.
Proof for the sandwich mannequin of AI’s productiveness results comes from a current paper on “Writing Code vs. Transport Code”. Throughout 100,000 builders on GitHub, the researchers discovered that AI brokers led to an eight-fold improve within the variety of strains of code written, in keeping with the concept that AI nearly fully compresses the Execute layer of the sandwich. However this led to solely 30% extra releases, strongly suggesting that human bottlenecks (the Determine and Ship layers) stay in place.
Can the sandwich be additional compressed? We don’t suppose so. At one finish of the pipeline, improvement groups must resolve what to construct. Some of the essential classes junior software program engineers study is that necessities specification (the career’s lingo for this layer) takes surprisingly lengthy, and whether it is compressed, it results in rather more ache down the road. This layer is tough to automate as a result of it requires fascinated about consumer wants, market alerts, organizational priorities, and in some circumstances regulatory constraints.
As AI capabilities enhance, the varieties of selections that may be delegated to AI improve over time. However this doesn’t make the “resolve” layer thinner — as soon as a choice will be delegated to AI, it’s now not a supply of aggressive benefit, and the worth of human decision-making migrates upward. Software program will increase in complexity over time, so there isn’t a ceiling to this course of.
On the different finish of the sandwich, human groups have to be accountable for what they ship. It’s attainable that some day sooner or later groups will ship mission-critical code with out absolutely testing and understanding it, however right now’s AI is so unreliable that such haphazard practices would characterize an existential risk to software program groups and their clients.
Even when the technical boundaries go away sooner or later, we don’t need to cede management to AI. A central perception of AI as Regular Know-how is that we will collectively select to maintain people accountable via shared norms, regulation, and coverage. This can be a rather more resilient strategy to management the velocity of AI impacts and enhance security than making an attempt to sluggish the event of technical capabilities. These velocity boundaries are already largely in place because of legal responsibility legal guidelines and sector-specific regulation, however will be additional strengthened. (For an extended model of this argument, see the authentic essay.)
On this imaginative and prescient, as increasingly more of the execution layer will get delegated to AI, the software program engineer’s function sooner or later turns into analogous to that of a crane operator. AI brokers will do many of the cognitive heavy lifting; supervising the agent and maintaining it in management turns into many of the human’s job.
Some commentators argue {that a} future with people staying in management is unlikely as a result of it’s too expensive to pay folks to take action. There have already been just a few viral tales of poorly-supervised coding brokers deleting manufacturing databases or inflicting different forms of injury. However we view these as “man bites canine” tales somewhat than an rising norm. They go viral exactly as a result of they characterize such irresponsible and weird conduct that they’ve shock worth, and function common reminders and studying moments serving to the group guard itself in opposition to over-reliance on AI. Because the aphorism goes, “if it’s within the information, don’t fear about it”. Nonetheless, with the ability to detect whether or not there may be an uptick in poorly-supervised use of AI for high-stakes duties — throughout the financial system, not simply in software program engineering — stays one of the essential information gaps we’ve got right now.
By the way in which, the sandwich getting squished is a brand new development and it isn’t uniquely because of AI. Over twenty years in the past, the Bureau of Labor Statistics began monitoring programming individually from software program engineering. Roughly talking, programmers are accountable just for execution whereas software program engineers handle an even bigger a part of the sandwich. Not solely has programming been shrinking, it’s also pays a lot much less as a result of it’s seen as grunt work. AI merely accelerates this long-existing development, additional devaluing purely technical expertise.
This sample — the place people stay closely concerned at each ends of the decide-execute-deliver sandwich, at the same time as AI more and more automates the center layer, appears to be broadly relevant to most information work, although it’s farthest alongside in software program. In spite of everything, advanced choice making and accountability are frequent to most fields. An absence of recognition of this phenomenon has led to many overconfident predictions about imminent job losses, comparable to amongst radiologists.
One cause for confusion concerning the extent to which software program engineering is altering is the sloppy use of the time period “vibe coding” to check with a large spectrum of practices, the ends of that are conceptually distinct and extra dissimilar than related.
In true vibe coding the consumer merely tells the agent what to do, doesn’t supervise it when it’s working, doesn’t evaluate the code — won’t even have the talents to take action — and doesn’t consider the output, past maybe noticing when issues are visibly damaged.
That is in distinction to how most software program engineers are literally utilizing brokers — as a device, with the human remaining in management and accountable for the output. Happily, the time period agentic engineering is gaining foreign money as a descriptor of this apply.
As agentic engineering has grow to be the norm, engineers are discovering that supervising coding brokers is surprisingly time consuming. For instance, Simon Willison, a distinguished developer and chronicler of the AI transition, has famous how he’s mentally exhausted by 11am from supervising brokers. That is in keeping with our expertise as properly.
Extra quantitative proof comes from SWE-chat, a dataset of coding agent interactions from open-source builders who opted right into a logging device. The research discovered that solely 44% of agent-produced code survives into consumer commits, that vibe-coded commits introduce vulnerabilities at 9 instances the human-only charge, and that the commonest consumer intent is knowing current code, not producing new code (19% vs 13%). The self-selected nature of the dataset signifies that we will’t draw robust conclusions based mostly on this research alone, nevertheless it does reinforce many different strains of proof that vibe-coding and agentic engineering patterns are fairly totally different.
To re-iterate, these will not be two distinct classes. They’re two ends of a spectrum, and there’s a blurry center. Not each undertaking is both a throwaway or mission-critical. Not each workflow suits exactly within the left column or the correct column of the desk. However the important thing implication for the roles query stays stable — corporations can’t ship manufacturing software program by hiring unqualified vibe coders as a substitute of software program engineers.
AI boosters would possibly declare that mass layoffs are coming; they only haven’t occurred but as a result of human-level software program engineering talents are very current (or haven’t been achieved but). But when the sandwich mannequin is appropriate, these predictions gained’t come true. AI has already largely compressed the center of the sandwich (and the compression really began a long time in the past). So even making the execution layer immediate and excellent will solely be a small change from the established order. The the reason why the opposite two layers have resisted AI is not due to functionality limitations.
The truth is, not solely are software program engineering jobs not going away because of AI, there would possibly even be a rise in demand for software program engineers. When software program (or the rest) will get cheaper to create because of technological productiveness enhancements, folks will purchase much more software program (in econ jargon, software program is very “value elastic”). And as we’ve got argued, AI doesn’t change software program engineers (the “elasticity of substitution” is low), so the demand for extra software program ends in a derived demand for extra software program engineers. A loosely associated however flashier economics time period, “Jevons’ paradox”, is usually thrown round within the AI discourse to explain this idea.
Traditionally, this has been the sample — programmer employment within the U.S. has grown from near-zero round 1950 to tens of millions right now. That is sharply totally different from occupations comparable to agriculture by which labor demand was famously decimated because of mechanization and automation. The distinction is that the quantity of energy folks devour is comparatively mounted — even a 25% improve led to the weight problems epidemic — whereas the quantity of software program produced has grown a millionfold. Trendy vehicles have one thing like a hundred million strains of code working on their varied on-board computer systems.
If there’s a ceiling to the demand for code, we’re nowhere close to it. Just about all cognitive work advantages from software program. As AI makes coding cheaper, individuals are creating all types of one-off utilities — whether or not for work or private use — that it by no means made sense to create till now.
To be clear, whereas we predict there can be much more software program sooner or later, and sure extra software program engineers, this doesn’t imply large tech corporations will get even greater. Nearly all of software program engineers right now already work in-house in non-software corporations, and that share would possibly develop sooner or later. Then there’s the concept of “AI rollups”, which refers to enterprise capital or personal fairness corporations shopping for “Fundamental road” companies — dentistry practices, accounting corporations, and whatnot — and rebuild them from the bottom as much as be “AI-native” by embedding software program engineers or AI engineers into these companies. After all, it’d find yourself being nothing greater than hype. It’s too early to inform.
Some folks predict that demand for software program engineering expertise will fall due to democratization. They acknowledge that there can be extra software program produced than ever earlier than, and in addition that extra human time can be spent producing software program than ever earlier than, however that this work can be performed by people who find themselves not software program engineers. The thought is that AI will democratize software program engineering to the extent that authorized software program, as an example, will be extra simply created by these with coaching in regulation than in software program engineering.
Perhaps. However we’ll guess in opposition to it. In our view, this falls into the identical lure of conflating vibe coding with agentic engineering, and the execution layer with the the entire decide-execute-deliver sandwich. The truth is, once we take a look at the historical past of programming, there have all the time been claims that we’re on the threshold of democratization — previous languages comparable to FORTRAN, COBOL, and SQL have been all accompanied by such distinguished hopes on the time of their introduction. It by no means occurred. The barrier isn’t really studying the syntax. It’s having sufficient expert judgment to make good selections whereas sustaining accountability.
In the end the excellence could also be semantic. It appears clear that the period of time folks spend on getting computer systems to do new issues will improve over time. This would possibly take the type of constructing software program, or managing advanced workflows utilizing brokers, or one thing else. It should require a mixture of software program expertise, AI expertise, and area experience. Whether or not it’s right now’s software program engineers who will greatest adapt to fill these new roles stays to be seen.
That final level concerning the want for adaptation units up the following essay on this collection. The truth that mixture labor demand in software program is prone to stay robust doesn’t imply that almost all particular person staff gained’t be affected. We’ll argue that AI will create large structural shifts in how software program is produced, which may have large impacts on which software program engineers stand to realize or lose — based mostly on the forms of corporations they work in, their geography, their seniority, the tempo at which they’ll adapt.
Additional studying
Deena Mousa factors out the superficiality of broad, economy-wide analyses of AI impacts based mostly on metrics like “AI publicity”, and as a substitute requires “cautious, occupation-specific work”. We hope that this collection of essays will play a job in establishing a nuanced understanding of AI’s transformation of software program engineering. We’ve earlier coauthored, with Justin Curl, a paper analyzing AI in authorized companies that significantly engages with regulatory and different bottlenecks that make that occupation distinctive. We plan to do extra occupation-specific deep dives sooner or later.
In a exceptional essay referred to as No Silver Bullet 40 years in the past, Fred Brooks distinguished between the “important complexity” and “unintentional complexity” of software program. He argued that a number of the complexity of software program is unintentional, arising from limitations of current know-how such because the clunkiness of programming languages, and will be alleviated over time as tooling improves. However a few of it’s important, as a result of specifying the right conduct of software program is itself onerous. He presents a forceful articulation of why the “resolve” layer of the sandwich is thick and resists automation. Curiously, hopes of boosting programmer productiveness via AI have been already distinguished again then! Brooks argues that as a result of AI or another know-how solely reduces unintentional complexity, it gained’t end in an order-of-magnitude productiveness enchancment. (Brooks is the creator of The Legendary Man Month, an essay assortment that’s nearly actually the perfect recognized and most influential writing on software program engineering of all time. No Silver Bullet later grew to become a part of the gathering.)
We’re grateful to Felix Chen for suggestions on a draft.
