Is AI Replacing Engineers? What the Headcount Data Shows

#Don’t learn to code?

In February 2024, Nvidia CEO Jensen Huang took the stage in Dubai and told the audience to stop telling kids to learn to code. For years, tech leaders called computer science a vital skill. Huang now believed AI would handle the coding, so young people should study biology or farming instead.

Enrollment in computer and information science programs at four-year universities fell 8.1% in 2025-2026, the steepest one-year decline of any major since at least 2020 (National Student Clearinghouse). The computer science major specifically fell 11.2%.

That’s a large number of eighteen year olds changing course. Contrary to the statements being made publicly: Engineering headcount grew.

The question underneath the question

The obvious thing to do is check the headcount. If software engineers were being replaced, it should show up in the number of people holding the job, and it should show up in the roles next to them, the designers and product managers who build software alongside engineers.

A useful finding came from trying to line those headcount numbers up against a measure of AI exposure.

The enrollment collapse is a response to statements from executives, not labor statistics. If you look into it, you’ll find that the underlying taxonomy splits software engineering into two occupations, rates them four points apart, and offers no principled way to choose between them.

Defining AI exposure

One of the two main data sources here is the Yale Budget Lab, which blends six distinct exposure metrics into a single, PCA-weighted composite score.

AI exposure metrics measure what share of a job’s tasks could be performed by AI or assisted by AI. High exposure does not mean the job disappears. It can mean AI replaces some of the tasks a worker performs, or augments them.

The Yale analysis found the six metrics mostly agree on who is not exposed. Plumbers, for example. The agreement breaks down elsewhere, and it breaks down hardest in the coding field itself.

Two codes, one job

The government SOC (Standard Occupational Classification) taxonomy splits software work into two occupations:

On a scale where plumbers sit near 1 and the most exposed office roles push past 7, Software Developers score 1.01. Computer Programmers, however, score 5.21.

While these two classifications essentially describe the same day-to-day work, they are separated by nearly four points on a scale that only spans seven.

This bureaucratic split by the Bureau of Labor Statistics (BLS) creates a misleading picture of AI’s impact: it suggests a massive risk to the profession by isolating “Computer Programmers” as highly exposed, even though the modern industry has almost entirely consolidated under the “Software Developer” label, which is rated as barely exposed at all.

Measuring headcount growth with LinkedIn

To layer real world activity over these exposure scores, I pulled LinkedIn headcount data. This measures the one year net change in headcount based on members listing a specific title. It reads whether a role grew or shrank. It’s not a count of hires.

LinkedIn is useful here because it lets me look at specific roles within specific companies, which is the granularity I needed to compare against role-by-role exposure scores. The trade-off is that it reflects self-reported profiles from a handful of companies over one year, making this directional rather than definitive.

Yale’s occupations use SOC classifications, so the names don’t match LinkedIn’s job titles. I built a crosswalk to address that.

Note: LinkedIn data is North America only.

amazon = pd.read_excel('amazon.xlsx', sheet_name='Titles')
amazon = amazon.rename(columns={'1y growth': 'growth'})
# growth = net one-year headcount change per title (not gross hires)

crosswalk = {
    'Software Engineer': 'Software Developers',
    'Product Designer': 'Web and Digital Interface Designers',
    'Senior Product Manager': 'Computer Occupations, All Other',
    'Data Engineer' : 'Database Architects',
    # ...35 title-to-occupation mappings total
}

Aligning software roles

Look at the first entry in that dictionary.

'Software Engineer': 'Software Developers'

I mapped LinkedIn’s “Software Engineer” to “Software Developers” (exposure 1.01), rather than “Computer Programmers” (exposure 5.21). This alignment better reflects the SOC definitions, O*NET’s task descriptions, and the actual professional identity of those holding the title.

Under that mapping, engineering headcount grew and was lower on the AI exposure scale. Growth at low exposure is what you would expect.

Note on Product Management: the same problem appears in a different shape. There isn’t a dedicated SOC classification for a tech product manager, so PM titles fall into a generic “Computer Occupations, All Other” bucket. The exposure score attached to PMs isn’t really measuring product management work. I treat that category with caution throughout.

Another entry worth looking at.

'Data Engineer' : 'Database Architects'

The SOC taxonomy has no occupation called data engineer. O*NET files the title under Database Architects (15-1243) for lack of anywhere better, but these are two different jobs. Database Architects develop and design database systems, while Data Engineers build the pipelines that move and transform data so organizations can use it.

Every observation in the Database Architects row comes from Data Engineer and Senior Data Engineer titles, so the growth is for data engineering, sitting under a label named for something else. It also means the highest exposure score in this set, 7.06, was calculated for an occupation the industry has largely stopped hiring for.

One more, and this one is problematic.

'Machine Learning Engineer': ???

There is no SOC occupation for machine learning. Not a mismatched label this time, just an absence. The taxonomy was last revised in 2018, and the job that most directly does AI work doesn’t have a code. With nothing to map to, ML titles get no exposure score, which means they drop out of any comparison built on that scale.

They are the fastest growing roles in the pull. Machine Learning Engineer and Senior Machine Learning Engineer grew 27.4% across the three companies, 16.6% when weighted by headcount, over roughly 500 people. That is more than six times the growth rate of software engineering. The occupation most defined by AI is the one the exposure data can’t measure.

The three company snapshot

The initial pull covered Amazon, Google, Shopify, Klarna, and Meta. Cleaning left three. Meta was a duplicate and was dropped. Klarna has only a few hundred employees, mostly in sales and operations, with almost no design or product management headcount, which made a three-role comparison impossible.

That left Amazon, Google, and Shopify. A small sample, and an illustrative case study rather than an industry census. It does offer a useful contrast: two incumbents against Shopify, a company that has publicly committed to AI.

In April 2025, CEO of Shopify, Tobi Lütke made AI use a baseline expectation, asking teams to prove a job could not be done by AI before requesting new headcount. If the displacement story held anywhere, it should be visible there.

While Shopify’s smaller headcount means that even a modest number of hires can inflate percentage growth compared to giants like Google or Amazon, the direction of the data is unmistakable. Despite a strict policy requiring AI-justification for every new role, Shopify’s engineering headcount outpaced its peers. A trend that directly contradicts the displacement that such an AI-first mandate would be expected to produce.

What the headcount shows

Every occupation I could match, sorted by exposure. Amazon, Google, and Shopify, one year net change in North America.

Rows are title-by-company observations. Weighted figures are headcount-weighted. Buckets built from few rows and small teams move substantially between the two columns.

# join LinkedIn titles to Yale exposure scores via the crosswalk
jobs = crosswalk_df.merge(combined_companies, on='title') \
                   .merge(tech_yale, left_on='yale_occupation', right_on='occupation')

# average exposure and 1-year headcount growth per occupation
chart_data = jobs.groupby('yale_occupation').agg(
    growth=('growth', 'mean'),
    exposure=('ai_score', 'mean'),
).reset_index()

Three things stand out.

Engineering grew. Software Developers were added at all three companies, including Shopify, which made AI justification a precondition for new headcount. The data shows a +4.3% average growth across the three companies when unweighted. However, when weighted to account for the significantly larger engineering headcounts at Amazon and Google, the figure lands closer to +2%. This weighted result more accurately reflects the impact of company scale on the overall growth trend.

Headcount declines were concentrated in design and web-related roles. Web Developers dropping 11%, Designers falling 3%, and product management decreasing 2%. This was not a uniform trend across all highly exposed occupations. For instance, Database Architects, which registered the highest AI exposure score in our set, defied this pattern with 10% growth.

Highly exposed roles still grew. Exposure scores don’t sort the top of the range. Data engineering roles, which map to Database Architects and carry the highest exposure score in the set, grew across all three companies. Design roles at effectively the same score contracted at all three. Security roles grew from the middle of the range, though the size of that gain depends heavily on one 24-person team at Shopify: weighted by headcount, the increase is 7.0% rather than 23.1%, with Amazon and Google both close to flat. What the data shows is that a high exposure score does not predict the direction of headcount change.

Caveats

• Observation, not causation. This analysis relies on a single-year snapshot of self-reported LinkedIn data from just three companies in North America. While there appears to be a contrast between growing engineering roles and declining design and product management roles, the data cannot explain why these trends diverged.
• No seniority, no hiring volume. Net headcount cannot separate slowed hiring from layoffs, and there isn’t a junior-versus-senior breakdown.
• Taxonomy limits, twice over. The engineering mapping is a subjective analytical choice made for this study. Additionally, the Product Management data is less reliable because it relies on a broad, generic job code that does not accurately reflect the role’s specific responsibilities.
• Not representative of the broader tech industry. This analysis is limited to three companies and features small headcounts in design and product management, meaning these findings may not reflect trends across the wider technology sector, or globally.
• LinkedIn data covers each company’s 100 largest titles. The employee threshold for inclusion differs by company (roughly 330 at Amazon, 136 at Google, 18 at Shopify), so smaller roles appear for Shopify that are cut off for the larger firms. Absence from a bucket means the title fell below that company’s threshold, not that the role does not exist there.
• Machine learning roles excluded. Machine Learning Engineer titles grew 27.4% across the three companies (16.6% weighted by headcount across Google and Shopify), the fastest growth in the pull. They are absent from the comparison because the SOC taxonomy has no machine learning occupation, so there is no exposure score to plot them against. The role most directly tied to AI is the one this scale can’t see.

What I think could be going on

A hypothesis, based on how tech functions operate and what I am seeing here, but not a finding.

The Engineering Paradox: Automation Breeds Work

Open engineering roles are up. Lenny’s Newsletter, drawing on TrueUp data and referenced by Business Insider, reported roughly 67,000 open engineering positions globally as of March 2026, about 78% above the early-2023 low, with the increase accelerating into 2026.

The likely explanation is that while AI automates specific coding tasks, it generates volumes of adjacent work like setting up models, debugging AI-generated output, and maintaining the complex systems built around these new tools. This creates a net increase in the demand for engineers to manage the lifecycle of AI-driven development. This is consistent with both the observed headcount growth and the rising number of open roles.

The Uncertainty of Design and Product Management

The trajectory for design and product management is far less certain. While the decline in design roles is evident in the data, the underlying cause remains ambiguous. It could be a result of direct AI substitution, where tools replace human designers for certain tasks. Alternatively, these reductions might be budget-driven, with design being classified as a “nice-to-have” during periods of fiscal tightening. A third possibility is role absorption, where design responsibilities are increasingly folded into the workflows of product managers or engineers. Product management itself remains the hardest to read, as the data continues to map to a generic job code that does not accurately describe the specific responsibilities of the role.

Bottom line

Exposure scores get the broad direction roughly right at the extremes. Software Developers, at the bottom of the range, grew 4.3%. Web and Digital Interface Designers, second most exposed, declined 3%. Both are what the scores predict.

Where it stalls is at the top. where near-identical scores produce opposite outcomes. Data engineering roles at 7.06 grew while design roles at 7.00 shrank, and the two occupations sit six hundredths of a point apart on a seven-point scale.

Data engineering roles grew modestly at all three firms. Whether that reflects AI systems creating demand for data infrastructure, or the same cloud and analytics trends that have driven data engineering growth since well before the current AI cycle, the data can’t say. The two appear identical here.

Measuring the shift is held back by an obsolete taxonomy. The government splits software work into two codes rated four points apart and offers no rule for choosing between them, and the metrics behind the lower code don’t agree with each other. The problem is that the government’s labels no longer describe what these jobs are. Without a category that matches the work, there’s nothing to measure the claims against.

CS enrollment fell 11.2%, and there’s still no statistic showing that the work itself is going away.

High exposure may be signaling that work is changing rather than ending. That’s a hypothesis, not a finding, and this data can’t distinguish between the two. Worth watching and checking again in the future.

How this was built

Yale’s exposure scores are matched to LinkedIn headcount data through a hand-built crosswalk of 35 job titles. Mapping to machine learning role didn’t work because there’s no SOC code (though that role showed up in the headcount data). Growth is reported both unweighted and weighted by headcount, since small teams move the unweighted average a lot. The sample size is small.

References

Business Insider. (2026, April 3). AI isn’t killing software coding jobs — they’re booming, according to TrueUp data. https://www.businessinsider.com/ai-isnt-killing-software-coding-jobs-booming-trueup-2026-4

Gimbel, M., Kendall, J., & Kulsakdinun, R. (2026, February 19). Labor market AI exposure: What do we know? The Budget Lab at Yale. https://budgetlab.yale.edu/research/labor-market-ai-exposure-what-do-we-know

O*NET OnLine. (n.d.). Summary report for: 15-1255.00 — Web and digital interface designers. U.S. Department of Labor, Employment and Training Administration. Retrieved June 2026, from https://www.onetonline.org/link/summary/15-1255.00

Rumage, J. (2026, March 12). Computer science degrees are losing popularity in the AI era. Built In. https://builtin.com/articles/computer-science-degree-decline-ai

Rachitsky, L. (2026, March 24). State of the product job market in early 2026. Lenny’s Newsletter. https://www.lennysnewsletter.com/p/state-of-the-product-job-market-in-ee9