Dev.to AI 🤖 Ai 👁 0 📖 11 min read

AI Makes Engineers Faster. Does That Mean Companies Need Fewer?

AI is helping developers complete certain tasks faster. Whether that means more software, fewer jobs or a different kind of engineering workforce depends on decisions made well beyond the coding environment. Imagine lea

AI is helping developers complete certain tasks faster. Whether that means more software, fewer jobs or a different kind of engineering workforce depends on decisions made well beyond the coding environment.

Imagine leading a software engineering team of 40 people.

There is a long product roadmap, customer requests waiting for attention and older systems that need maintenance. The team has more work than it can comfortably handle.

Then AI coding tools begin making a difference. Routine changes take less time. Developers can prepare tests, investigate unfamiliar code and complete some implementation tasks faster.

At first, the improvement looks like an obvious advantage.

Then someone asks a question that changes the conversation.

If 40 engineers can now accomplish more, does the company still need all 40?

It is a reasonable business question. Companies have always invested in technology to improve efficiency, and sometimes that has meant reducing labor costs.

But I would challenge any decision to reduce a team based primarily on a coding productivity figure. Not because engineering jobs are immune to automation, but because that number tells management very little about the work the company still needs to accomplish.

There is a larger business decision hidden behind the productivity measurement.

The productivity numbers need context

In November 2026, research organization METR published an independent assessment of AI-assisted research and development at Anthropic.

The researchers estimated that Anthropic’s rate of improvement on AI capability benchmarks had accelerated by approximately 1.9 times compared with an earlier period. They considered AI a likely major contributor, while acknowledging substantial uncertainty about the causes. Important parts of the supporting evidence remained confidential or redacted.

It was a significant result, although it measured the rate of progress in AI capabilities rather than the productivity of individual engineers.

METR’s November 2026 assessment explains the measurement and its limitations.

Studies of software developers offer a different perspective.

Three randomized field experiments involving 4,867 developers found that AI coding assistance increased completed tasks by approximately 26%. Another randomized study by METR found that experienced open-source developers took 19% longer to complete the particular tasks tested with early-2025 AI tools.

These findings came from different working environments, tools and tasks. They should not be treated as competing estimates of one universal improvement.

Sources: Developer field experiments and METR’s experienced-developer study.

The table shows what each study measured, where, and what it leaves out.

Study What was measured Setting What it does not show
METR assessment of AI-assisted R&D at Anthropic (November 5, 2026) The rate of improvement on AI capability benchmarks, about 1.9 times faster than in an earlier period (95% interval 1.5 to 2.8 times), for well-specified tasks Anthropic’s own AI research and development. METR said Anthropic redacted most of the evidence it found most informative How much faster individual engineers work. METR says individual speedup estimates do not correspond to acceleration rates
Three randomized field experiments (paper dated June 2025) Completed tasks, about 26% more for developers using an AI coding assistant (standard error 10.3%) 4,867 developers at Microsoft, Accenture and a Fortune 100 company, during their ordinary work Delivered product, customer value or headcount. The paper says each experiment is noisy
METR experienced-developer study (July 2025) Time to complete real issues. Tasks took 19% longer when AI tools were allowed 16 experienced open-source developers, 246 issues, early-2025 AI tools Today’s tools. In February 2026 METR called its newer data an unreliable signal of the current effect and said developers are likely more sped up now than in early 2025
NBER Writing Code vs. Shipping Code (September 2026 revision) Autonomous coding agents associated with about 240% more commits, 80% more projects and 30% more releases, in a matched event study More than 500,000 GitHub developers Customer value. Total usage did not rise across the marketplaces examined, and the study does not establish why releases rose less than commits

The studies are listed under Go deeper. METR’s February 2026 update is the source of the note in the third row. The rows measure different things, so the percentages are not comparable.

A further distinction becomes visible when we look beyond individual tasks.

The September 2026 revision of the NBER working paper Writing Code vs. Shipping Code examined activity among more than 500,000 GitHub developers.

Using a matched event study design, the researchers estimated that adoption of autonomous coding agents was associated with approximately 240% more commits and 30% more software releases.

These are different measures of software production, not interchangeable percentages of productivity. A commit records a code change, while a release reflects a later stage of delivering software. Neither directly establishes customer value.

The difference in estimated gains is nevertheless informative. It shows that greater coding activity was accompanied by a much smaller increase in completed releases.

The study does not establish precisely why those increases differed. Review, testing, integration and release coordination are possible constraints that organizations would need to examine in their own environments.

Read Writing Code vs. Shipping Code.

For an engineering leader, this changes what is worth measuring. A team may be producing more code without completing proportionally more work that customers can actually use.

That distinction becomes important when productivity starts influencing budgets and staffing plans.

A smaller team can sometimes be the right business decision

Consider an organization maintaining a mature internal application.

Its requirements change slowly. Most of the engineering effort involves maintenance, routine enhancements, compliance updates and occasional production issues. The business has no substantial plans to expand the application.

Suppose AI reliably reduces the time required for some of that work.

Over time, the organization may be able to meet the same obligations with fewer engineering hours. Management might reduce contractor spending, leave positions unfilled when employees depart or consolidate responsibilities.

The team could become smaller without a major restructuring announcement.

That possibility should not be dismissed simply because architecture, security and operational judgment still require human attention. Those responsibilities affect how far staffing can be reduced safely, but they do not guarantee that the original team size will remain necessary.

There is also a practical limit to cost reduction. A smaller team must still handle unexpected failures, complex changes, production accountability and continuity when experienced employees leave.

Management has to determine whether the efficiency gain is dependable enough to justify accepting those risks.

For organizations with stable requirements and limited demand for additional development, the case for reducing labor requirements may be legitimate.

Lower development costs can also change what is worth building

A company with a growing product roadmap faces a different decision.

Perhaps several promising projects have remained on hold because development was too expensive. Customers are requesting capabilities the team cannot deliver quickly enough, and hiring another group of engineers would make the economics unattractive.

AI may change those calculations.

A feature that previously required months of implementation might become affordable. Smaller product ideas can be tested without committing as much engineering capacity. Improvements that were repeatedly postponed may become realistic.

In that situation, retaining the engineering team could create more value than reducing its cost.

There is a condition, though, and it has little to do with how capable the AI tool is.

The additional software must solve problems that customers or the business genuinely value.

The NBER research found that greater software-production activity did not produce a corresponding increase in total usage across the application marketplaces it examined. This does not establish that new applications were worthless, but it cautions against assuming that greater supply will create equivalent demand.

A company with valuable unmet customer needs may benefit from expanding development. Another company may have little commercial reason to build more, even when implementation becomes cheaper.

The difference lies in the opportunity available to the business.

The first employment effect may be fewer openings

Workforce changes are often judged through layoff announcements. Those numbers matter, but they do not capture everything happening inside engineering organizations.

Imagine a team that normally recruits several graduates each year. Its existing engineers begin completing more work with AI assistance, and management decides to postpone the next round of junior hiring.

Nobody has lost an existing job. The team’s headcount may remain stable.

For someone trying to enter the profession, however, the opportunity has disappeared.

A July 2026 analysis by Indeed Hiring Lab offers relevant context. U.S. software-development job postings had increased by almost 15% from late February 2025, but 71% of the increase between May 2025 and May 2026 came from senior positions.

This was a change in advertised opportunities, not proof that AI caused employers to favor experienced engineers.

Indeed Hiring Lab’s analysis provides the underlying figures and comparison periods.

Federal Reserve researchers have also documented slower employment growth in coding-intensive occupations since ChatGPT’s introduction. Other evidence is more reassuring about aggregate employment. A Danish administrative-data study found no detectable effects exceeding 2% on earnings or recorded hours two years after ChatGPT’s launch, while the U.S. Bureau of Labor Statistics projects continued growth in software-development-related employment through 2035.

These sources examine different populations, time periods and outcomes. None establishes a universal causal estimate of AI-driven engineering job losses.

Sources: Federal Reserve employment research, Danish labor-market study and BLS occupational outlook.

What deserves attention is the possibility of a gradual change in the workforce’s composition.

A company might continue employing experienced engineers while recruiting fewer people at the beginning of their careers. Its reported headcount could remain relatively unchanged even as opportunities to enter the profession narrow.

The available evidence does not establish how widespread that outcome will become. It is still an important possibility for leaders planning their future engineering capacity.

The decision I would challenge in a leadership meeting

Return to the original 40-person team.

After several months of AI adoption, its engineering manager reports that selected implementation tasks now take 30% less time.

A proposal follows: reduce the engineering budget by a similar percentage.

I would question that calculation immediately.

Suppose implementation previously accounted for half the effort involved in delivering a feature. If the same implementation work now takes roughly 30% less time, the direct saving would be approximately 15% of the original total effort, assuming the other activities remain unchanged.

Interactive figure: view it in the original article

Accessible text alternative for this figure

Before AI, the effort to deliver one feature is shown as one bar of 100 percent: implementation is half of it and everything else is the other half.

After AI, implementation takes 30 percent less time, so it shrinks from 50 percent to about 35 percent of the original total. Everything else stays at 50 percent, so the bar reaches about 85 percent and the dashed block marks about 15 percent saved.

Everything else includes testing, integration, architecture decisions and operational review. They stay the same in this example and may limit how much faster releases get. This is an illustration of the example in the text, not measured data.

The effort to deliver one feature before and after AI in the 40-engineer example. If implementation is half the effort and now takes 30% less time, the saving is about 15% of the original total, assuming the other activities remain unchanged. These are illustrative figures from the text, not measured data.

That is already very different from a 30% reduction in the entire team’s workload.

And the calculation still does not tell us whether the organization can deliver releases faster. Testing, integration, architecture decisions and operational reviews may limit the improvement.

The next conversation should concern the product roadmap.

Are there customer problems the team can now solve economically? Could saved capacity reduce maintenance backlogs or improve reliability? Would the company gain more by releasing important features earlier?

If the answer is yes, reducing the team might sacrifice an opportunity that is worth more than the immediate payroll saving.

If the roadmap contains little valuable additional work, the calculation changes. Management may have a legitimate reason to reduce costs, provided the remaining team can safely support the systems it owns.

This is where leadership judgment matters.

A workforce decision should follow an assessment of delivered output, available demand, operating risk and long-term capability. A coding tool’s productivity dashboard cannot supply those answers.

The consequence for engineers

Individual engineers do not control whether their employers use additional capacity for growth or cost reduction.

They do influence how effectively they contribute to the work that remains valuable.

Learning to use AI tools is part of that responsibility. So is developing the judgment required to understand requirements, evaluate generated code, make architecture decisions and deliver reliable systems.

Those responsibilities are not permanently protected from automation. AI capabilities will continue to change how much human effort they require.

The useful career response is therefore not to compete solely on how quickly code can be produced. It is to become effective across the engineering decisions that determine whether the resulting software works and creates value.

That is a direction for professional development, not a promise of employment security.

A productivity gain is not yet a workforce strategy

I do not think companies should avoid reducing engineering costs when technology genuinely allows them to operate more efficiently.

But I also do not think an impressive AI productivity result is enough to justify a staffing decision.

The choice has consequences beyond the next budget cycle. Reducing a team may save money while also losing knowledge, weakening future capability or delaying products that have become affordable to build. Keeping the same team without worthwhile work is not automatically better.

Leaders need to make that trade-off explicitly.

The responsibility is not to preserve every existing engineering position or eliminate every hour that AI can save. It is to decide which engineering capability the business needs, what that capability can now accomplish and whether the value justifies the investment.

That is the decision a productivity benchmark cannot make.

Go deeper

METR: Anthropic AI R&D Acceleration Assessment: Independent assessment of AI capability progress and uncertainty about its causes.

The Effects of Generative AI on High-Skilled Work: Randomized field experiments evaluating AI assistance and software-developer task completion.

METR: Experienced Developer Productivity Study: Controlled evidence showing why productivity depends on task and environment.

NBER: Writing Code vs. Shipping Code: Research comparing coding activity with later stages of software production.

Indeed Hiring Lab: AI and Job Postings: Analysis of software-development job postings, including changes by seniority.

Federal Reserve: AI and Coder Employment: Research into employment trends and causal uncertainty across coding-intensive occupations.

NBER: Still Waters, Rapid Currents: Administrative labor-market evidence on AI adoption, earnings and working hours.

U.S. Bureau of Labor Statistics: Software Developers: Official employment projections for software developers and related occupations.

Related TechiesJournal reading

The Cost of AI: How Technology Companies Are Rethinking People and Computing Power

When AI Does the Junior Work, Who Develops the Next Generation of Professionals?

AI Writes the Code. Who Understands What Happens in Production?

Research cutoff: November 16, 2026. Research measures, employment observations and author interpretations are distinguished throughout the article.

Originally published on TechiesJournal.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.