The New Software Economics: Why Cheaper AI-Assisted Development May Increase Demand for Developers, Not Shrink It
For three years, the dominant question about AI in software has been how many jobs it would eliminate. A different hypothesis is now gaining ground, and the U.S. labor data leans toward it: when the cost of building software falls, the volume of software worth building rises.
The U.S. Bureau of Labor Statistics puts numbers on that idea. Overall employment of software developers, quality assurance analysts, and testers is projected to grow 15 percent from 2024 to 2034 — much faster than the average for all occupations — with about 129,200 openings projected each year over the decade. BLS attributes that demand specifically to the continued expansion of software development for artificial intelligence, IoT, robotics, and other automation applications.
At the same time, the same agency projects a decline for a related occupation: employment of computer programmers is expected to fall 6 percent from 2024 to 2034, with some higher-skilled programming tasks shifting to software developers.
That split is the whole story in two data points. Task-level automation is real. Occupational collapse is not.
Adoption is near-universal; trust is not
| Metric (Stack Overflow Developer Survey 2025) | Result | Prior year |
|---|---|---|
| Use or plan to use AI tools | 84% | 76% |
| Professional developers using AI daily | 51% | — |
| Actively distrust AI output accuracy | 46% | 31% |
| Trust AI output accuracy | 33% | — |
| Report "high trust" | 3% | — |
| Not using AI agents, or using simpler tools only | 52% | — |
Source: Stack Overflow Developer Survey 2025 (49,000+ respondents, 177 countries).
The survey found that 84% of developers use or plan to use AI tools, up from 76% a year earlier, while more developers actively distrust the accuracy of those tools (46%) than trust it (33%) — only 3% report high trust in the output. The top frustration, cited by 45% of respondents, is AI output that is "almost right, but not quite," and 66% say they spend more time fixing almost-correct generated code.
For an operations or engineering leader, that gap is not a reason to wait. It is a budgeting instruction: verification hours are part of the project cost, not an overhead line you can trim.
The economic argument, stated plainly
This is a Jevons-paradox scenario applied to code. When unit cost falls, total consumption tends to rise.
Consider a mid-sized U.S. manufacturer that needs to digitize inventory reconciliation, build a custom app for field service technicians, or replace a process still running on spreadsheets and email threads. Three years ago, the cost of specifying, coding, testing, and maintaining that tool put it below the investment threshold. It got shelved.
With AI accelerating code generation, documentation, and test creation, that same request re-enters the budget conversation. Multiply it across thousands of shelved requests, and the addressable universe of software expands substantially.
"Productivity gains do not necessarily mean reduced demand for talent. If we can develop software in less time with a better cost-to-value ratio, we also expand the number of problems that are viable to solve with technology. AI may let us do more projects, not simply do the same projects with fewer people," says Fabio Caversan, Global Chief Technology Officer at Stefanini Group, a technology consultancy operating in 46 countries.
The same logic explains why higher individual productivity doesn't automatically shrink teams. An organization producing more with the same headcount can redirect that capacity toward modernizing legacy systems, automating processes, or launching services that were previously deprioritized.
The strategic question shifts from "how much can we save with AI?" to "what can we build now that wasn't feasible before?"
Three stages, three levels of maturity
| Stage | What it does | Industrial example | Maturity |
|---|---|---|---|
| Generative AI | Responds to a prompt | Generates code, documentation, a recommendation | Mainstream (84% adoption) |
| Agentic AI | Receives a goal, reads context, executes a task sequence | Analyzes a ticket, generates code, runs tests, flags errors, coordinates with other tools | Early (52% not using agents) |
| AI + physical capability | Converts a decision into an action | Agents detect an operational anomaly and dispatch a robot for inspection | Emerging, vertical by vertical |
Sources: Stack Overflow Developer Survey 2025; Stefanini Group (August 2026).
A majority of developers (52%) either don't use agents or stick to simpler AI tools, and 38% have no plans to adopt them. Among those already running agents, roughly seven in ten report time savings on specific tasks, according to figures cited by Stefanini.
Skepticism follows adoption here too. Stack Overflow reported that 87% of respondents agreed they are concerned about the accuracy of information provided by AI agents, and 81% cited security concerns.
"Artificial intelligence is also software. To generate real value, it needs to be integrated into processes, rules, and architectures capable of guaranteeing that the technology works safely and consistently," Caversan notes.
That constraint is familiar to anyone who has deployed an MES or SCADA integration. Agents need system access, clean data, business rules, security controls, and defined workflows before they do anything useful in production. It is the same maturity curve documented in our coverage of how industrial AI moved from experimental "toy" to operational tool.
When software starts acting on the plant floor
The next frontier connects that intelligence to machines. An agent can analyze data and reach a conclusion; connected to a physical system, that conclusion becomes an action.
On a production line, distributed agents could monitor operational data, identify an anomaly, and coordinate a response while a specialized robot performs the inspection. Stefanini's view is that this evolution will advance first through robots and specialized systems built for concrete tasks — manufacturing, logistics, agriculture, infrastructure — rather than through general-purpose humanoids. Specialization lets you design the system around the specific environment it operates in.
For U.S. manufacturers, the financial case for that layer is already being modeled rather than theorized; our robotics ROI framework for manufacturers walks through the payback math, and the digital twin deployments now running at production scale show what happens when simulation and physical operations are wired together.
The emerging competency is orchestration: coordinating people, virtual agents, and machines simultaneously. That is a management skill, not a coding skill.
The developer's role changes; the developer doesn't disappear
As AI absorbs repetitive work, human value concentrates in architecture, integration, validation, domain knowledge, security, and supervision of increasingly autonomous systems.
Developers themselves are drawing that line clearly. Per the same survey, 76% do not plan to hand off deployment and monitoring to AI, and 69% do not plan to hand off project planning.
That caution aligns with what the BLS projection actually says: the occupation growing 15% is software developer — the role that owns design, integration, and accountability — while the narrower coding-execution role is the one projected to contract.
What this means for U.S. industrial operations
- Reopen the shelved backlog. Pull the internal-tooling requests killed on budget between 2021 and 2024. A meaningful share of them now clear the threshold.
- Start where spreadsheets and email still run production. Inventory reconciliation, quality logs, maintenance scheduling, supplier onboarding. Highest ratio of value to effort.
- Budget verification explicitly. With 46% of developers distrusting AI output accuracy, review hours belong in the quote — not discovered later as rework.
- Sequence integration before agents. An agent without governed access to your data and business rules is a demo, not a system.
- Hire for judgment, not keystrokes. The scarce profile is the engineer who understands your process, your compliance environment, and your risk tolerance — and can validate what the model produced.
- Treat software capacity as capacity, not cost. If your team can now ship three projects instead of one, the question is which two additional problems are worth solving.
The paradox worth holding onto
The better AI gets at producing software, the larger the universe of software organizations will want to produce. On that reading, the future of development is defined less by a contest between humans and machines than by a different economics: when building technology costs less and takes less time, more problems become worth solving with it.
The hypothesis is reasonable, not proven. The trust gap — 46% against 33% — is a live indicator that the industry is still calibrating how much it can safely delegate. But the ten-year U.S. labor projection currently sides with expansion.
Frequently asked questions
Will AI replace software developers in the U.S.? The federal projection points the other way for the developer role. BLS projects 15 percent employment growth for software developers, QA analysts, and testers from 2024 to 2034, with roughly 129,200 annual openings. Computer programmers, a narrower occupation, are projected to decline 6 percent over the same period.
What is agentic AI, and how is it different from generative AI? Generative AI responds to a specific prompt. An agent receives an objective, interprets context, interacts with multiple systems, and executes a sequence of tasks to reach it.
Are AI agents ready for industrial production environments? Not broadly. 52% of developers don't use agents or stay with simpler tools, and 38% have no plans to adopt them. Data governance, system integration, and security controls need to be in place first.
Why is developer trust in AI falling while adoption rises? The leading frustration is output that is almost right but not quite — cited by 45% of respondents — with 66% reporting more time spent correcting AI-generated code.
What's the real business impact beyond cost savings? The larger effect isn't cheaper delivery of current projects. It's the expansion of which projects are viable at all: internal apps, automations, and legacy modernization that previously failed the budget test.