When Intelligence Becomes Cheap, What Happens to Ordinary People's Jobs?
A question that matters to me more than model rankings
On September 10, 2026, DeepSeek officially released V4.1 Flash and announced a corresponding reduction in API prices. It brought me back to a question I had when V4 Flash first appeared: when a model is reasonably capable and very cheap to use, how should we think about its effect on everyday work? [1]
In developer communities, every new model naturally prompts questions about how much better it is at coding, whether it can solve harder problems, and whether it stays reliable over longer tasks. Those discussions matter. As a developer, I care about those capabilities too.
But much of the work that keeps a business running does not require a model at the top of the rankings.
Organizing Excel spreadsheets, consolidating data from different sources into reports, producing documents from existing material, comparing two versions, or turning meeting notes into a task list. These are ordinary tasks, yet someone has to keep spending time on them.
Once a model can handle a substantial share of that work, at a price low enough for frequent use and repeated attempts, the question becomes very practical: how many full-time employees will companies still assign to this work?
The work certainly has value. Data needs organizing, documents need producing, and business needs moving forward. But whether work needs to be done and whether a company will retain a position to do it are two separate questions.
Once capability is good enough, cost changes where it gets used
Even a very powerful model may be reserved for a handful of important tasks if every use is expensive. When the cost is low enough, small tasks that once seemed unworthy of automation can become candidates for experimentation.
Consider a small report produced once a week. In the past, it might not have justified paying a developer to maintain dedicated software. Having an employee spend an hour preparing it manually was tedious but easy to arrange. If a model can interpret material in different formats and organize it with existing tools, the barrier to trying automation may fall.
I believe that once capability crosses the threshold of practical usefulness, falling costs will broaden the range of work affected. A business does not need to buy the strongest reasoning capabilities for every task. It needs a solution capable enough to finish the job and cheaper than the current approach.
Companies will ultimately compare the total cost of getting the work done: model calls, systems integration, checking results, correcting errors, and handling exceptions. An employee’s ability to chase missing data and reconcile definitions across departments belongs in that calculation too. My concern is this: as more of those steps become automated, how much of the cost justification for existing staffing levels will remain?
Imagine a team that originally needed five people to handle routine material. If, after automation, two people can do the remaining checking, coordination, and exception handling, the company has an incentive to change its staffing. This hypothetical illustrates the mechanism that concerns me: as long as the remaining work can be concentrated among fewer people, demand for labor can fall.
AI does not have to perform every part of a job independently for that job to be affected.
The first to lose out may be people who never get hired
I believe AI’s earliest impact on employment is likely to show up as fewer opportunities to get hired.
Companies may be in no hurry to dismiss existing employees. Those people know the business, keep daily operations running, and do plenty of work that has yet to be automated. But when a department asks for more staff, managers have an increasingly easy request to make: try AI first and see whether the current team can take on the work.
A position that was about to be advertised gets put on hold. There is no layoff announcement, and nobody leaves, but someone who might have gotten that job loses the chance to join the organization.
Recent months of U.S. employment data have made me more concerned about this shift. Earlier reports of weakening employment, followed by another strong nonfarm payroll report, show that short-term figures remain subject to fluctuations and revisions. But resilience in total employment does not ease my concern about access to jobs. Companies can retain existing employees while remaining cautious about adding staff; growth in food service and education does not mean ordinary office jobs offer the same opportunities. [2][3]
My short-term assessment is that companies may remain in a pattern of limited layoffs and limited hiring. Existing employees’ experience is still useful, and integrating AI into the business takes time. At this stage, companies can keep operating while testing which additional tasks no longer require additional hires.
For employers, this is a relatively inexpensive experiment. For job seekers, the consequences are already real. Graduates cannot find their first job, and people changing careers wait for opportunities that do not arrive. Those already employed may feel that the job market has not changed much. Those outside the door can already feel it narrowing.
What worries me more is that this may only be a transition.
For now, a model’s ability to organize spreadsheets and generate documents does not mean it can take over an entire job. Information is scattered across systems. Processes need coordinating, anomalies need addressing, and someone must be accountable for the results. These requirements keep companies reliant on existing staff for the time being.
But if agents become increasingly suited to different workflows in the period ahead, retrieving information, operating software, checking results, and handing a small number of exceptions to people, the conditions for companies to recalculate their staffing needs will gradually fall into place.
At that point, “We won’t hire for this new position yet” may become “We don’t need to replace this person when they leave,” and eventually, “Does this team still need to be this large?”
I worry that the absence of a clear rise in layoffs today may lead us to underestimate the risk of fewer jobs tomorrow. Today’s stability may reflect a lag while companies experiment, connect their workflows, and reorganize. We cannot treat another strong jobs report as evidence that ordinary jobs are secure for the long term.
This does not mean I can use a few employment reports to prove that AI has already caused a hiring slowdown. I am making a judgment about the direction ahead: as models become useful enough and cheap enough, and agents gradually bring those capabilities into business processes, the incentive to reduce staffing needs will become increasingly concrete.
Are we using this time to create new routes into work, training opportunities, and income protection for the people who may be affected? Or will we wait until staffing starts shrinking before telling them to learn new skills and find more valuable work?
I worry that by the time the impact becomes unmistakable in layoff figures, many people’s opportunities will already have disappeared through one hiring request after another being put on hold.
Agents are bringing companies closer to rethinking staffing levels
I believe the key to when this impact reaches existing staffing levels is how quickly agents become integrated into real workflows. Model capabilities and costs have already given companies a reason to experiment. The next step is connecting the individual stages of everyday business operations.
Take a monthly business report. The actual work might involve retrieving information from different systems, checking data definitions, spotting anomalies, contacting people for missing information, updating spreadsheets, and finally submitting the report for review. Every step connected to the workflow may remove another need for someone to transfer information manually, check a status, or follow up on progress.
By agent, I mean an AI system that can call tools, execute multiple steps, and use the results to continue a task. To work within a company’s daily processes, it also needs appropriate permissions, the ability to stop and hand work to a person when necessary, traceable records, and manageable recovery costs when something goes wrong.
These are engineering and management problems companies need to solve when implementing AI. My concern is that we are treating integration work that remains unfinished today as a reason certain jobs will endure. As the connections between business data, enterprise software, and human review mature, companies will find it easier to translate model capabilities into staffing decisions.
Within such a workflow, a person’s role may shift from performing every step to reviewing the output of several agents and handling a small number of exceptions. The work still exists, but fewer people may be needed to do it.
This process may take time and will differ across industries. Work involving disorganized data, numerous exceptions, and significant accountability requirements will take longer to transform. Clearly defined processes with inputs and outputs that are easy to verify are better candidates for early experiments.
Whether new business demand can absorb the labor freed up by these changes must be answered by actual hiring. I am unwilling to treat “greater efficiency will naturally create more jobs” as a promise that has already been fulfilled. For people losing opportunities, when and where new jobs appear, and what skills they require, determine whether those jobs are actually within reach.
In the future, growth in a company’s business may no longer produce the same proportional growth in hiring that it once did.
If entry-level work shrinks, where does experience begin?
A familiar response at this point is that people can move on to more valuable work.
That answer troubles me because it skips the most practical questions: how does someone develop the ability to do that work, and who will make sure they have the time and income they need while learning?
Many careers begin with organizing information, preparing documents, and checking spreadsheets. Through that work, people gradually learn the business, discover which figures tend to go wrong, understand what customers really care about, and learn when to ask further questions and when to escalate an issue.
As these basic tasks become extensively automated, I worry that newcomers will also lose opportunities to build experience. Companies need to explain how those people will get enough practice.
An organization can rely on a few experienced people to review AI output in the short term. It still needs to train the next group of people who can review, exercise judgment, and take responsibility. If entry-level positions keep shrinking without new ways to develop talent, the supply of experienced workers may become a problem a few years later.
Companies and the education system need to design those paths together. Beyond teaching people to use tools, training must provide opportunities to work on real business problems, handle exceptions, and receive feedback. Otherwise, “learning AI” may simply give everyone similar tools without creating enough routes into a career.
Smaller teams and harder questions about who benefits
If a small team using AI can do what once required an entire department, companies have room to rethink their size and the way they organize jobs. Starting a business or providing professional services may become easier. But my greater concern is how many ordinary people such organizations will still offer stable employment.
At the same time, companies that already have customers, data, capital, and distribution channels can use the same tools to expand. The future may bring both more small teams and companies running larger businesses with fewer employees.
I worry that as intelligence becomes cheaper to use, customers, data, capital, and distribution channels will matter even more in determining who captures the returns. Someone who loses an office job may have access to the same inexpensive model without having customers, startup capital, or the financial capacity to weather an unpredictable income. Asking everyone to become an independent entrepreneur is no substitute for society’s arrangements for employment and social protection.
Companies might use the gains to lower prices, expand, raise wages, or shorten working hours. They might also use them primarily to increase profits. Those choices mean very different things for ordinary people’s lives.
If we continue to rely mainly on full-time employment to distribute income, protection, and opportunities to participate in society, changes in the demand for workers become more than an internal efficiency issue. A person may lose a stable income, a path to building experience, and their expectations for the future.
When I ask whether society can prepare, these are the questions I mean: can people retrain for a different career without losing their income? Can newcomers get real opportunities to grow? Can workers share in productivity gains? If the same output requires fewer working hours, can we turn some of that saving into more breathing room in people’s lives?
In my view, these preparations must happen before jobs are restructured on a large scale. A model can see rapid adoption after a single update. Rebuilding someone’s professional capabilities or redesigning a society’s support systems takes much longer. That difference in speed is what worries me most.
Before the next hiring request is put on hold
V4 Flash, and now the V4.1 Flash update, have made me increasingly concerned that once model capabilities and costs together cross the threshold of practical usefulness, companies may adjust staffing faster than society can create new paths forward. Businesses do not need to wait for AI to do everything before recalculating how many people they need.
At first, the change may be just one more hiring request put on hold. One company hires one fewer person. One team does not replace someone who leaves. Each decision looks small and has its own business rationale. But for people waiting to enter the workforce, opportunities that never materialize can change a life too.
As agents gradually take over entire workflows, those decisions may reach further into existing staffing levels and organizational structures.
I believe we already need to take this question seriously: as everyday work requires fewer people, how will we ensure that ordinary people still have access to income, opportunities to grow, and opportunities to participate in society?
Sources
The following sources are cited based on information available as of September 10, 2026; monthly employment figures may be revised further. The judgments about hiring, staffing, and future organizational structures express my views and concerns. They do not imply that the cited data has established these changes.