AI is abundant. Reliable AI labor is not.
I first started thinking about this through a question that sounds almost childish: if Earth has so much water, why do people still live without clean drinking water?
From space, the planet looks almost impossibly blue. Oceans wrap around continents. Clouds move above them. Rivers cut through land. It feels strange that a planet covered in water can still have a water crisis.
But the contradiction disappears once we separate existence from usability.
About 71% of Earth’s surface is covered by water, and oceans hold around 96.5% of all the water on the planet. Only around 2.5% is freshwater, and much of that is locked away in glaciers, ice caps, or underground reserves. The water exists, but most of it is not immediately usable by the people who need it.[1]
That distinction stayed with me.
Water is abundant. Drinkable water is not.
Later, when I started building and thinking more seriously about AI systems, the same pattern kept showing up in a different form. We now have more machine-generated intelligence than at any point in history. Models can write, summarize, code, translate, classify, search, reason, and interact with tools. Every product seems to have an AI layer. Every company seems to have an AI strategy. Every founder deck has some version of the same promise: fewer people, more output, faster execution.
And in many ways, that promise is real.
AI can already do things that would have felt impossible a few years ago. The cost of capable inference has also dropped quickly. Stanford’s 2025 AI Index reported that the inference cost for a system performing at a GPT-3.5-like level fell more than 280-fold between late 2022 and late 2024.[2]
So the obvious question becomes: if intelligence is becoming this cheap and this available, why has AI not replaced far more human work already?
I do not think the answer is simply capability.
I think the answer is conversion cost.
By “desalination problem,” I do not mean that AI needs one magical process that makes it safe or useful forever. I mean that raw abundance has to be converted into usable reliability, and that conversion has a cost. For AI, that cost shows up as evaluation, grounding, workflow design, human review, monitoring, accountability, infrastructure, and trust.
The ocean is full of water, but the work is turning it into something people can drink.
AI output is abundant, but the work is turning that output into something people can depend on.
That is the desalination problem of AI.
Capability is not replacement
A demo proves that something is possible. A business process has to survive reality.
That is where a lot of AI conversations become confused. When we watch an AI system perform a task once, the question is simple: can it do it? But when a company depends on that system every day, the question becomes much heavier.
Can it handle messy inputs? Can it work when the user is angry, vague, or wrong? Can it deal with exceptions? Can it explain why it made a decision? Can it fail safely? Can it know when not to answer? Can it be monitored, audited, corrected, and trusted?
These are not small details. They are the difference between output and outcome.
A model can generate a customer-support reply in seconds, but a support operation needs account history, refund rules, escalation paths, CRM access, tone guidelines, legal boundaries, logging, monitoring, and fallback. A model can generate code quickly, but production-safe code still needs tests, architecture awareness, security review, deployment discipline, and someone accountable when it breaks. A model can summarize a legal document, but a legal decision still needs context, judgment, verification, and liability.
Companies do not really pay for text, code snippets, summaries, or recommendations. They pay for trusted results: a ticket resolved, a bug fixed, a contract reviewed, a lead qualified, a patient note made accurate, a hiring decision made fair and explainable.
So the useful question is not only whether AI can do a task. The better question is whether AI can deliver the outcome at a lower total cost than a human.
That is where the economics begin.
The real unit is a reliable outcome
A common mistake is comparing the salary of a human employee to the subscription cost of an AI tool. That comparison is convenient, but too shallow.
A human is not just a salary. There are hiring costs, training, management, benefits, inconsistency, mistakes, attrition, onboarding time, communication overhead, and coordination drag.
But AI is not just an API call either. AI also brings integration, prompting, evaluations, monitoring, privacy controls, tool access, hallucination management, human review, compliance, maintenance, incident response, user trust, and model drift.
The fair comparison is not human salary versus AI subscription. It is
human full-stack cost per reliable outcome vs AI full-stack cost per reliable outcome.
That phrase - reliable outcome - is the important part.
Not output. Not token count. Not task completion in a demo. A reliable outcome means the business problem was actually solved to the standard the task requires.
For some tasks, that standard is low. If AI writes a mediocre first draft of a social media caption, the cost of being wrong is small. For other tasks, the standard is extremely high. If AI misses a legal risk, approves a defective machine part, misclassifies a medical note, or rejects a candidate unfairly, the cost of being wrong can dominate the cost of doing the work.
This is the AI version of potable water. Intelligence in the abstract is not enough. The question is whether the intelligence can be safely used where the work actually happens.
A simple way to express the replacement point is:
AI replaces human labor when the total cost of an AI system at the required reliability is lower than the total cost of human labor at the required reliability.
Inside that sentence is most of the hard part: compute, integration, supervision, correction, security, compliance, maintenance, trust, and the cost of failure.
The reliability tax
If raw model output is seawater, the reliability tax is the energy bill.
Desalination is not expensive because water is rare. It is expensive because making water usable requires energy, infrastructure, filtration, distribution, maintenance, and quality control. AI has a similar tax. The model may produce an answer cheaply, but making that answer dependable can require evaluations, domain constraints, retrieval, human review, monitoring, audit logs, fallback paths, and continuous calibration.
For low-stakes work, this tax can be small. For high-stakes work, it can become the main cost of the system.
Consider legal document review. A simple AI workflow may read a contract, identify clauses, summarize risks, and produce a report. That is useful. But if the system misses an important liability clause, the cost of that miss may be much larger than the cost of the review itself.
So the question is not only, “How much does the AI cost to run?” The real question is, “How much does the AI cost after mistakes are priced in?”
Imagine 100 contracts need to be reviewed. A lawyer reviews them at 95% accuracy and costs $10,000. A simple AI workflow reviews them at 85% accuracy and costs $500. At first, AI looks 20 times cheaper. But that comparison is incomplete. If each missed issue costs only $50, the AI system may still win. If each missed issue costs $50,000, the cheap system becomes very expensive.
This is why accuracy alone is not enough. The cost of error changes the economics.
There are narrow legal tasks where AI has already performed very well. In a well-known NDA review study by LawGeex, an AI system reportedly achieved 94% average accuracy compared with 85% for lawyers, while completing the task much faster.[3] That is important because it shows that AI can beat humans in constrained, well-defined workflows.
But the lesson is not that AI is always better than lawyers. The lesson is that task shape matters.
AI is strongest when the task is narrow, the success criteria are clear, the input format is stable, and the output can be checked. The more the task moves from extraction to judgment, from issue-spotting to advice, and from draft to accountability, the more expensive reliability becomes.
Going from a poor system to a useful system may be cheap. Going from useful to dependable often requires serious engineering. Going from dependable to trusted in a high-stakes environment can require process design, domain experts, evaluations, monitoring, legal accountability, and human review. The last few percent of reliability are often where the economics become difficult.
A model can be cheap at 85% accuracy and expensive at 95%. It can be impressive at 95% and still unacceptable at 99%.
That is the reliability tax.
The cost of failure scales differently
Humans make mistakes too. Any serious argument about AI has to admit that.
Humans forget, misread, rush, misunderstand, overpromise, and confidently say wrong things. AI is not uniquely flawed because it can be wrong. The issue is that AI failure has a different shape.
It can be confident. It can be non-obvious. It can be hard to trace. It can repeat the same error many times. And because AI systems scale quickly, their mistakes can scale quickly too.
A human support agent might make one bad refund decision. A poorly designed AI workflow can make thousands of bad refund decisions before anyone notices. A human reviewer might miss one legal citation. An AI-assisted workflow can generate a pattern of mistakes that looks polished, formatted, and authoritative.
That is one of the uncomfortable parts of AI. It can make work feel cheaper while making failure more scalable.
For low-risk workflows, this may be acceptable. For high-risk workflows - finance, healthcare, legal, hiring, infrastructure, security, safety inspection - the cost of being wrong can dominate the cost of doing the work.
This is where many AI replacement arguments become too simple. They focus on the cost of generation and ignore the cost of verification. But in many knowledge-work systems, verification is the expensive part.
The rubber-stamp problem
There is another hidden cost of AI that is harder to measure: what it does to the human beside it.
Most serious AI systems are sold with some version of the same reassurance. The AI will draft, summarize, inspect, classify, flag, or recommend. A human expert will review the result before anything important happens.
That sounds safe. But it depends on what we mean by review.
A human in the loop is not enough. The human has to be cognitively in the loop.
At first, the reviewer may check the AI carefully. They compare the summary with the source. They inspect the raw document. They look at the evidence. They ask whether the model missed something. Then the AI is correct ten times, then fifty times, then two hundred times. Slowly, the human starts to trust the pattern. The review becomes faster. The source is opened less often. The raw evidence is checked only when the AI sounds uncertain.
The human stops asking, “Is this correct?” and starts asking, “Does this look reasonable?”
That is a dangerous shift.
The workflow still has a human reviewer on paper, but in practice the human can become a rubber stamp. This is not because people are lazy in a moral sense. It is because people adapt to tools. If a system is usually right, the brain conserves attention. If the interface presents an answer confidently, people treat it as more trustworthy. If the organization rewards speed, review slowly becomes approval.
Human-factors researchers have studied automation bias and automation complacency for decades. NIST’s Generative AI Profile also lists risks such as automation bias, over-reliance, and people making poor decisions based on erroneous AI outputs.[4] Modern AI makes the problem more subtle because the output is not just a warning light or a number. It is a fluent explanation. It sounds like judgment.
Imagine a safety inspection system for a machine part. The AI scans images, checks measurements, identifies possible defects, and generates a report. The engineer receives the report and signs off after review. On paper, the engineer is the safety layer. But what if the engineer mostly reads the AI summary? What if the AI misses the defect? What if the interface says “No critical issues found” with confident language? What if the engineer has hundreds of reports to clear that day? What if management rewards speed more than independent inspection?
In that case, the human is no longer a safety layer. The human is an accountability layer.
The AI generated the miss. The human signed the approval. The system failed between them.
This pattern is already visible in law and software. Lawyers have been sanctioned for submitting AI-generated legal citations that did not exist. Stanford researchers found that general-purpose language models hallucinated heavily on certain legal queries, with rates varying by model and task.[5] Developers know the smaller version of this problem too: accepting generated code because it looks right, then discovering later that they never really understood it.
Bad AI workflows create mistakes. Bad human-AI workflows create mistakes that look reviewed.
So “human in the loop” is not a safety guarantee. It is a design challenge.
A good AI workflow should keep the human anchored to the underlying evidence. It should show uncertainty, reveal what the model checked and did not check, make disagreement easy, require deeper verification for high-risk cases, and avoid turning the AI summary into the only thing the human reads.
The goal is not to remove AI from high-stakes work. The goal is to prevent AI from quietly turning experts into approvers.
The margin problem of AI startups
The same conversion cost also appears in startup economics.
It is not enough to access the ocean. You have to purify and deliver the water at a price the market will pay.
Classic SaaS has beautiful margins because serving one more user is often cheap. Once the software is built, usage usually helps. More customers usually means more revenue, and the marginal cost of serving them is relatively small.
AI SaaS is different. Every user action may trigger real compute cost.
In normal SaaS, more usage is usually good. In AI SaaS, more usage is good only if revenue grows faster than inference cost.
That changes the business model.
A lot of AI startups today are built from a familiar stack: a frontier model, a thin workflow UI, some prompt engineering, maybe retrieval, and a subscription page. That can still become valuable if the product owns distribution, workflow, data, trust, or a painful niche. But if the startup owns none of those things, it is fragile.
The model provider can copy the feature. Competitors can rent the same intelligence. Users can switch to a cheaper tool. Margins can disappear as usage grows.
“AI wrapper” is not automatically an insult, but it is a warning. A wrapper can become a real business if it becomes the place where work happens. If it owns the workflow, captures feedback, builds trust, improves over time, and becomes hard to remove, it can become valuable. But a wrapper that only forwards prompts to a model and returns formatted output is living on borrowed intelligence.
The winning AI companies will understand inference economics. They will know when to use frontier APIs, when to use cheaper models, when to cache, when to route tasks, when to self-host, and when not to use AI at all.
The naive version is to use AI everywhere. The better version is to use the right amount of intelligence for the value of the task.
Not every task needs the strongest model. Extracting an email from a resume may need regex. Summarizing a section may need a cheap model. Scoring a resume against a job description may need a stronger model. Handling a disputed result may need a frontier model and human review.
The future of AI products is not just better prompts. It is routing, evaluation, caching, monitoring, feedback loops, and cost control.
AI products are not only product businesses. They are margin businesses.
Why demos feel magical and deployments feel messy
There is a reason AI demos are often more impressive than AI deployments.
Demos are clean. Real work is not.
A demo has a clear prompt, a controlled environment, and a forgiving audience. The model only needs to show possibility. A deployment has unclear inputs, legacy systems, edge cases, angry users, incomplete data, internal politics, privacy constraints, and business rules that exist in someone’s head but not in any database.
This is why companies can be excited about AI and disappointed by AI at the same time.
McKinsey’s 2024 global AI survey found that 65% of respondents said their organizations were regularly using generative AI, nearly double the share from around ten months earlier.[6] Adoption is real. This is not just hype.
But adoption does not automatically mean deep transformation.
A company can use AI everywhere and still not redesign the core workflow. Employees may use AI to draft, summarize, brainstorm, or search while the actual business process remains mostly human-owned. That is not failure. It is usually the first stage of diffusion.
AI enters as a tool before it becomes infrastructure. It becomes a helper before it becomes an operator. It becomes a productivity layer before it becomes a replacement layer.
AI will replace tasks before it replaces jobs
The strongest argument against a simple “AI will not replace humans yet” view is that AI does not need to fully replace humans to reduce human labor.
A company does not need one AI agent to replace one employee. It only needs AI to make a team smaller.
Before AI, ten analysts might produce a hundred reports per week. After AI, four analysts with good tools might produce the same number, or more. No single analyst was replaced by a robot sitting at a desk, but six future roles disappeared from the hiring plan.
This is probably how much of the first wave of AI labor impact will appear: not always as dramatic layoffs, but as hiring freezes, smaller teams, fewer junior roles, higher output expectations, more pressure on existing employees, and more work routed through senior reviewers.
That is still labor replacement. It just does not look like the science-fiction version.
This also rhymes with what economists describe as task-based automation: technology does not usually replace a job as one complete block. It changes the economics of specific tasks inside the job. Some tasks become cheaper, some become unnecessary, some become more valuable, and some new tasks appear around the new technology itself.
Goldman Sachs estimated in 2023 that generative AI could expose the equivalent of 300 million full-time jobs globally to automation.[7] The word “expose” matters. It does not mean all those jobs disappear. It means parts of those jobs contain tasks that could be automated or transformed.
Jobs are bundles of tasks. AI attacks the bundle unevenly.
Some tasks disappear. Some become easier. Some become more valuable. Some become bottlenecks. Some new tasks appear around supervising, evaluating, integrating, and governing the AI system itself.
So the future of work is not simply humans versus AI. It is a rebundling of work.
The dangerous middle
The most interesting period is not the world before AI replacement or after AI replacement. It is the middle.
The middle is where AI is good enough to be useful but not reliable enough to be left alone. This is where we are now in many domains.
In this phase, AI can generate a lot of work quickly, but humans still have to judge it. That changes the shape of expertise. The human is no longer only a creator. The human becomes a reviewer, editor, verifier, architect, exception handler, and accountability layer.
This can be powerful, but it can also be exhausting.
Anyone who has used AI for serious work knows the feeling. The model gives you something plausible. It is not obviously wrong. But you cannot fully trust it. So now you are reading with suspicion, checking facts, testing code, looking for missing assumptions, and asking whether the confident answer quietly skipped the hard part.
In low-stakes work, this may still be a win. In high-stakes work, this can become a tax.
AI speeds up generation, but it increases the need for judgment.
And judgment becomes the scarce resource.
What this means for people
This does not mean humans are safe. It means the danger is more specific.
AI will not replace everyone at once. It will replace the parts of work where reliable outcomes can be produced cheaper by machines than by people. That is still a very big deal, especially for people whose work is mostly predictable, digital, repetitive, text-based, and easy to evaluate.
But it also means the most valuable human skills will shift.
The future belongs less to people who merely produce output and more to people who can define outcomes. People who can ask better questions. People who can judge quality. People who can understand systems. People who can handle ambiguity. People who can design workflows. People who can own accountability. People who can decide what should not be automated.
In a world of abundant generation, taste becomes valuable. In a world of cheap answers, judgment becomes valuable. In a world of fast output, responsibility becomes valuable.
What this means for companies
For companies, the lesson is simple: do not ask only where AI can be inserted. Ask where reliable outcomes are expensive.
Then ask whether AI can reduce the full-stack cost of those outcomes.
That framing avoids two mistakes. The first mistake is rejecting AI because it is imperfect. Humans are imperfect too. If AI is cheaper, faster, and good enough in a low-risk workflow, it should be used.
The second mistake is adopting AI because the demo looks impressive. A demo is not an operating model. A chatbot is not a department. A generated answer is not a resolved business process.
The companies that win with AI will not be the ones that put AI everywhere. They will be the ones that know where AI actually changes the economics.
The final thought
I do not believe AI replaces humans the moment it becomes capable.
I believe AI replaces humans when capability becomes economically reliable.
That distinction matters because the AI debate is often framed as a question of intelligence. Can it reason? Can it code? Can it write? Can it plan? Can it act?
Those questions matter, but they are not the final questions.
The final question is whether it can own the outcome.
Until then, AI is leverage. Powerful leverage, dangerous leverage, transformative leverage - but still leverage. The moment it can own outcomes cheaper than humans, the conversation changes.
That is when the ocean becomes drinking water.
That is when abundance becomes usable.
That is when AI stops being a tool beside the worker and starts becoming the worker.