The chatbot is a terrible way to measure the future of AI.
I started thinking about what happens when everyone has a personal AI agent. Not a chatbot you occasionally ask questions, but something that actually works on your behalf. It researches purchases, arranges appointments, compares insurance, negotiates with suppliers, manages administrative tasks, and generally takes care of the thousands of little things that consume our time. There are billions of potential users. That's already a substantial market.
But then I started thinking about the other side of those interactions. If I have an AI agent contacting a business, who exactly is going to answer it? The business will have its own agents. One might handle customer service. Another verifies identity. Another checks inventory. Others handle pricing, scheduling, fraud detection, payment processing, logistics, and compliance. My one agent could easily generate work for ten more.
And those ten might need to communicate with another company's systems. Suddenly, the number of interactions taking place between machines could dwarf the number of interactions between humans and machines. That was the original thought.
One person. One agent. Ten more agents on the other side.
Not a prediction that every interaction requires precisely ten agents. Just the recognition that demand could multiply across both sides of a transaction. Then the architecture started changing. Systems can divide objectives among specialized AI workers. Very large swarms are a possible architecture, not a description of typical enterprise deployment. Instead of answering a question, an AI system can break a problem into pieces, distribute work, compare competing solutions, test outcomes, and commission additional research.
At that point, estimating the future of artificial intelligence by counting chatbot users starts looking rather like estimating the size of the internet by counting how many emails people send. The more important question becomes:
How much economically valuable work will machines eventually perform for other machines?
I suspect the answer is vastly greater than today's conventional understanding of the AI market. And if that suspicion is correct, the consequences reach far beyond software. They reach into electricity, industrial manufacturing, communications networks, physical infrastructure, and businesses that have never described themselves as AI companies. This is the Agent Multiplication Effect.
I. From answering questions to accomplishing objectives
The first commercial wave of generative AI was built around a familiar interaction. A person asks a question. A model produces an answer. The interaction ends. Even when the model performs something sophisticated, such as writing software or analyzing financial statements, a human generally initiates and supervises the work. Agents introduce a different relationship. Instead of providing detailed instructions for each individual step, a person assigns an objective.
The system identifies tasks, accesses tools, gathers information, executes actions, and evaluates progress. The distinction sounds minor until you consider the economics. A chatbot might help an employee prepare a supplier comparison. An agentic system might be asked to evaluate every supplier the company uses, identify contract discrepancies, investigate alternatives, calculate switching costs, and return a prioritized set of recommendations.
The second assignment can generate orders of magnitude more computational activity than the first, depending on the scope and implementation. And it may execute much of that work without a human initiating each individual operation. This is no longer an entirely speculative direction.
McKinsey's November 2025 global survey found that 62% of respondents reported their organizations were experimenting with or scaling AI agents. But only 23% reported scaling agentic systems in at least one business function, and deployment across individual functions remained limited.
That distinction is important. Experimentation is not mature commercial adoption. But it establishes that organizations are actively exploring a different way of using artificial intelligence.
Anthropic's June 2026 Economic Index provides another useful signal. The company reported that Claude sessions increasingly consisted of long-running agentic tasks with the growth of Claude Code and Cowork. This is evidence from one provider, not a census of the AI economy. It also acknowledged that traditional chat transcripts were becoming insufficient for understanding how people actually use these systems.
Source: Anthropic, Anthropic Economic Index: Cadences, June 26, 2026.
That last observation is especially revealing.
The way we measure AI usage may already be falling behind the way AI is being used.
If more work happens across autonomous sessions, tools, and delegated subtasks, counting human conversations increasingly misses part of the picture.
II. The multiplication begins when agents meet agents
Consider a relatively ordinary transaction. A consumer wants to purchase a vehicle. Their personal agent receives an objective: find an appropriate vehicle, compare financing and insurance, and arrange a purchase within an agreed budget. The personal agent contacts dealerships. The dealership's systems check inventory, pricing, promotions, and available delivery dates. Financing requests reach lenders. Insurance quotes are obtained. Identity, fraud, and payment checks occur. Logistics systems coordinate delivery.
A single purchasing decision can involve numerous companies and systems, each performing its own work. That isn't entirely new. Existing commerce already runs on enormous networks of APIs, databases, payment processors, and automated business software. The important potential change is that AI makes a much wider range of previously human-mediated activities economically automatable. Negotiation, interpretation, unstructured research, exception handling, and coordinating complex workflows may become increasingly accessible to software.
Those capabilities can increase the number of tasks worth performing. An insurer might evaluate more potential policies. A customer might compare more suppliers. A manufacturer might monitor more operational variables. A retailer might adjust purchasing decisions more frequently. And one agent's output can become another agent's input. The multiplier isn't necessarily the number of independent AI personalities involved.
It's the amount of computational work produced by the original economic objective.
A process that previously required five database queries might now require ten model calls, three external tools, and a verification step. Or improved technology might collapse those operations into one highly efficient model call. Both outcomes are possible. What matters is that the relationship between human activity and machine computation is changing. We need to stop assuming that one human interaction corresponds to one meaningful unit of AI demand.
III. The swarm changes the scale of the question
Now move beyond ordinary commercial transactions. Imagine a business asks its AI system to identify every material inefficiency in its supply chain. A single model could attempt to answer. Alternatively, the system could divide the assignment into thousands of subtasks. Different workers analyze suppliers, purchasing histories, contracts, freight expenses, inventory levels, market prices, and operational constraints. Other workers independently challenge their conclusions. Some simulate alternative purchasing arrangements. Some investigate inconsistencies.
Others assemble findings into a coherent plan. The architecture begins to resemble an organization conducting a large research project rather than a person asking a chatbot a question. Theoretically, these systems can become very large. But there's an important complication.
More agents do not automatically mean better results.
In January 2026, Google Research published findings from a controlled evaluation of 180 agent configurations across different models, tasks, and organizational structures. The results showed that multiple agents could improve performance on tasks that benefited from parallel work. But on certain sequential tasks, adding agents actually made performance worse. Coordination imposed costs. Communication consumed resources. Errors could spread across workers.
The researchers also found that architecture mattered substantially: different arrangements produced different levels of performance and error propagation.
Source: Yubin Kim and Xin Liu, Google Research, Towards a Science of Scaling Agent Systems: When and Why Agent Systems Work, January 28, 2026. The underlying research paper, Towards a Science of Scaling Agent Systems, is available publicly.
That finding doesn't weaken the central economic question. It makes the question more precise. The future may not belong to the company deploying the most agents. It may belong to the company that can organize machine intelligence most effectively. Some tasks will use one powerful model. Others will involve specialized models. Some may benefit from extensive parallelization. Others will require strict coordination and verification. The winning architecture will depend on the task.
And organizations will have powerful incentives to discover which arrangements create the most useful work for the least cost.
We should measure productive computation, not count digital workers.
IV. The efficiency paradox
Here's where the economics become particularly interesting. One of the strongest arguments against sustained compute demand is that artificial intelligence is becoming dramatically cheaper. That's true.
Stanford's 2025 AI Index documented a more than 280-fold decline in the price of querying a model at roughly GPT-3.5-level MMLU benchmark performance between November 2022 and October 2024.
The quoted inference price fell from $20 to $0.07 per million tokens. This measures price at a benchmark level, not a 280-fold reduction in electricity per task.
At first glance, that sounds bearish for infrastructure demand. If models become hundreds of times cheaper, shouldn't we need proportionately less hardware and electricity? Possibly. But falling costs can also unlock demand that didn't previously exist. Imagine a business process that requires an hour of analysis but produces only a few dollars of potential value. Hiring someone to perform that analysis makes no economic sense. Using an expensive AI system might not make sense either.
But if the same analysis can be performed reliably for a fraction of a cent, the calculation changes. Now imagine that business has ten million similar decisions available to investigate. A task that was previously uneconomical becomes an opportunity for continuous automation. And once the company starts analyzing those decisions, it might discover entirely new categories of useful work.
This is where the logic begins to resemble Jevons paradox.
The rebound mechanism is straightforward: greater efficiency can lower costs enough to expand usage. It is a possibility, not a law requiring demand to rise. The same economic mechanism could operate in artificial intelligence. Cheaper computational intelligence creates more potential customers and applications. More applications create additional work. Additional work creates demand for computation. The essential question is whether expanded usage outweighs efficiency improvements. There's no guarantee it will.
But the potential feedback loop is enormous.
If intelligence becomes cheap enough to apply everywhere, demand may stop being constrained primarily by how often humans choose to ask questions.
Instead, it becomes constrained by how much useful work exists to be performed. And that market may be substantially larger.
V. What happens when intelligence becomes a production input?
For much of modern economic history, expanding production meant adding some combination of people, machinery, capital, energy, and physical resources. Software changed that relationship. One software product could serve millions of customers with comparatively low marginal distribution costs. Artificial intelligence introduces another possibility. Instead of merely distributing software, businesses may increasingly purchase computational work. Research, classification, analysis, design, customer communication, forecasting, and countless other activities can potentially be performed as services.
This doesn't eliminate human judgment. It does change the economics of getting certain kinds of work done. Consider a manufacturer operating hundreds of machines. Today, it might perform periodic maintenance checks and analyze selected failures. In a more automated environment, it could continuously investigate equipment behaviour, evaluate emerging problems, simulate maintenance decisions, and compare its findings against operational records.
Now consider the same concept applied across transportation fleets, agricultural equipment, energy infrastructure, hospitals, financial institutions, and logistics networks. These applications would differ enormously in their technical requirements and commercial viability. Some would use specialized small models running locally. Others would require substantial cloud computation. Many would be limited by sensor quality, data availability, regulation, or the difficulty of acting on recommendations. But they share a common economic possibility:
When the marginal cost of analysis falls sufficiently, businesses can afford to investigate problems they previously had to ignore.
That is the expansion of computational demand we consider most important. Not necessarily people asking more questions. Machines performing more economically justified work. And because much of that work involves monitoring, verification, coordination, and repeated decisions, consumption may become more continuous.
VI. Specialized models could make the market bigger, not smaller
Another apparent contradiction deserves attention. The largest frontier models attract enormous investment because they can perform increasingly diverse and difficult tasks. At the same time, smaller models have become substantially more capable. These trends aren't necessarily competing visions of the future. They may reinforce each other. A large model might plan a complex objective or resolve an unusual problem. Smaller specialized models might handle routine classification, extraction, monitoring, or domain-specific operations at lower cost.
A coordinated system could use different models according to the complexity of each task. That doesn't mean every company needs to train a proprietary model. Many will use existing models, fine-tuned systems, conventional software, or services supplied by specialized vendors. Nor does specialization guarantee superior performance. The opportunity depends on whether a particular model can deliver sufficient accuracy and reliability at an attractive total cost. But consider the economic implication.
If smaller models make millions of narrowly defined tasks affordable, adoption can expand well beyond the applications that justify expensive frontier-model inference. The number of potential applications grows as the price of performing them falls. And increasingly sophisticated coordination systems can combine those capabilities into larger workflows. There is no contradiction in believing that individual AI tasks will become dramatically cheaper while aggregate demand for AI computation continues increasing. The question is which effect dominates.
VII. The physical world eventually sends the invoice
None of this happens in a vacuum. Computational work requires physical resources. Servers need processors, memory, storage, networking, cooling, and electricity. Data centres require buildings, transformers, switchgear, power-distribution equipment, and access to electrical grids. Manufacturers require production capacity, materials, skilled labour, and financing. Even a largely digital economic transformation eventually encounters physical constraints.
The International Energy Agency estimated that data centres consumed approximately 415 terawatt-hours of electricity in 2024, representing about 1.5% of worldwide electricity consumption.
Its 2025 Energy and AI report projected approximately 945 terawatt-hours by 2030 in its base case.
That forecast includes growth from AI and other digital services. It is not a direct forecast of autonomous agent swarms.
Source: International Energy Agency, Energy and AI, 2025.
This creates an interesting economic situation. A business may have the financing, customers, and equipment required to expand computational capacity. But if it cannot secure electricity, the capacity cannot operate as planned. The constraint is no longer the sophistication of the AI model. It is something considerably less glamorous. Physical infrastructure. And that raises a question central to Sorites Research:
What happens to the economic value of an ordinary industrial business when its existing products become essential to an extraordinary technological expansion?
The answer depends on competition, manufacturing capacity, customer relationships, pricing power, and valuation. But it is an important question. The company manufacturing the intelligence may not capture all the economic value created by demand for intelligence. Some of that value can migrate toward businesses removing the constraints that prevent the industry from expanding. And those businesses may have existed for decades.
VIII. The investment opportunity is not necessarily where everyone is looking
The first approach asks which company will build the most capable model. The second asks what every successful model provider and enterprise customer will need in order to operate. Those questions can lead to different businesses. Processors attract attention, but a working data centre also depends on power distribution, cooling, connectors, enclosures, networking, construction and maintenance. A shortage in one modest component can delay a much more expensive installation.
This is the hidden bill of materials: the industrial capabilities beneath a digital service. The research opportunity is to trace demand through that bill, then establish where supply is difficult to expand. Qualification requirements, installed relationships and manufacturing lead times may matter more than whether a supplier markets itself as an AI company.
A bottleneck can confer pricing power. It can also invite competitors, substitution and excess investment. A company can increase revenue while its margins deteriorate. A scarce product can become abundant before an investor's purchase pays off.
Being correct about the future of an industry is not the same as being correct about the value of its securities.
This thesis therefore makes no stock recommendation. It supplies a question for subsequent research: where does additional useful computational work encounter a constraint, and who can remove that constraint at an attractive economic return? Answering it requires company accounts, customer evidence, competitive analysis and a valuation. The structural story alone is insufficient.
IX. The strongest argument against our thesis
There is a credible alternative future. Models improve so rapidly that substantially less computation is required for each completed objective. A capable single model replaces elaborate teams. Better software removes redundant calls and repeated reasoning. Conventional automation handles predictable steps without invoking a language model at all.
Organizations may discover that the fully loaded cost of autonomous work is much higher than the advertised inference price. Integration, monitoring, permissions, verification, failed actions and human intervention all belong in that calculation. A cheap answer is not a cheap business outcome if someone must check every step.
Reliability also compounds across a workflow. A system that performs individual steps well can still fail an extended assignment. Security controls can restrict access to the very data and tools that would make autonomy useful. Liability, regulation and accountability can keep humans in the loop even when the technical demonstration looks impressive.
Gartner's June 25, 2025 forecast said over 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. This is a forecast, not a measured cancellation rate. Source: Gartner, June 25, 2025.
McKinsey's November 2025 survey also described a gap between widespread experimentation and enterprise earnings impact. Adoption counts cannot settle the return-on-investment question. Nor can a provider's usage growth establish that customers will keep paying once subsidies, trial budgets or novelty fade.
Even durable demand growth can disappoint infrastructure investors. Capacity can arrive ahead of customers. Utilization can fall. Competition can transfer efficiency gains to buyers. Depreciation, financing costs and rapid equipment obsolescence can absorb operating profits. Electrical equipment suppliers may expand production just as shortages ease.
There is also a valuation problem. Much of the anticipated expansion may already be reflected in capital spending plans and securities prices. An investor can identify the direction correctly and still pay too much for exposure to it.
The efficiency rebound is therefore a conditional hypothesis. We have to observe whether the growth in useful work outweighs the reduction in resources required per task. We cannot infer that result merely from lower prices, more agents or a larger demonstration.
The strongest bearish argument is not that AI will fail. It's that AI may succeed without generating the economic returns investors currently expect.
X. What would change our minds?
A foundational research thesis should be capable of being wrong. We would weaken our view of the Agent Multiplication Effect if evidence increasingly showed that:
- Enterprise agents deliver insufficient returns to justify broad deployment.
- Agentic systems remain heavily concentrated in a narrow set of applications rather than expanding across industries.
- Improving single-model capabilities eliminate much of the expected growth in multi-agent computational work.
- Total inference consumption grows more slowly than efficiency improvements reduce computational requirements.
- Infrastructure utilization, investment returns, and customer spending consistently disappoint despite increasing model capabilities.
These are review conditions, not numerical forecasts with invented precision. We would assess them across multiple reporting periods and distinguish temporary deployment friction from persistent economic failure. The thesis would strengthen with independently measured business outcomes, repeat customer spending, growth in automated task execution and economically justified infrastructure expansion. We would watch whether lower inference costs make additional categories of work viable, rather than merely making existing usage cheaper.
API consumption and token volumes are useful signals, but they need context: model mix, caching, task completion, error rates and resources per successful outcome. Hardware utilization and electricity demand are different measures again. No single indicator proves the whole chain from adoption to computation to investor returns. Future updates will record the evidence, what changed and whether our confidence increased or decreased. The original publication will remain available alongside those dated assessments.
XI. The Question We're Leaving Open
We are not predicting that every business will run a swarm of thousands of agents. We are asking whether the amount of economically useful work performed by machines can grow beyond the activity initiated directly by people.
If it can, counting chatbot users will miss a consequential part of the market. A person assigns an objective; machines research, compare, verify and coordinate; other businesses' systems respond. The human decision is the beginning of a chain of work, not its measure.
The size of that chain depends on reliability, cost and value. Some architectures will collapse into simpler systems. Some applications will fail. Others may become economical precisely because intelligence becomes cheaper. The balance between those effects is the question we are leaving open.
Our next investigations will follow the physical consequences: which constraints matter, how quickly supply can respond, and who already possesses the capacity to remove them. They may reveal attractive businesses, expensive securities or reasons to abandon our original expectations. The most important changes rarely arrive fully formed. They accumulate: one technical improvement, one customer deployment, one new application, one infrastructure constraint. Eventually, the accumulation becomes impossible to ignore.
The Future Has a Past.