Imagine a company that can do the same work with fewer person-hours. The result looks good in the presentation: faster delivery, lower costs, higher capacity. Then the biggest customer calls. Your rival has started using the same tool and cut its price. The customer now expects a discount from you too.
The company really is more productive. But it is not yet clear how much of that gain will stay with the company. As I assess artificial intelligence investments over the next three to five years, this is where the more interesting question begins for me: when production gets cheaper, what will customers keep paying for?
Who keeps the productivity gain?
My central thesis is this: turning the savings created by AI into profit depends on how quickly rivals catch up, how strongly customers demand lower prices and which assets the company has that cannot be easily substituted. Using a good model can create an advantage at first. Once everyone has access to the same capability, the source of the advantage changes.
Reaching customers, solving a particular problem reliably, having a body of data that the company is allowed to use, or taking responsibility for the service may become more valuable. Each of these claims still has to be tested. A company having a lot of data does not mean that the data is useful or unavailable to others. A strong brand does not show that customers will pay a price premium under every condition.
That is why I am not ranking the winners in advance as chipmakers, model developers or application companies. The same question applies at every layer: What is it that the customer cannot easily get elsewhere, and how much will the customer pay for it?
The strongest counterargument: Total gains can grow as prices fall
The strongest objection to this thesis is that a lower cost can create new demand. A service that was not economical to sell to a small business may be offered at a lower price with AI. Existing customers may use more of it. Faster delivery or better service may also make it possible to keep the price unchanged.
Consider a hypothetical company that can provide specialist support only to large customers. If some parts of the service become cheaper, it may also reach smaller customers. Contribution per unit may fall, but total contribution can rise if the number of customers grows enough. The critical conditions are that the new customers actually arrive and that the additional cost of serving them does not overturn the calculation. Potential market size is not realized sales.
Moving early can also have value. As access to models spreads quickly, changing how the work is organized can take longer. Customer relationships, experience and product knowledge gained during that period may carry into the next one. But a temporary first-mover advantage should not be treated as pricing power that can last for years. The first opens a window of opportunity; the second needs a reason to survive after the window closes.
The gap between saving time and making money
Research shows that even the first step varies by task. The customer-service study published by Brynjolfsson, Li and Raymond in 2025 reports an average increase of about 15% in issues resolved per hour when 5,172 workers had access to AI. The gains were concentrated among less experienced workers. Customer sentiment and escalation outcomes also improved; the study does not show that customers would pay a higher price. This is a result about operational productivity at one company, not a forecast of long-term profit or the sector as a whole. The study.
In METR's small experiment using tools from early 2025, 16 experienced developers working on their own open-source projects took 19% longer on average when they had access to AI across 246 tasks. The result is limited to particular tasks and the tools available at that time. It cannot be carried over to all software work today. The METR study.
In its February 2026 follow-up note, METR said its follow-up study using newer tools could not produce a reliable estimate because participant and task selection had changed. We therefore cannot draw a straight line from the older result to today. The 2026 update.
We cannot subtract these two percentages and produce an AI average. The workers, tasks, systems and measured outcomes differ. My lesson from this comparison is that the investment calculation should start with the company's own work. A speed increase calculated without checking, rework, explanation to the customer and error-correction time can easily mislead.
The way the saved time is used also changes the result. If the same team can complete more orders, the additional capacity can be sold. If there are no orders, an empty hour does not create revenue by itself. If people have to be moved to another job, there is a training and transition cost. Whether labor spending actually falls depends on working hours, contracts and organizational decisions. The difference between opening a gap in the calendar and reducing cash outflow should appear in the budget.
When savings become a price cut
Let me make the mechanism concrete with a simple example that does not use real company data. A service sells for 100 units and has a variable cost of 80 units. That leaves 20 units of contribution per sale. This is the contribution before fixed costs, financing and taxes; it is not net profit.
AI reduces the existing cost of the work by 10 units. Model use, checking and error correction add 5 units. The new variable cost is 75. If the selling price does not change, contribution rises to 25. If the price falls to 90 after a move by rivals, contribution falls to 15.
| Assumption for the same service | Price | Variable cost | Unit contribution |
|---|---|---|---|
| Before AI | 100 | 80 | 20 |
| If the company keeps the savings | 100 | 75 | 25 |
| If price competition begins | 90 | 75 | 15 |
In the last row, the company does the work more cheaply but earns less per sale. To recover the previous total contribution, volume must rise by roughly one-third: 20 / 15 = 1.33. This calculation assumes that the unit cost does not change with higher volume and that additional fixed-capacity investment is not required. If more sales require a new team or facility, the threshold rises.
For the customer, a lower price is a gain in its own right. A development that limits the company's increase in profit is not necessarily a failure for the economy as a whole. The same technology may bring consumers cheaper services, workers higher pay or suppliers more sales. The creation of value and the sharing of value are separate questions.
Where does bargaining power go when the model gets cheaper?
Stanford's 2025 AI Index summary reports that the price of a million-token query for models reaching a 64.8% MMLU performance level fell from $20 in November 2022 to $0.07 in October 2024. This is the price at a particular performance threshold; it does not apply to the total cost of implementation or to every task performed by the most advanced models. Stanford HAI.
If this kind of deflation continues, it can open new territory for companies developing applications. But the same opportunity opens for rivals. That is why, rather than tying the budget only to the model's current capability, I find it more useful to ask why customers would choose us when that capability is cheap and widely available.
A January 2025 staff report from the U.S. Federal Trade Commission discusses possible competition concerns in the partnerships between large cloud providers and AI developers that it examined, including switching costs and access to critical inputs. These are risks identified in the report, not a settled legal finding or a ruling on who will receive the profit. FTC report announcement.
A fall in the unit price of a service is not the same as an easier exit from that service. Moving data, rebuilding an evaluation system, training employees and maintaining uninterrupted service can be costly. When I choose a supplier, I include those costs alongside the price. A portable system can sometimes provide more bargaining power than today's lowest invoice.
Savings, pass-through, persistence
I suggest that a management team test an investment decision with three questions. The first concerns savings: has the total cost of completing a job at the same quality standard really fallen? Without a baseline, representative tasks and the cost of errors, the answer is often only an impression. Choosing the easiest tasks for a successful pilot can also make the scaling calculation look better than it is.
The second question is pass-through: which party receives the gain? We need to look at realized selling prices, discounts, usage volume, customer losses and the supplier invoice together. If we measure price from the list price while leaving out extra service given to the customer for free, we can miss the transfer. Similarly, higher usage is not the same as higher revenue; if the package price is fixed, more usage may raise costs.
The third question is persistence: why will the customer stay when a rival can access the same model? The answer may be a particular integration, reliable results, permitted and useful data, or a strong distribution relationship. If the answer is “we use AI too,” the investment's defense looks weak. We need to ask separately how quickly each advantage can be copied and how much the customer values it.
This assessment can be read as a continuation of [the article on managing AI workflows](/en/writing/ai-strategy-operating-layer). There, the focus was the company's internal structure of decisions and responsibility. Here, I am looking at how much value that structure's output carries when it meets a customer and a rival. A well-functioning process may be necessary for commercial advantage; it may not be sufficient on its own.
What would I invest in over the next three to five years?
I use this period to question the durability of an investment, not to make a technology forecast. At the first stage, I would measure the net cost and quality with a small but representative use case. I would set the success criterion around the cost per accepted result and the customer's experience, rather than producing more output. Making usage volume the only target could increase the risk of [KPIs distorting behavior](/en/writing/when-kpis-turn-into-midas).
I would tie the next allocation of resources to evidence from the customer side. Does faster service raise the renewal rate? Does the new price actually bring in new customers? How much discount pressure does the sales team feel? A scaling plan that cannot answer these questions turns technical success into an assumption of commercial success. The comparison also needs to separate the product mix, overall demand and other price changes.
Over the longer term, I would put money into two things: capabilities that customers value and rivals find hard to copy, and preserving the option to change suppliers. The second does not mean building everything in-house. Making critical data exportable, testing an alternative service at small scale and knowing the transition cost can all widen the decision space.
The conditions for scaling, correcting or stopping an investment can be written in advance. For example, if unit cost falls while realized price falls faster and quality is maintained, I would reconsider differentiation or the target customer group rather than put more resources into the same product. If savings are weak but the price customers pay and their loyalty increase, the investment case built around service quality becomes stronger.
What would change my mind?
My expectation is that ordinary cost advantages will be more exposed to competition in areas where access to similar models becomes widespread. Evidence that would weaken this view would be for companies without a clear data, distribution or customer-retention advantage to keep higher contribution for years after rivals catch up. If new demand and better quality consistently outweigh price pressure, I would put less weight on the need for a separate scarce asset to support lasting gains.
At the same time, I do not assume that all the gain will stay with a few infrastructure providers. The spread of cheaper alternatives and easier switching can move bargaining power to customers or application companies. Whatever result emerges, we cannot read it from usage volume alone; we need to track price, quality, cost and customer behavior together.
The question I ultimately want to ask about an AI investment is this: when our rivals can do the same work at the same speed, what difference will the customer pay us for? A company with a concrete answer has a stronger reason to believe it can turn today's savings into tomorrow's gain.
Sources
- Brynjolfsson, Li, and Raymond, Generative AI at Work, Quarterly Journal of Economics, 2025.
- METR, Early-2025 developer productivity experiment, July 10, 2025.
- METR, Experiment design and follow-up update, February 24, 2026.
- Stanford HAI, 2025 AI Index summary, April 7, 2025.
- FTC, AI partnerships staff report announcement, January 17, 2025.
