In 2026, Uber spent its full-year budget for AI coding tools in four months.
Fortune reported it on May 26. Andrew Macdonald, Uber's president and COO, talked about AI usage on a podcast. He mentioned the kinds of AI statistics companies like to cite, such as the share of code commits driven by AI, or how much usage has grown. Connecting numbers like these to how many more useful consumer features actually get built, he said, is not easy.
"That link is not there yet."
Bloomberg then reported that Uber had set a cap for each AI coding tool: $1,500 per employee per month.
Uber is not alone. McKinsey's 2026 global survey asked the same respondents two things. One: has AI improved your own productivity? Two: has AI increased your company's profit? 80% said their productivity had improved. Only 37% said company profit had increased, about the same as a year earlier. Profit here means EBIT.
"These individual gains have yet to translate into broad financial impact for organizations."
Individuals got faster. The company's profit did not move with them. What happened in between?
Productivity does not turn into profit by itself
Two researchers at MIT published an experiment in 2023. They recruited 453 professionals: marketers, HR staff, data analysts, consultants, and managers. The tasks were things they already write at work: press releases, short reports, analysis plans, and difficult emails. Half were randomly chosen and given access to ChatGPT.
The group with access took 40% less time, and their output was rated 18% higher.
This is what an AI investment buys: knowledge work such as reading, writing, organizing, and judging takes less time per task. In the same time, people can do more, and do it better. That is productivity.
But productivity is not profit. Profit is what is left on the income statement after costs and expenses are taken out of revenue. For productivity to turn into profit, someone has to decide what the freed-up time is used for.
Productivity turns into profit along four paths
Extra productivity turns into profit along roughly four paths. On each one, the profit comes from a different place.
Path 1: Spend less
On a January 2026 earnings call, Bank of America CFO Alastair Borthwick described what the bank does now. Every time someone leaves, it evaluates whether the role needs to be replaced.
Shopify is going the same way. An internal memo from CEO Tobi Lütke became public in April 2025. It requires teams, before asking for more headcount, to show why the work cannot be done with AI.
Both companies are doing the same thing: spending less. This is the most direct path from productivity to profit. Hire one fewer person, and one salary is not paid that same period.
Payroll is not the only thing a company can spend less on. There are also things it used to buy and can now build itself. In McKinsey's 2026 survey, 32% of respondents said their company decided not to buy a software product or feature because it could build it with AI coding tools. That fee, too, goes unpaid that same period.
On this path, the company pays out less. That is savings. The condition is that the company really pays less. If roles are still filled and software is still bought, spending has not gone down.
Path 2: The same people do more
A study published in 2025 followed more than five thousand customer support agents at one company. An AI assistant was rolled out in stages. Agents who got it resolved 15% more issues per hour on average. The headcount did not grow. The amount of finished work did.
The study measured issues resolved, not how much the company's profit grew. For the extra work the same people can do to become profit, there must be somewhere to use it.
Somewhere to use it means more customers to serve and more orders to take. Companies short of workers are in exactly this position: the work is there, and people are what is missing. The Market Intelligence & Consulting Institute surveyed Taiwan's electronics and IT manufacturers (in Chinese). The companies said the clearest improvements after adopting AI included higher revenue, less pressure from the labor shortage, and lower costs.
The Liberty Times reported in September 2026 (in Chinese) on how the Commerce Development Research Institute read the business indicators. The institute said companies facing a labor shortage are speeding up their adoption of automation and AI equipment, substituting machines for people to raise efficiency.
In that situation, the extra work becomes revenue. Customers pay in more. That is revenue.
When there is nowhere to use it, the extra productivity sits idle, and revenue does not grow.
Path 3: Do it better and faster
Researchers from Harvard Business School and other universities published an experiment in 2023 with 758 consultants at a large consulting firm. On tasks AI could handle, the group using AI worked a quarter faster, and the quality of their output was rated more than 40% higher.
AI can make the work better and faster. The next question is whether the customer cares about that "better".
The customer cares, and is willing to pay more. This is the best case. Lead times shorten and quality rises, so the customer orders more or accepts a higher price. That is revenue.
The customer cares, but will not pay more. This case is more common. Better quality becomes the standard customers use to choose suppliers. Revenue does not grow, but the business is kept. Competitors who cannot match it drop out, and new ones find it harder to get in. That is retention: keeping business that would otherwise have been lost.
The third case is that the customer does not care. The work is better than the customer needs, and the extra part makes no difference. Project managers call this gold plating.
AI makes "a little better" cost almost nothing, so gold plating happens more easily.
Whatever the customer's reaction, the company also saves on its own side. Fewer errors mean lower costs for rework, returns, and complaints. That is savings. Less back-and-forth inside the company frees up time. That freed-up time still has to go down Path 1 or Path 2 to become profit.
Path 4: Do what was not worth doing before
Some things are useful but were not worth the cost, so they did not get done. Once AI lowers the cost of doing a task, they can be done.
Of the four paths, this one starts first, and the company does not have to arrange it. When employees get AI, the first things they do are often the ones that used to be too much trouble to be worth the time.
There are two kinds of such things. In both, the profit arrives by a roundabout way.
The first kind is comparisons that people wanted to make but had no time for. A web page used to get one layout, because that was all the time allowed. Now five can be made at once, opened side by side, and one chosen. In McKinsey's 2026 survey, half of the respondents said AI helps them make better decisions.
Better decisions show up first in hit rates. Marketing looks at conversion rate. Sales looks at close rate and bid win rate. Pricing looks at how far the actual margin on a won bid is from the estimate. The company already measures these numbers. A higher hit rate brings revenue.
The second kind is work that should be done but was not worth doing before. It could not get staffed. Now it is affordable: reading three years of customer complaint records from start to finish to find causes, or adding tests and documentation to every code change. The error rate comes down, and rework falls with it. That is savings, but it takes a while.
Both kinds end up on the income statement, as revenue and gross margin that slowly improve. But there are several steps in between, it takes time, and no line on the income statement will say "AI".
So on this path, a company cannot wait for the numbers on the income statement. It has to look at the steps in between first: is the error rate falling, is the hit rate rising? If neither is, the company is only doing more things because they are cheap.
Four paths, three places profit comes from
The profit from all four paths comes from only three places. What the company no longer pays out is savings. What customers newly pay in is revenue. Business kept that would otherwise have been lost is retention.
| Where it comes from | How soon it shows | |
|---|---|---|
| Savings | Hiring fewer people, buying less, fewer errors | Hiring fewer people and buying less mean paying less that same period. Fewer errors has to wait until the error rate comes down |
| Revenue | Taking more orders, customers willing to pay more, higher hit rates | More orders become revenue in the period they ship. Higher prices and hit rates are slower: lead times, renewals, and hit rates move first, then revenue and gross margin |
| Retention | Customers care but do not pay more, and the business is kept | The income statement shows nothing gained. It has to be compared with the case of not investing |
Not many small and mid-sized companies in Taiwan know how AI productivity turns into profit. Taiwan's Ministry of Economic Affairs surveyed more than a thousand of them in 2025 (in Chinese) (10 to 199 employees, excluding wholesale and retail, accommodation and food services, finance, and ICT). In the questionnaire, the highest level of AI awareness was a company that fully understands AI and knows how it can bring the company revenue growth, cost savings, or both.
About 4% of the companies were at that level.
How to look at the investment
Evaluating an investment means asking four things: what it costs, compared with what, how you know it is working, and how long to wait. Applied to AI, each one has a spot that is easy to misread.
What it costs
Manufacturers split a product's cost into two periods: what is spent once during development, and what keeps being spent after mass production starts. An AI adoption can be split the same way.
What is spent once, during adoption, is the investment before AI is in real use. The hours of the people involved count, and that means everyone involved, not only the IT department. Then there is the time spent on trials, the effort of building the workflows, and the fees for training and consultants.
Of what keeps being spent after AI is in use, the largest item is the tools. The other is administration: someone has to manage accounts, security, and usage rules.
How the tool cost is counted depends on how the company gets its AI.
| Subscription | Pay per use | Self-hosted | |
|---|---|---|---|
| How you pay | Per seat, a fixed amount each month | For what you use | Hardware up front, then maintenance and power |
| When nobody uses it | You still pay | You pay nothing | The equipment still depreciates |
| What to watch | Seats that are paid for and unused | Heavy use means a high bill, and it is hard to estimate in advance | A large sum at the start, and someone has to maintain it |
Subscriptions usually have a usage limit. Past it, the company buys more or switches to paying per use. Pay per use has no limit of its own; the company has to set one. At high volumes, the same usage usually costs more on pay-per-use than on a subscription.
A pay-per-use bill is also hard to estimate in advance. Uber, the company in the opening, used up a year's budget in four months and then set a monthly cap per employee for each tool.
Pricing models change, too. GitHub's coding assistant, Copilot, moved to usage-based billing on June 1, 2026. The monthly fee stays the same, but the allowance it includes is now drawn down by actual usage. When it runs out, the user pays extra or stops using the tool. In McKinsey's 2026 survey, about 20% of respondents said AI operating costs were already limiting their company's use of AI.
What is spent during adoption has to be earned back from the extra profit each month afterward. Fewer users and fewer uses stretch that payback out.
Compared with what
Evaluating an investment means comparing two things: what happens if the company invests, and what happens if it does not.
The second one is where mistakes are made. What happens without the investment is hard to estimate, so it is often treated as "staying where we are": this year's revenue, this year's customers.
That only holds if competitors do not invest either. Once competitors invest, and really turn it into an advantage in cost or lead time, three things usually happen.
First, the competitors get stronger. Their costs come down, or their lead times get shorter and their quality better. They use this to win business. EY published a survey of senior leaders at US companies in late 2025. Among companies reporting AI-driven productivity gains, 29% were using the gains to reduce prices and gain market share (respondents could choose more than one answer).
Next, the gains reach the customer. On JPMorgan's earnings call on July 14, 2026, CEO Jamie Dimon, answering an analyst's question, talked about whether a company gets to keep the benefits of AI.
"…you don't uniquely benefit from AI. The ultimate beneficiary of AI will be our customers."
He went on to say that a bank cannot simply claim that AI will raise its margins and that it will keep all of the gain:
"If that were true, our margins would be 80% today because of computerization over the last 20 years."
Then customers take all this as the new standard and go back to every supplier asking for it: the same price, lead time, and quality as the best one. On price, manufacturers know this well. Supply chains have a convention called the annual price-down (every year the customer asks the supplier to lower the price a little more, whether or not the supplier's costs have fallen). Once competitors' costs have come down, the customer asks with more confidence. Suppliers that can lower their costs use the savings to cover the price cut. For those that cannot, gross margin gets thinner every year, and the customer may switch suppliers.
So not investing does not mean staying where you are. It means going down. Competitors bring the price down and the standard up, and the company has not kept up.
① vs. today: only this part② vs. not investing: the whole gain
A company that invests, and really turns it into cost or lead time, at least gets retention. Look at retention alone: for the same product, the price is lower every year, so that product's revenue is actually going down, and the income statement shows nothing gained. What is kept is the customer. Only while the customer is still there is there a chance to sell new products, adjust the product mix, and build revenue back.
Back to the comparison at the start. What happens with the investment: look at the four paths. Some are savings, some are revenue, some are only retention. What happens without it: the result of the three things above. What this investment buys has to be counted from the not-investing side, not from where the company is today.
How you know it is working
When evaluating results, the numbers at hand are of two kinds: process numbers and outcome numbers. Process numbers are the easiest to get, but results have to be read from outcomes.
Process numbers: easy to get, but not results
There are two common ones: usage, and individual speed.
Usage: how many people use it, and how much
Usage is the easiest number to get. The Uber COO from the opening was citing this kind of number: the share of code commits driven by AI, how much usage has grown. His point was that these numbers do not connect to how many more useful consumer features the company actually built.
Heavy use does not mean something got built. Nor does it mean that what got built is something customers will pay for.
Gartner published a survey of finance leaders in May 2026. Speaking about AI in finance functions, analyst Marco Steecker said activity must not be mistaken for impact. Counts of pilots, tools rolled out, or use cases in production show that the function is moving. They do not prove that AI is delivering the value boards now expect.
Individual speed: how much faster each person is
An O'Reilly Radar article on software development from July 2026 describes a situation. Every engineer on the team sincerely believes they are 50% more productive, and the ship dates have not moved.
The reason is the bottleneck: the slowest step in the whole process. A factory line works the same way. The output of the line is set by the slowest station. Speed up the other stations, and work in progress only piles up in front of the slowest one.
How much faster each person is does not mean the product ships sooner.
Outcome numbers: the operating numbers the company already measures
Outcome numbers are not new numbers. Expenses, shipment volume, lead time, return rate: the company already measures them. Which ones each of the four paths moves can be matched up.
- Spend less: personnel and purchasing costs go down
- The same people do more: with headcount unchanged, more orders taken, more shipped, more customers served
- Do it better and faster: lead times shorten, returns and complaints fall, customers renew and order more
- Do what was not worth doing before: the error rate falls, and hit rates (conversion rate, close rate, bid win rate) rise
On the first two paths, when the numbers move, profit soon moves with them. The last two are further from the income statement: the numbers move first, and the profit follows after a while.
Write down what these numbers are before adoption. Only then is there something to compare with.
How long to wait
Gartner surveyed 160 senior finance leaders from January through April 2026 and published the results in September. The simpler use cases in finance generally deliver returns within nine to ten months. Examples are data extraction, accounts payable and receivable automation, and report creation. More complex ones, such as data management, insight generation, and forecasting, take longer.
Marco Steecker, the analyst mentioned earlier, gave an interview to CFO Dive. He said that if the value has not appeared by that point, and the reasons have been examined, the investment has probably been placed in something that will not deliver.
In the press release, he also warned about the other side. AI investments in finance tend to concentrate on use cases that return quickly and raise productivity. CFOs should not let the appeal of quick returns crowd out the more complex use cases. Those mature more slowly, but they can improve decisions, manage risk, and support revenue growth.
This lines up with the four paths. What returns quickly is mostly Path 1, spending less: fewer hires and fewer purchases mean paying less that same period. What matures slowly is more like Path 4, doing what was not worth doing before. The profit arrives by a roundabout way: hit rates and error rates show first, and revenue and gross margin only after a while.
One more thing affects how long to wait: how much was spent during adoption. A few dozen subscription seats and a few training sessions do not cost much up front. Changing workflows, connecting systems, and self-hosting are a large sum from the start. The more is spent, the longer the payback.
Questions to ask before adopting
The same 2025 Ministry of Economic Affairs survey (in Chinese) asked small and mid-sized companies what challenges they met in adopting AI. One in five chose "the results of adoption are hard to evaluate".
In April 2026, the Artificial Intelligence Foundation and Qualcomm released the 2026 Taiwan Industry AI Survey (in Chinese). The foundation's CEO, Greta Wen, also addressed this in the press release. After adoption, she said, companies also need to consider how to evaluate and accept the results, including when to confirm that an experiment has failed and stop, and when to continue.
The points above can be gathered into a few questions to ask before adopting.
- What do we plan to do with the extra productivity: spend less, do more, or do better?
- The same people can now do more. Are there more orders to take and more customers to serve?
- If we do it better, does the customer care? Will they pay more?
- Have we counted all the costs: what is spent once during adoption, and what is spent every month after?
- If competitors invest and we do not, can we keep our current customers and business?
- What will we use to judge whether it is working: usage and individual speed, or outcome numbers? What are those numbers today, and have we written them down?
- How long are we willing to wait?
Further reading: Private LLM vs API: 3 Ways to Deploy AI in Your Company — Cost, Security & Trade-offs Compared