The Pilot Is Done. Where Is the Profit? Measuring AI on the P&L, Not on Productivity

The year-end deck says “productivity is up”, and not a single line on the P&L has moved. Hours saved aren't profit until they turn into headcount or work you sell. How to sort your pilots with that in mind.

Alp SuleimanGrowth PartnerArticle 02 / 0210 min read

Published 14 October 2026

December, budget meeting (example). The operations director opens the third page of the deck: the AI pilot that matches supplier invoices to purchase orders has saved 400 hours a month. The slide is green. The CFO looks at the income statement. The accounting team is still 12 people. Overtime is the same as last year. Revenue hasn't moved beyond price increases. “Where did those 400 hours go?” the CFO asks. Nobody answers, because nobody planned to ask that question.

This piece is about that silence. Both sides can be right. The hours may really have been saved, and profit may really not have changed. Both can be true at once, because saved hours are not a financial result on their own. Hours saved aren't profit until they turn into a change in the headcount plan or into output you sell.

Earlier on this site we looked at the real reading of “AI sped us up”: perceived speed and measured speed can diverge. This piece starts one step further on. Say the speedup is real and correctly measured. It may still never reach the income statement.

The productivity measure can stand in for the profit measure

The Türkiye arm of KPMG Türkiye's study “Yapay Zekâ Dönüşümü: Türkiye ve ABD Perspektifleri” (AI Transformation: Türkiye and US Perspectives; published in Turkish in November 2025) was carried out with 31 senior executives in the second quarter of 2025. The survey paints a picture in which the return on AI is tracked through productivity more than through profitability.

Thirty-one people don't represent Türkiye; this is the answer of one group of executives. But the situation it describes is familiar. The measure you use to track the return decides which question you ask. A company that measures productivity asks “how many hours did we save?” and finds the answer. The question “which line changed?” often never makes it into the meeting.

A study from Denmark shows a related picture. Anders Humlum and Emilie Vestergaard matched survey answers from around 25,000 workers in 11 occupations and 7,000 workplaces with the country's official records of earnings and hours worked. People who use chatbots say they save time, but the time saved is small. There is no meaningful effect on earnings or on recorded hours.

A distinction needs protecting here. This study measures the worker's pay and hours. Company profit can rise without pay rising. Still, the picture points the same way: time was saved, and measured earnings and recorded hours didn't change. The authors partly explain this by productivity gains passing through weakly to wages. So the gain may not have reached the worker but may have stayed in the company; the study doesn't measure company profit.

The strongest counterargument

The most serious argument against this thesis is about time.

Erik Brynjolfsson, Daniel Rock and Chad Syverson's “productivity J-curve” work argues that measured productivity follows a misleading path with general purpose technologies. The argument goes like this: adopters of a new technology first make invisible complementary investments. Processes get redrawn, teams get trained, data gets cleaned. Most of this is booked as an expense today and isn't measured as an intangible asset. So measured productivity first looks lower than it really is, and later higher. The paper's finding is at the level of sectors and the economy. If we carry the same logic over to a single company's profit (interpretation): “no profit today” and “there will be no profit” are not the same sentence.

The J-curve may be right. But it gives you a right to wait, not a blank cheque. For the curve to bend upwards, the complementary investment has to have a name: which process will change, which headcount plan gets rewritten, how many months it will take. Waiting without a name is just waiting.

Three places saved hours can go

A saved hour goes to one of three places. If nobody decides, it goes to the third one by default (rule of thumb).

The calculations below are examples; put in your own numbers. Shared assumption (example): one FTE, a unit of headcount equal to one full-time employee, works about 160 hours a month and costs the employer 90,000 TL a month in total.

1. Headcount: fewer hires, less overtime

Go back to the accounting team from the opening (example). The company is growing, and next year's plan includes two new accounting specialists. 400 hours is roughly 2.5 FTE. If the person running the pilot and the person signing the budget sit down and rewrite the plan, the two hires don't happen. A cost of 180,000 TL a month, 2.16 million TL a year, never appears in operating expenses. Operating expenses are the cost of running work not directly tied to the product sold: administrative staff, rent, software.

Nobody has to be let go. Growing headcount more slowly while you grow is enough. This is a small version of operational leverage: revenue grows while costs don't grow at the same rate.

Two things to watch. The saving only exists if the hiring plan actually changes; if the plan stays the same and the two people are hired anyway, the 400 hours go to the third place. And don't read the TL saving in nominal terms. As salaries rise over the year, the “cost that never appeared” also grows in nominal terms, and next year's deck may show it as “savings up 40 percent” (example). Most of that increase may be price effect. Make the comparison at the same period's prices, in real terms. If you pay for the tool in foreign currency, the same rule applies on the cost side.

2. Output you sell: the same team sells more work

The second place is the revenue line (example). A six-person proposal team prepares 60 proposals a month. Ten hours per proposal, 600 hours in total. AI brings that down to 6 hours per proposal. The same 600 hours could produce 100 proposals.

Say the win rate is 20 percent, the average order is 250,000 TL and gross margin is 30 percent (example). The extra 40 proposals mean 8 extra orders. That's 2 million TL in extra revenue a month and 600,000 TL in extra gross profit. Gross profit is what's left of revenue after cost of goods sold (COGS: the direct cost of producing the product or service).

On paper the calculation looks good. But it depends on one condition: the market has to want those extra 40 proposals. If you already respond to every request that comes in, preparing proposals isn't the bottleneck. A faster team won't produce more proposals; it will finish the same 60 sooner, and the remaining hours go to the third place again. So first look at which input actually drives the result; that's what the Metric Tree is for. Check in your unit economics whether the extra orders actually leave a profit, too. Eight orders won on discount can pull gross margin below 30 percent.

The measure to track on this path is revenue per employee. In subscription businesses its common form is ARR per employee: annual recurring revenue divided by the number of full-time employees. If the ratio rises without headcount growing, hours are turning into work you sell.

3. Nowhere: the hours dissolve

The third place has no owner, because nobody decides to go there. The 400 saved hours (example) spread across the weeks. Meetings run a little longer. Reports get one more round of review. Month-end close still finishes on the same day, just more calmly.

For the team this can be a real improvement; a less exhausted team is nothing to sneer at. For the income statement it is zero: same personnel cost, same revenue, same profit.

Rule of thumb: if nobody is moving the hours into headcount or sales, this is where they go. There's no bad intent, only a missing decision. In the Danish data, time was saved and the records didn't change. But that data doesn't separate hours that dissolved from hours that stayed in the company; separating them is up to your own measurement.

The test question

Ask every AI initiative the same question, at the start and in every budget cycle:

If this pilot succeeds, which line of the income statement changes, within how many months, and on whose signature?

Each part of the question filters on its own. “Which line” forces the answer into three options: revenue, COGS or operating expenses. “Productivity” isn't a line. “Within how many months” puts a limit on the J-curve; 18 months can be reasonable too, as long as it's written down.

“On whose signature” is, I think, the part that filters out the most (interpretation). The person running the pilot is often not the person who signs the hiring plan or the sales target. Moving the hours depends on the budget owner saying “I'm taking these two hires out of the plan” or “I'm raising this team's target from 60 proposals to 100”. If there's nobody to say that sentence, the result won't change however well the pilot works.

The first of the ten growth questions for founders looks at the same place: which line of your AI spending shows up on the income statement? This piece is about what to do when that question has no answer. If a team comes to you claiming “productivity is up X percent”, the Growth Claim Check shows in five questions whether the number is ready for a real-terms reading.

Continue, redefine, shut down

The answer to the test question takes every initiative to one of three decisions.

Continue. The line, the target number, the timeline and the signature are written down. The early indicator (hours saved, proposals processed) is moving towards the target. Give it the budget and look at the line in 90 days.

Redefine. Hours really are being saved, but they have nowhere to go. This may be the case you meet most often, and it may be where most of the value is hiding. Rewrite the target. “400 hours a month on invoice matching” becomes “not making the two hires planned for next year, owned by the finance director”. This is also where the J-curve argument is right: name the complementary investment and write down how long it takes.

Shut down. The line can't be written, nobody will sign, or the hours themselves can't be measured. Shutting down doesn't have to be a declaration of failure; it closes as a lesson paid for out of the experiment budget. The cost of shutting down is usually low (interpretation). The cost of carrying on is the same silence at the next budget meeting.

One expectation to set: this test can be unfair to some initiatives. Work that lowers error rates or improves customer experience reaches profit indirectly and late, so it looks weak in the test. Before shutting those down, look for the indirect line. If the error rate falls, does the returns, discounts or support cost line change? If you can find a line that changes, the initiative continues. If you can't, it's an experiment too, and it should be labelled as one.

Try it today

Set aside an hour, with last quarter's income statement next to you.

Write every AI initiative running in the company on its own row: pilots, enterprise licences, tools teams use on their own.
Next to each row write the income statement line that should change (revenue, COGS or operating expenses), the 90-day target number and the owner. The owner should be the person who signs that line's budget. To see which line depends on which input, the Metric Tree makes the job easier.
Mark any initiative you can't write a line for as an “experiment”.
Make one decision for each experiment: redefine or shut down. “We'll see” isn't an option.
Split the saved hours three ways: headcount, sales and neither. If the “neither” share is more than half the total (rule of thumb), look at the missing decisions before you look at a new tool.

Write these six fields for each initiative. Three examples:

Invoice matching (example). Line: operating expenses. 90-day target: 2 planned hires not made, 180,000 TL a month. Owner: finance director. Hours saved per month (headcount / sales / neither): 320 / 0 / 80. Decision: redefined, continue.
Proposal drafts (example). Line: revenue. 90-day target: from 60 to 100 proposals a month, 8 extra orders. Owner: sales director. Hours saved per month: 0 / 240 / 0. Decision: continue.
Team chatbot licences (example). Line: couldn't be written. Target and owner: none. Hours saved per month: unknown. Decision: experiment; redefine or shut down.

The first version of the list will be incomplete; the owner may be the field left empty most often. Every empty field already shows you the question that will be asked at the next budget meeting.

Take care.