What is the value of your AI spend?

The question everyone is working through right now.

AuthorAuthor

Ellen Köhler & Simone Rüschenberg
August 27, 2026 • Estimated Reading Time: 6 minutes

Read of the week

One question is occupying CFOs and companies alike right now: what is the value of our AI spend?

Ask how much goes into the AI budget for 2027, and the answer is almost always the same. “A lot.” Then come the harder questions. Where do revenues rise at the same rate? Where do the savings come from?

Nobody has a clean answer yet, and we will report more through the second half of the year. The budget round itself gets its own issue soon. But some switches can be set now, so here is some food for thought.

00· · In this issue

01· · When the estimate is wrong: three examples
02· · What is already measured, and what is not
03· · You own this one: COGS, OpEx or capitalised
04· · Before you approve the next AI line

01· · When the estimate is wrong: three examples

Start with three companies everyone knows, where a wrong estimate was no small internal mistake but showed up in growth, in budgets and in the share price.

Canva launched Canva AI 2.0 in April. Demand ran past the plan and each task cost too much to serve, so Canva slowed its own rollout and moved onto its own models, cutting cost per task by roughly 90%. Still, in August it told shareholders growth would be 20% instead of 30%. Canva is private, so no share price fell. But with an IPO targeted for 2027, cutting your growth forecast by a third is the more expensive version.

Uber first encouraged usage, including a leaderboard ranking teams by AI consumption. By April the year's budget for AI coding tools was gone, single engineers billing up to 2,000 dollars a month. In June came the cap, 1,500 dollars per employee per tool. Uber's own COO said in May that the link between usage and better consumer features is “not there yet”.

Meta raised its capital expenditure guidance in April by another 10 billion dollars, to a range of 125 to 145 billion, most of it AI infrastructure. Asked what all that investment returns, the CEO called it a very technical question and left it there. The quarter beat expectations, and the shares still fell more than 6% after hours, which analysts put down to the raised guidance.

💡 All three could see their costs. What none of them could show in time was the value. Answer that before someone asks.

02· · What is already measured, and what is not

None of those three was flying blind on cost. Almost every company we speak to can see the cost side now: token boards, usage boards, cost by provider and model, and which employees work with their licences. Most have a P&L line for it, rare at the start of this year.

You know what was consumed and who consumed it. What the boards cannot show is what came back.

Interesting numbers come from ICONIQ, a well-known US investment firm. It surveyed around 305 executives at software companies building AI products, from under 5 million to over 1 billion dollars in annual recurring revenue, about 85% of them in North America. These are the companies furthest down this road. Not your peer group, but your preview:

If revenue does not climb at the same pace, the gap lands in EBITDA, and in net profit once depreciation hits.

On the value side, SAP and Oxford Economics surveyed 2,600 director and C-suite executives in 13 countries in July. Most of what they report are self-assessed expectations, but one number is a control issue, and that is our field: 69% see shadow AI, tools used without approval, at least occasionally. That spend never reaches the dashboards, so cost and value are both undercounted.

Counting is step one. Booking it right is step two, and that one is yours.

💡 Almost everyone can count the spend now. Very few can say what it bought.

03· · You own this one: COGS, OpEx or capitalised

This is not the finance team's tool budget. As CFO you own the AI budget for the whole company, because you stand behind the number in the end. That starts with booking it correctly.

The data exists. Claude Enterprise reports cost per user and per department. ChatGPT Enterprise exports analytics per user and project. Microsoft shows Copilot adoption in the admin centre. Only coding tools in your own cloud need a gateway or your own telemetry.

Then one question for every Euro you book.

  • Did a customer trigger it, inside your product? COGS, and it hits gross margin.

  • Everything else follows the team using it. Engineering into R&D, Sales and Marketing into S&M, Finance and HR into G&A. Never one central AI line.

  • Teams already sitting in cost of revenue, support and professional services at most software companies, take their AI with them. That is why customer care feels ambiguous: what decides is where the function sits in your P&L.

Engineers are the harder case. Under IFRS, IAS 38 expenses research and capitalises development only when all six criteria are met (technical feasibility, intention to complete, ability to use or sell, probable future benefit, sufficient resources, reliable measurement). AI woven through a core product rarely clears that bar. And capitalising only moves the cost to the balance sheet. It never leaves the cash flow, and boards back it out anyway.

💡 If AI still sits inside a blended COGS and DevOps number, you are running margin blind.

04· · Before you approve the next AI line

Nobody has a clean answer on value yet. But most AI spend gets approved without anyone naming what it should do, and that part is fixable today. Five questions, asked before the money goes out.

  • 1 · Revenue or cost? Not efficiency, not enablement. If a line cannot be assigned to one side, it is not a business case yet, it is a pilot. Fund it as one.

  • 2 · Which number moves? One named metric, not a category. Revenue side: price, attach rate, conversion, retention. Cost side: cost per ticket, close days, cost per claim. “Sales will be more productive” is not a metric.

  • 3 · What is that number today? Baseline it before the tool goes in. This is the one we see missed most often. Without a before there is no after, and the vendor's number becomes the only number in the room.

  • 4 · Who owns it? The business owner whose metric it is, not IT and not finance. Finance owns whether the spend is booked correctly, not whether the tool works.

  • 5 · When do we stop? Set the date at approval, because nobody sets it afterwards. Ninety days: either the number moved and the line scales, or it comes out.

Back to our three examples Uber had no stop date, so the budget ran out before anyone reviewed it. Canva had no baseline for cost per task, so the overrun only showed up in the growth forecast. Meta could not name the number the spend was meant to move, and got asked in public. None of them lacked data. All three lacked the question.

💡 You cannot measure the value of an AI line nobody defined before approving it.

Bottom Line

  • If you cannot yet see tokens, usage and cost by model, set that up first.

  • Then show it correctly in the P&L: COGS, OpEx, capitalised where allowed. Without that baseline there is no value to measure.

  • And before the next AI line is approved: name the side, the metric, its number today, an owner and a stop date.

We will keep sharing what we see through the second half of the year.

CFO Watchlist

European CFOs are more pessimistic than at any point since the 2022 energy crisis

Deloitte surveyed 1,136 finance chiefs in 12 countries. Cost discipline, resilience and scenario planning are back on top:

  • The share feeling less optimistic than three months earlier doubled, 25% to 48%.

  • 76% rate external uncertainty as high or very high, back to 2022 levels.

  • Only 26% plan to raise capex, 85% say this is not the time for balance sheet risk.

  • 36% expect margins to decline even though more than half expect revenue growth.

  • Cost reduction is a top three priority in every country surveyed, a first.

Surveyed March and April 2026, published 7 August 2026

Klarna cuts its outlook on weak German retail

Klarna beat its second quarter and still cut full-year revenue guidance, blaming soft discretionary retail spending in Germany, its largest market by volume. Worth a closer look if Germany sits in your 2027 plan.

Q2 results and guidance cut announced 18 August 2026

Podcast: Ray Dalio on the AI bubble and what comes next

The man who predicted the 2008 crash says the warning signs are flashing again. Ray Dalio, founder of the world's largest hedge fund, on why we are in an AI bubble, why he thinks the US has entered its decline, and what comes next.

The Diary of a CEO, published 30 July 2026

FROM SIMONE

Finance Collective goes Nordic

15 September 2026 · Stockholm

Finance Collective is the peer network Simone runs for finance leaders. We match you with the right peers by company size, segment and industry. Live Q&A on our platform, expert sessions, and regular meetups in our local hubs. In September we are adding the Nordics.

WHO WRITES THIS

Two CFOs, two vantage points.

Ellen sits in the operating team of a PE growth fund and works closely with scale-up companies. Simone spent 20 years running finance functions, then switched sides to build Finance Collective, a network of CFOs. Every second Thursday we write about what's on finance leaders' desks right now, and talk to people from our networks who have already solved it.

Feedback, suggestions, or something you want us to cover? Write us on [email protected].

CFO Playbook reflects our personal opinions, not those of our employers and is not professional advice.