Why your AI cost went up but not productivity?

“You can see the computer age everywhere but in the productivity statistics.” When Nobel laureate Robert Solow wrote that in 1987, he could not have known that four decades later the same line would describe AI.

We can see AI on every CFO’s agenda, in every board paper, in virtually every annual report. Where we still cannot find it is the P&L. A recent McKinsey study states that while 80% respondents say AI has improved their individual productivity, the share reporting any impact on EBIT sits at 37%, same as last year.

I have watched this productivity paradox play out across most of the firms I advised this year, and most importantly in mid-sized ones. Nearly all of them have been through the first phase of AI excitement. They signed up for enterprise Copilot, Gemini or Claude licences. Individuals and teams now write faster emails, churn out PowerPoints, and draft variance analysis (which, in most cases, even the CFOs admitted was largely AI slop). A few tech-savvy people-built agents to take care of repetitive work. I have seen many green dashboards showing hundreds of hours saved.

Then comes what I have started calling the great realisation phase. The CFOs and CEOs I work with are asking the same question: where is my productivity? If the team is saving that much time, where is it going? Why are people still working overtime, why do stakeholders still struggle to get information when they need it, and why are heroics still the only way to deal with uncertainty? Meanwhile the cost of owning AI keeps moving. Though a subscription arrives as part of the enterprise licence but soon the bill shocks start. At one firm I worked with, number of employees used up the entire monthly usage quota in just one day.

The bill shock and the missing productivity are the same problem. It is usage without a design. Tokens go wherever people point them. If pointed at writing faster email, that is where they stop, and with them the time saved as well. The hours saved remain with the employee saving them and never converts into any P&L impact, since the workflows are not designed to convert. The most rigorous study so far, which tracked 25,000 workers across eleven occupations in Denmark, including accountants, found that most users saved time and that earnings and hours did not move in any occupation.

The silver lining is however that the realisation phase is making CFOs more deliberate. An AI feature bolted onto an existing tool is no longer enough to earn an approval, and that is exactly as it should be.

The firms that have come out of this phase share a pattern, and it has nothing to do with which tools they bought. Their leadership was already in the realisation phase: a number of AI subscriptions in place and no business or people metric moving. The diagnosis was almost identical every time: significant shadow AI, a sprawl of tools, and a belief that “some AI tool” will solve the problem. Three things happened that set these firms apart:

The future operating design came first

AI changes the way we work. That is a cliché by now. How it changes the way we work is the question, and this is where these leadership teams spent most of their time. They defined what the business would look like in three years: how it would capture, create and deliver value to clients when every competitor could produce the same output faster. When speed is universal, the client stops paying for it. A premium has to come from something else i.e. the firm’s own knowledge of its clients, its own redesigned way of working, and its own judgement. Everything worked backwards from that. The entire future operating model and where value would be created fitted on a single page. That page was the asset. The software underneath is just a rented piece of “equipment”.

Partners were chosen against the design

An almost knee-jerk reaction to any target operating model is to start finding the vendors who can deliver that. It was fine for discrete software products, but with Gen AI and agentic AI being a fundamentally pervasive technologies requiring an almost human-like co-ordination, a couple of vendors are not enough. We need to be able to step back and connect the end-to-end workflows within the organisation. For example, we looked at how does a client proposal sends the right signals to pertains team while enabling FP&A to re-cut the forecasts before any surprises come up.

Look for partners who are flexible enough to work through to solutioning the achievement of your specific operating design as opposed to just selling a product. However, also be open to have multiple partnerships as no single tool exist that solve for everything. You will need to build a thoughtful stack of right AI solutions which can work with each other and humans, within the right frameworks and guardrails that you define.

Execution happened – thoughtfully

The line between experimentation and a pile of dead pilot-projects is finer than we think. The entry barriers have never been lower. A few thousand dollars buys a good working proof of concept, and some vendors will give you one for free. That is precisely the problem. Without the big picture, point solutions accumulate that never work with each other. The firms that got through did something unfashionable first: they inventoried what their existing stack could already do and spent the first few months experimenting with nothing new. One team I worked with proposed around seventy use cases within four weeks; from centralised knowledge bases and redesigned workflows to role-specific assistants. The realisable value was around half a million dollar. Personal experimentation was free, but anything that touched a firm-wide workflow or client, went through a small digital council before being approved. This ensured we did not end up with a pile of dead pilot projects.

Solow’s paradox was eventually resolved. Productivity took off in the late 1990s, and the economists who studied it concluded that the gains arrived once firms reorganised work around the technology, years after they bought it. AI will reach the P&L the same way. A mid-sized firm does not need a decade. It needs a design, a gate, and one honest question: where does your productivity currently sits?