Should you still invest in AI? What the sceptics say

Less than 0.66% cumulative productivity gain over ten years: that is the estimate by Daron Acemoglu, the 2024 Nobel laureate in economics, for AI across the US economy. A Goldman Sachs report airs doubts about an estimated USD 1 trillion of expected capital spending. Djtal has read this work closely. Scaled down to a single company, the arithmetic points to a sound way to invest, one task at a time.

What the sceptics say

In May 2024, Daron Acemoglu (MIT) published ‘The Simple Macroeconomics of AI’ as a National Bureau of Economic Research working paper (no. 32487). He estimates a total factor productivity gain of less than 0.66%, cumulative over ten years, for the US economy. Leave out the tasks that are hard to learn and it falls below 0.53%. That is an order of magnitude below the projections of investment banks and consultancies.

In June 2024, Jim Covello, Goldman Sachs’s head of global equity research, was interviewed for the bank’s report ‘Gen AI: too much spend, too little benefit?’. The report put the capital spending expected over the following years at an estimated USD 1 trillion (data centres, chips and the power grid), for a technology Covello considers poorly suited to the complex problems that would justify the bill. Two years on, he admits he was wrong about some things, but on the central question, whether that spending earns a return, he has only grown more convinced (Fortune, 6 May 2026). Robert Gordon (Northwestern) completes the picture. In his view, the share of workplace tasks AI can automate is too small to move the wider economy much, and its gains are minor next to the switch from horses to trucks (Marketplace, March 2025).

The hasty conclusion would be to shelve your AI budgets. Their own calculations point the other way.

The arithmetic behind the figure

Acemoglu’s estimate rests on two factors: the share of tasks an AI can take over, multiplied by the saving made on each one. His caution comes from the first factor. Across a whole country, the share of exposed tasks stays modest. The second factor holds up. On a well-chosen task, the saving is real and measurable.

Everything turns on that framework. AI’s gain is a sum of tasks long before it is a national percentage.

At your company’s scale, the first factor is yours

At the scale of a national economy, the share of automatable tasks is something that happens to you. In a single company, you choose it. An admin team that repeats the same data entry, follow-ups and reconciliations in its ERP every day has, in one place, exactly the tasks an AI handles well.

The national figure hits a ceiling because it averages out exposed jobs and unaffected ones. Your company is free to choose which side of that average it sits on.

The investment method that follows

1. Take stock of the repetitive work where it lives: ERP, CRM, shared mailboxes.
2. Put a figure on what each task costs every month: hours spent multiplied by the fully loaded hourly rate.
3. Start by automating the tasks that pay for themselves within a few months, on the systems you already have.
4. Measure the gain achieved, in hours and in Swiss francs, before extending anything.
5. Extend one task at a time, as fast as the observed paybacks allow.

Nearly all of the 95% of pilots with no measurable return (MIT NANDA, 2025; Fortune, 18 August 2025) did the opposite. They bought the revolution without counting the tasks. The gap between their expectations and their results exactly mirrors the gap between the consultancies’ projections and Acemoglu’s calculation.

The academic debate can carry on without you

Before your next AI budget, list your tasks, what they cost and what they would return once automated. That one-page document is worth more than any market growth projection, whichever way it points.

---

Djtal’s AI strategy audit produces this inventory for your company in 24 hours of work spread over two weeks, with every task costed. Get in touch.