Does Enterprise AI Actually Pay Off?
The J-curve against the 95%, and the number that would settle it.
The Question
Can organisations show, at firm level (not anecdote), that AI spending pays?
Optimist
The Optimist's Case
Returns lag by design. Evidence This is the J-curve pattern: productivity dips before it climbs, exactly as it did with electrification. The 5% seeing returns redesigned workflows rather than automating in place. That is the signal, not the noise.
Capability compounds. Unit costs fall quarterly. Buy-and-partner succeeds more than internal builds. Interpretation The firms that will show returns are the ones that treated AI as infrastructure redesign, not feature addition.
"No payoff yet" is exactly what the early J-curve looks like. The question is not whether it pays off, but whether you lose nerve before the curve turns.
Sceptic
The Sceptic's Case
MIT NANDA study (2025)
Evidence By April 2026, 44% of enterprises were funding the next AI wave from savings that had not yet materialised (paywalled). In May 2026, corporate America began rationing AI access (paywalled) as costs outpaced budgets. The UK government's Copilot trial found no robust evidence that time savings were leading to improved productivity.
InterpretationCircular financing flatters the picture: AI providers spend on infrastructure, enterprises spend on AI, both call it growth. "The J-curve will deliver later" is unfalsifiable faith.
The burden of proof sits with the optimist. The proof, so far, is mostly anecdotal. At firm level, attribution remains elusive. Evidence The pattern from previous technology waves is that most spending does not pay off, and the winners are identifiable only in retrospect. When tier-1 enterprises like Uber blow through annual AI budgets in four months, the J-curve story becomes harder to defend.
Synthesis
House viewWhat Would Settle It
A durable, attributed, firm-level value number at named companies. Not "productivity improved" but "revenue rose by X, cost fell by Y, and we can show the AI contribution with this method". Audited, not asserted.
Interpretation The tell will be attribution coverage: the proportion of AI spending for which value can be demonstrated. If that number rises over time and the value gap shrinks, the optimist wins. If coverage stays low while spending scales, the sceptic wins.
Track attribution coverage across multiple reporting periods. That is the scoreboard.
If We Get It Right / If We Get It Wrong
Right + act: Compound advantage. The firms that redesigned early and measured honestly will show durable returns and pull away from peers.
Right + lose nerve: Premature cancellation. The J-curve turns after you stop funding it, and the advantage goes to whoever held steady.
Wrong + act: Capital redeployed. If AI does not pay off, the firms that measured will know it first and redirect spending before the damage compounds.
Wrong + ignore: Good money chases bad. Unmeasured spending continues because no one can prove it failed.
The Author's Honest Position
Both the J-curve and the bubble risk are real, and they are not contradictory. The resolution is measurement, not optimism or scepticism.
Insist on attribution coverage as the running scoreboard. Redesign rather than automate. Interpretation The 5% showing returns are the ones that changed how work happens, not the ones that added AI to unchanged workflows.
I will not defend spending at scale while measuring nothing. But I will also not declare the J-curve broken before it has had time to turn. The evidence is still being written.
No declared winner. Watch the attribution coverage number.