Why Rising AI Spend Isn't Producing Growth

For about a hundred years, running a company came down to one piece of arithmetic. Production was the constraint. If you wanted more growth, you hired more producers. Every org chart, every funding round, every annual plan was built on that equation: headcount in, output out, growth follows.

The equation just broke, and it broke from a direction nobody planned for. Production stopped being scarce.

You can see the break in the spending data before you see it anywhere else. Gartner puts worldwide AI spending at roughly $2.5 trillion for 2026, up from about $980 billion in 2024. Ramp's card data across 70,000 companies shows the top 1% of spenders paying a median of $7,400 per employee on AI as of July, nearly triple January's figure. Uber reportedly exhausted its entire 2026 AI budget by April. Companies are pouring money into producing more, exactly as the old equation instructs.

And the growth is not arriving. Deloitte surveyed 3,235 leaders across 24 countries and found 66% reporting productivity gains, but only 20% seeing AI-driven revenue growth. Fewer than 1% of organizations report returns above 20%. The strangest data point comes from OpenAI itself, which analyzed more than 1,500 organizations this year and found that revenue per employee was not meaningfully associated with how many tokens those employees burned. The company selling the tokens measured it and published it. Consumption does not predict growth.

Busy is the new stalled

Inside these companies, nothing feels wrong. That is what makes this expensive.

Everyone is producing more than ever. Drafts, decks, variants, campaigns, code. The feeling of work is fully intact, in most places intensified, because generation invites more generation. A team that could make three concepts now makes thirty, and reviewing thirty feels like a heavy, productive week.

It is fake work if the market doesn't move, and mostly the market doesn't move. Sometimes it is worse than indifferent. When everyone's output comes from the same models prompted with similar words, the results converge, and buyers have learned to recognize the convergence. The research showed this early: Doshi and Hauser found in Science Advances that AI raises individual output quality while narrowing what a group collectively produces. Everyone gets better and more alike at the same time, and alike doesn't win anything.

So the old equation now runs in reverse. Buying more production doesn't buy growth. It buys volume the market has already priced at zero, plus a token bill that compounds monthly.

Where the constraint actually went

If production is abundant, the constraint has to be somewhere else, and it is not hard to find. It sits in everything that happens on the way to the product. Knowing what is worth making. Choosing between the thirty variants. The taste that kills twenty-eight of them. The distinct, slightly strange execution that a model trained on everyone's average cannot produce, because average is precisely what it is trained toward.

Which means the companies getting real returns are using AI in a completely different shape than the ones stalling. The stalled ones bought AI as a product-stamping machine, bolted onto the end of the process to generate the deliverable. The winners thread it through the path instead: a small assist inside research here, a summarization inside review there, a variant generator inside one niche workflow, each piece attached organically where it removes friction, with people holding the spine of the work the whole way. PwC's finding fits this exactly. The 20% of companies capturing 74% of AI's value are distinguished by having redesigned how the work flows, not by generating more at the end of an unchanged process.

The distinction sounds subtle and is everything. AI at the end of the process replaces the deliverable. AI along the path amplifies the people making it. Only one of those compounds.

Growth was always people. Now it's only people.

Strip the equation to what remains scarce and you get an uncomfortable answer for anyone whose plan is a bigger token budget. The best talent is scarce. The best thinking is scarce. Ideas that didn't come from the same distribution as everyone else's are scarce. Execution with an actual signature is scarce.

Those were always the real inputs to growth. The old world let you approximate them with volume, because when production was hard, whoever produced most won by default. That subsidy is gone. Now the market only pays for what the abundant thing cannot supply, and every one of those things lives in people, in how they think together, and in whether the path from idea to shipped work makes them sharper or just busier.

This is what we built ALLO for, and the connection is direct. If the constraint is the path, the path needs somewhere to live. On a canvas, the thirty variants sit side by side where choosing between them is a real act instead of a scroll. The research, the argument, the rejected directions, and the decision stay attached to the work, so the team's thinking compounds instead of evaporating into threads. The token bill buys generation. The canvas is where generation gets turned into the one thing the market still pays for, a choice somebody stood behind.

Companies are starting to perceive this. The budget conversations are shifting from how much AI to where in the workflow, and the ones asking the second question are pulling away. The equation didn't stop working. It moved. Growth still comes from investing in the constraint. The constraint is just no longer the making.

FAQ

Why isn't AI spending producing revenue growth? Because spending is aimed at production, which is no longer the constraint. Deloitte's 2026 survey found 66% of leaders report productivity gains but only 20% see AI-driven revenue growth, and OpenAI's own analysis of 1,500+ organizations found token consumption doesn't meaningfully track revenue per employee.

How fast is enterprise AI spending rising? Gartner estimates worldwide AI spending at about $2.5 trillion in 2026, up from roughly $980 billion in 2024. Ramp's data shows top-spending companies nearly tripled per-employee AI spend in the first half of 2026, and Uber reportedly used its full-year AI budget by April.

What is the new bottleneck if not production? The path to the product: deciding what is worth making, choosing between abundant options, and executing with enough distinctiveness that the output doesn't blend into everyone else's. Research shows AI narrows collective diversity even as it raises individual quality.

How should companies actually deploy AI? Woven into specific workflows along the path, small assists in research, review, and iteration with people holding the spine, rather than bolted onto the end as a deliverable generator. PwC found the minority of companies capturing most of AI's value redesigned their workflows this way.

What does ALLO have to do with this? ALLO is where the path lives. Generated options land side by side for real comparison, and the research, feedback, and decisions stay attached to the work, so a team's judgment compounds instead of disappearing into chat threads.

--
Business ran for a century on one equation: production was the constraint, so adding people bought growth. AI broke the equation from the production side. Output is now cheap and abundant, spending has rotated from headcount to tokens, and growth has not followed: Deloitte's 2026 survey of 3,235 leaders found 66% report productivity gains while only 20% see AI-driven revenue growth, and OpenAI's own study of 1,500+ organizations found revenue per employee is not meaningfully associated with token output. The constraint moved. It now sits in the thinking, the deciding, and the distinctiveness of execution, and that is where investment has to follow.