Capability Accelerates, Investors Extrapolate, Payoff Lags

Every few months a new release clears a benchmark that was supposed to be years away. The models are improving on a curve almost nobody predicted, and the market makes the obvious leap: that same curve will carry productivity and financial results, across a wide range of businesses, and will happen very soon.

That leap has the timing all wrong.

Capability moves on one clock. The business that has to absorb it, moves at a much slower pace. A new model lands in an afternoon. Turning that model into productivity, and productivity into earnings is the slow part — it takes the organization through a series of gates, and the gates open at the speed the organization moves, not the speed the model does. The more that a gate matters, the longer it takes, because each one forces risk and reward to be weighed at the same time.

The market has no patience for this. It prices the thesis the moment the headline hits, unwilling to wait for a result that shows up diffuse and late. It wants the payoff on capability’s timetable; the payoff runs on that of adoption. Underwrite near-term margin expansion off the capability curve and you are reading the right trend at the wrong tempo.

I made the supply-side version of this case in Pent-Up Drag¹. This is the demand side. Even the companies successfully deploying all that compute will not see the gains reach their financials for many quarters — and the market is crediting them today. The bigger the need, the greater the hope. A recipe for disappointment.

The First Gate: Cost Rises Before Efficiency Arrives

For two years, enterprise AI ran on unconstrained compute. No caps, the best model on every problem, the bill a rounding error against the option value of figuring it out first. That produced perverse habits like token-maxing — defaulting to the biggest model and the longest context on reflex, because the cost of overkill was invisible.

That phase is ending, and how it ends is the tell. Companies are rationing: capping consumption, throttling access, pulling back the expensive models. Rationing is not sophistication. It is a blunt reaction to a cost that spiked well past budget — if it was budgeted at all.

The cap is a stopgap. The real fix is to match each task to the right model — the cheap one for simple work, the expensive one only where it earns its price — and to build the rails that route reliably between them. That is not a knob you turn; it is infrastructure you build, and it takes engineering, measurement, and time.

So the near-term signature of AI adoption is not margin expansion. It is a step up in cost — the spike that forced the rationing, plus the build required to replace the cap with something smarter. Margins may compress before they improve.

The Second Gate: The Tool Speeds Up the Wrong Step

Suppose the tool works and the cost is managed. Cycle times still barely move, and the reason is structural, not cultural.

Every process is paced by its slowest binding step, and the tool rarely replaces that step. Drafting was never what made contracting slow; review, negotiation, and the counterparty’s response were, and a model that drafts in seconds leaves all three untouched. The work that was already fast gets faster. The constraint that governs the clock sits exactly where it was. Speed applied to a non-binding step is invisible in the result.

Moving the real constraint is the harder project: rebuilding approval chains, redrawing handoffs, retraining people, reassigning accountability. That work is slow and political because it touches roles and headcount. I argued in The Layoff Lie that AI is mostly giving executives the courage for restructuring what they already wanted². The corollary holds here: the restructuring is the part that unlocks the gain, and it runs in years, not release cycles.

The capability is installed the day you buy the license. The result waits on the constraint being moved.

The Third Gate: Gains Have to Be Harvested and Held

Now assume the constraint has moved and real productivity exists. For most companies it still does not reach the financial statements quickly. A productivity gain is a cost saving, and a saving becomes earnings only if someone collects it — and then manages to keep it.

Most never collect. The freed capacity is absorbed into more scope and more meetings rather than removed. An engineer who works thirty percent faster does not become thirty percent of a salary the company can bank. Reaching the income statement means deliberately harvesting the gain — taking headcount out, or redirecting it to revenue — and that decision gets deferred for the same reasons the restructuring did.

Collecting is necessary but not sufficient. I argued in Efficiency Masquerading as Inflection that the gains do not spread evenly³. Early movers capture the outsized share; laggards subsidize the transition; and once everyone holds the same tool, the saving gets competed away to the customer through lower prices. Holding the gain requires a position competitors cannot quickly erode — pricing power, a structural moat, proprietary data or workflow ownership. Where those conditions hold, the gain is banked and durable. Where they do not, it is absorbed or competed away. The result reaches the print only where the business was built to keep it.

What This Means for the Print

This is not an argument that AI fails to deliver. It is an argument about tempo — and about mispricing the result. Converting adoption into durable financial leverage is hard work, the execution risk is real, and the first effects are likely to be negative before they turn.

I should say where this breaks. The lag is not uniform. Where the binding constraint is the very step the model replaces — and in a few narrow, software-native workflows it is — the payoff can arrive on something close to the capability curve. Those cases are real, and if they generalized I would be wrong. My bet is that they do not, because most enterprise value sits behind constraints that are organizational rather than technical, and those move at organizational speed. So here is the observable that settles it: broad, cross-industry margin expansion credited to AI within the next few print cycles, not a handful of native exceptions. I do not expect to see it — but I would rather name the test than let every outcome confirm the thesis.

Look to the management teams doing the slow work quietly, rather than the ones promising the curve will arrive next quarter. They win not because they hold the best model, but because they are built to keep what it gives them.

It only looked free. The payoff is real, but it lags the curve that promised it. The market wants instant gratification. The companies that earn it are built to wait.

Footnotes 

¹ Mark Scalzo, “Pent-Up Drag“, Validex.co, May 26, 2026.
² Mark Scalzo, “The Layoff Lie“, Validex.co, March 11, 2026.
³ Mark Scalzo, “Efficiency Masquerading as Inflection“, Validex.co, May 8, 2026.