If you invest in AI apps, run an agent company, or are measuring AI software with a SaaS ruler — here’s a structural change the market badly underrates: software has a cost of goods sold for the first time.

Traditional software sells a tool: you buy a better hammer (a seat of CRM, an IDE); how you use it is your business; pricing is per-seat. The agent changes the verb — it’s no longer a tool you use, it’s an executor that does the work for you. Pricing slides from “how many use it” to “how much work got done” — per-outcome.

That verb change is two inversions in the underlying economics of software: one up, one down.

Up: TAM Jumps From the “Software Budget” Into the “Labor Budget”

The software industry spent thirty years fighting over one pot: the enterprise IT budget. Selling outcomes punctures that ceiling — when an agent delivers “completed work,” it’s benchmarked not against a software subscription but against the money you’d have paid a person to do it. And the labor budget is an order of magnitude larger than the IT budget.

That’s the hard meaning of “agents rewrite software”: not that software got better, but that software is for the first time eligible to chase a far bigger pot it never could reach. An agent that replaces a junior analyst or a support rep is priced not “a bit more than comparable SaaS” but “a bit less than that person.”

If the story ended here, agents would be the best business in history. But there’s the downward inversion.

Down: Software Has a COGS for the First Time

SaaS is one of the best business models ever invented, for one reason: zero marginal cost. Write the code once, sell another seat for almost nothing — hence 80%+ margins, hence bigger-is-more-profitable.

The agent adds cost back. Every agent run truly burns tokens — calling the model, retrieving context, multi-step reasoning, tool calls, each paid upstream. Serving one more user, completing one more task, actually costs money. Software now has what manufacturing has: variable cost, a COGS that grows linearly with scale.

The consequence is structural, and most agent valuations haven’t digested it: margins fall from 80% toward a much lower number set by token cost; scale no longer auto-improves margin. The only external salvation for margin is betting that token cost falls faster than price gets competed down — a business racing the upstream deflation.

And a Hidden Liability: The Reliability Tax

In the tool era, software errors are the user’s problem; in the outcome era, when the agent errs, the seller of the agent pays. That’s the reliability tax — it shows up as “human backstop” (giving back the labor you saved), or “error liability,” or “only daring to take low-risk tasks.”

This explains the recurring pattern: many agents dazzle in demos, catch fire in pilots, then stall at scale. What stalls isn’t capability — it’s that “error cost × error rate” eats the entire premium of selling outcomes. A 95%-correct agent, in a high-error-cost scenario, can be more expensive than a tool a human watches the whole time.

The Only Profitable Seam

Put the three accounts together and the condition for an agent business is an inequality:

Outcome value > token cost + reliability tax (error cost × error rate)

This also explains why “code” was among the first to scale: outcome value is high (replacing expensive engineering time), reliability tax unusually low — code ships its own free referee (did the tests pass), errors caught instantly. A clear right/wrong signal is what lets “selling outcomes” stand up economically.

Three closing lines. Don’t value agent companies with a SaaS multiple indiscriminately — first ask the unit economics: is per-outcome token cost on a declining track, does the reliability tax require human backstop. Most overvalued: the agent that dazzles but scales on human backstop (bigger = more losses). Most undervalued: the agent in a scenario with a free referee, cost falling with deflation, benchmarked to high labor cost. The market prices agents at SaaS multiples without discounting their dirty income statement — that mismatch reprices eventually.

The agent company you’re looking at — is its per-outcome token cost on a declining track? Does it need heavy human backstop? Comments open.

Sources (framework argument, June 2026): token-cost deflation (9–900x/year, Stanford AI Index 2025); ~70% of spend flowing upstream (Epoch AI).

— From Chapter 9 of a book in progress, working title The Deflation Sandwich

#AI #Agents #DeflationSandwich