.png)
.jpg)
Frameworks, core principles and top case studies for SaaS pricing, learnt and refined over 28+ years of SaaS-monetization experience.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
There's a seductive narrative forming in the AI agent market right now: that agents can be priced like people, and that this pricing power is durable.
Tomasz Tunguz articulated it best in a recent post. His argument is sharp, data-backed, and - for the next 12 months - largely correct. But it mistakes a temporary arbitrage window for a structural advantage. And for founders building agent businesses, confusing the two could be fatal.
Let me steelman his case before I explain why the clock is already ticking.
The Arbitrage Case (And Why It Works Today)
Tomasz observes that in labor-shortage markets, AI agents are already commanding 75%, 85%, even 100% of a human equivalent salary. And he stacks a genuinely compelling case for why buyers are willing to pay.
First, the obvious: the work gets done. In markets where you can't hire fast enough, an agent that ships output today beats a recruiter who might find a candidate in 90 days. That urgency alone creates willingness to pay at or near human rates.
Then he layers second-order benefits that are hard to argue with. Training agents is dramatically faster - you feed them all materials at once, in parallel. They need less management overhead. They work 24/7, scaling up or down as the team needs. Capacity becomes a spend decision, not a headcount problem.
He backs this with macro data that's hard to dismiss. Goldman Sachs found that low-labor-cost stocks outperformed high-labor-cost ones by 8 percentage points in 2025. Labor's share of GDP hit a record low of 53.8% in Q3 2025. Across the S&P 500, labor costs sit at roughly 12% of revenues while software costs hover around 1-3%. As agents absorb labor, that ratio inverts. Software TAM grows at labor's expense, and profitability grows with it.
His conclusion: "No pricing competition on a per-agent basis. Vendors aren't racing to the bottom; they can price at par to a person."
Every link in this chain is factually correct. But the conclusion confuses an arbitrage with a moat. Here's the distinction that matters.
"Value" Alone Doesn't Set the Price
In theory, pricing is based on value delivered. Every pricing textbook says so. And Tomasz is absolutely right that agents deliver enormous value relative to their cost.
But in practice, real market pricing doesn't hold at value. It holds at the level set by competition and alternatives.
This is the part that gets lost in the excitement around agent economics. Yes, an agent that does the work of a $100K employee is delivering massive value. But the price a buyer actually pays isn't determined by that value ceiling. It's determined by the next-best alternative available to them.
If you're the only agent vendor in a category, you can price at value. Congratulations - you have a monopoly, and monopolies set prices however they like.
But the moment a second vendor shows up - or the buyer realizes they can build their own - the pricing anchor shifts. It moves from "how much value does this create?" to "what's the cheapest way I can get comparable value?" Those are two very different numbers.
Tomasz's argument implicitly assumes the pricing anchor stays at "human salary" indefinitely. It won't. Because the anchor only holds as long as the alternatives aren't in the picture. And in the agent market, the alternatives are arriving at unprecedented speed.
Tomasz's argument describes an arbitrage: the spread between the cost of running an agent and the cost of hiring a human. Today, that spread is enormous. Buyers are happy to pay 75-100% of a human salary for something that costs a fraction of that to deliver.
But nothing in his argument explains why that spread is defensible. And the data on inference costs says it isn't.
The Spread Is Collapsing - Fast
AI inference costs are dropping at a rate that has no precedent in technology history. This is the force that turns today's arbitrage into tomorrow's commodity.
Andreessen Horowitz coined the term "LLMflation" to describe what's happening: for a model of equivalent performance, inference cost is decreasing by 10x every single year. The cost of using GPT-3 quality inference has fallen 1,000x since 2021. And this isn't limited to budget models - the cost curve for frontier-class performance follows the same slope.
Stanford's 2025 AI Index quantified the collapse. The inference cost for GPT-3.5-level performance dropped from $20 per million tokens in November 2022 to $0.07 by October 2024. That's a 280-fold reduction in roughly 18 months.
Epoch AI's research across multiple benchmarks shows price-performance improving anywhere from 9x to 900x per year, depending on the task. On PhD-level science questions, the cost to match GPT-4's accuracy fell by 40x per year.
For context, this is faster than Moore's Law ever was. Faster than the bandwidth cost declines during the dotcom boom.
And every force driving this decline is accelerating, not decelerating. DeepSeek shattered pricing assumptions with inference 20-50x cheaper than OpenAI's equivalent models.
What does this mean for the arbitrage?
If your agent costs the equivalent of $20K/year to deliver today, the same capability will cost $4K to deliver next year and $500 the year after. Your competitor sees the same cost curve. They will price accordingly. And if they don't, someone else will.
Brian Roemmele's Preview of Where This Goes
If the macro data feels abstract, Brian Roemmele is making the endpoint concrete - right now.
Roemmele has been running what he calls a "Zero-Human Company" - an enterprise operated entirely by 100+ AI agents. No human employees. The CEO is Grok 4. The agents run on local hardware: Raspberry Pi boards, old laptops, commodity GPUs.
Using agent swarms - collaborative networks where AI instances specialize in subtasks and coordinate through shared protocols - Roemmele demonstrated a 97% task automation success rate across 240 diverse projects. These are the same categories of work (game development, architecture, data analysis, coding) that the widely-cited Remote Labor Index said frontier AI models could only handle at a 2.5% rate.
The cost? His Grok-4 CEO said the output would be worth it "even at $100 a month per employee." But the actual operating cost is far less. Pennies. Just the electricity your machine was already drawing while it sat idle.
The trajectory this reveals is the part that matters for pricing strategy: the floor for AI agent operating costs is converging toward the cost of electricity and commodity compute.
This is where every agent vendor's cost structure is headed within 24 months. And when it gets there, the arbitrage against human labor costs doesn't just shrink. It becomes irrelevant. Because you're no longer competing against the cost of a human. You're competing against another vendor whose cost floor is the same as yours.
The SaaS Pricing Parallel
We've seen this movie before. It's the exact arc SaaS followed, just on a compressed timeline.
In the early 2000s, the first cloud CRM could charge a premium because the alternative was Siebel on-prem - a multi-million dollar deployment. The pricing anchor was "cheaper than the old thing." And for a few years, that worked beautifully.
But within a cycle, the anchor shifted. Buyers stopped comparing Salesforce to Siebel and started comparing Salesforce to HubSpot, to Pipedrive, to Zoho. The relevant alternative wasn't the old expensive thing anymore. It was the new cheap thing. Margins compressed. Differentiation had to come from somewhere other than "cheaper than what you used to do."
Agents will follow the same arc, but radically faster. Because the underlying cost structure isn't falling 10-20% per year like SaaS infrastructure did. It's falling 10x per year. The compression that took SaaS a decade will happen to agents in 2-3 years.
Tomasz is describing Year One of this arc and calling it the steady state. It isn't. It's the peak of the arbitrage.
From Arbitrage to Dominance: What Actually Holds
None of this means the agent market isn't massive. It is. Tomasz is right that the labor-to-software shift is structural, that the TAM is enormous, and that the market rewards this transition. The opportunity is generational.
But the winners won't be the vendors who rode the human-parity pricing window and watched their margins evaporate when the arbitrage closed. They'll be the ones who used the window to build something that holds up after the cost floor drops.
Here's what I think actually converts the arbitrage into lasting dominance:
Re-value the human
This might sound counterintuitive in a piece about AI agents, but hear me out.
As long as humans are the ones buying, they will pay a premium for humans on the other side.
People want to buy from people who understand their context. They want support from someone who actually cares whether their problem gets solved. They want the judgment and expertise of someone who's been in their shoes. When everything is automated, the "right" human touch becomes the scarce resource - and scarcity is what supports pricing power.
The companies that will command premium pricing in the agent era will be the ones that bring the most empathetic humans, the most expert humans with discernment to the table. The most human humans. The most creative humans. The most heart-led humans. With AI, these humans are who people will want to do business with. As long as there are people on the other side.
Think about it: if every vendor's agent can do the same task at roughly the same cost, what differentiates? The human layer.
Agents commoditize execution. Humans hold pricing power on judgment, trust, and care.
The winning strategy isn't to eliminate the human from the equation. It's to make the human the premium tier of your offering.
Vertical specialization with infrastructure depth
Every generation of software has been built on horizontal layers. Cloud infrastructure. Databases. CRMs. Project management tools. The pattern repeats: build a horizontal platform, compete on features, watch margins compress as the category matures.
The next moat isn't horizontal. It's vertical - and it goes deeper than most founders realize.
The agent vendors that will sustain pricing power are the ones who stack unfair advantages that can't be replicated by a competitor spinning up the same open-source model on the same commodity hardware. That means owning the full vertical:
Your own fine-tuned version of open-source models, trained on proprietary data that no competitor has access to. Your own inference infrastructure - maybe even your own data center - so your cost floor is lower and your performance is tuned to your specific workloads. Your own frameworks and orchestration patterns, hardened through thousands of real-world deployments in your specific industry.
But infrastructure alone isn't enough. The moat gets real when you combine it with industry-recognized expertise and distribution. The agent vendor who is also the trusted name in compliance, or insurance underwriting, or construction project management. The one who doesn't just sell an agent - they bring the relationships, the channel partnerships, the integrations into the industry-specific platforms where buyers already live.
This is the stack that survives the arbitrage closing. Not "we have an AI agent that's cheaper than a person." That's a race to the bottom. But "we are the definitive AI-powered platform for [specific industry], built on proprietary data and infrastructure, backed by the deepest domain expertise in the market, and integrated into every workflow our customers depend on." That's a moat.
Horizontal agent vendors will compete on price and lose. Vertical agent vendors with real infrastructure depth will compete on irreplaceability and win.
The Bottom Line
The age of pricing agents like employees is a short-term arbitrage. Long-term dominance belongs to those who re-invest the arbitrage gains into two things: humans worth paying a premium for, and vertical depth too specialized to replicate.
Humans are your premium tier.