This framing maps perfectly onto ServiceNow. They're trying to sell the work (AI agents resolving tickets) instead of selling the seat. Problem is the cost structure flips, seats had zero marginal cost while each AI "Assist" has a real inference cost waterfall through model providers and hyperscaler GPU. The CFO guided gross margins down 200bps and said they'd offset it with opex cuts. You can sell the work, but the work has a cost that seats never did.
Good call out Jose. Maybe worth a larger case study on ServiceNow? I haven't written a real breakdown of a company like them for this publication...anything you would want to see covered or questions you would like answered?
Yes, it would be interesting to se how different scenarios may play out for them. From AI agent utilization exploding to revenue per seat compression. Difficult to know how it plays out, I guess that is why the stock has been selling off recently
The 'opacity of production' section is the crux of this whole argument and I think it's underappreciated. McKinsey's pricing power was never really about the quality of the PowerPoint... it was about the black box around how it got made. AI doesn't get that black box. The buyer can see the seams.
Your insight about 'work' becoming hard to price aligns perfectly with what I'm exploring—the precision paradox. We articulate AI tasks with extreme clarity (detailed prompts), yet when delegating human work, we're vague. This ambiguity makes outcomes unmeasurable. Clear communication frameworks matter whether you're briefing an AI or building a pricing model for work. Check it out: https://creatism.substack.com/p/we-prompt-machines-better-than-we?r=177ve
The strongest point in this essay is that “work” becomes hard to price once production is inspectable. The next question is:
In non-binary decisions, what is the buyer actually purchasing? Output, or permission to act on output?
If quality can’t be fully contracted, then value shifts to decision infrastructure: policy constraints, exception handling, evidence traceability, and accountable sign-off when the model is wrong. That seems like the real moat, not outcome markup.
Evan, I concur with your assessment of the fuzzy economics of "selling the work."
You’re right that it fails when treated as a simple pricing mask on top of transparent token costs; if the factory is visible, the margin is always negotiable.
However, the flaw in your logic is the absence of runtime economic observability: the inline capability of correlating the cost of each agent interaction and result to business value-added.
Using available technology for any of four and counting providers for runtime economic observability for each agentic outcome, your argument collapses.
This infrastructure enables the fundamental transition from "software as a tool" (priced by the seat) to "software as talent" (an autonomous labor force).
The "support exception" you noted isn't an outlier—it’s a preview. It works because those outcomes are measurable today. The failure elsewhere isn't a flaw in outcome-based logic, but the lack of economic infrastructure to govern it.
It was an epistemic mistake to conflate deterministic work with probabilistic resolution (agents at work) and the judgment—if not "sagacity"—of a McKinsey, a firm currently in the process of monetizing agent-based deliverables (ask me how).
You cannot realistically price a "seat" for an agent that performs 10,000 tasks an hour. Nor can your arbitrage tokens, API and MCP calls, reinforcing learning, etc.
We don't need a retreat to seats; we need a System of Record that provides the runtime observability required to link non-deterministic AI activity to verifiable P&L value.
Loved this analysis Evan - in a way it feels like it's going full circle back to raw value and it will become harder and harder for most, bar a few companies, to sell seat-based subscriptions
This framing maps perfectly onto ServiceNow. They're trying to sell the work (AI agents resolving tickets) instead of selling the seat. Problem is the cost structure flips, seats had zero marginal cost while each AI "Assist" has a real inference cost waterfall through model providers and hyperscaler GPU. The CFO guided gross margins down 200bps and said they'd offset it with opex cuts. You can sell the work, but the work has a cost that seats never did.
Good call out Jose. Maybe worth a larger case study on ServiceNow? I haven't written a real breakdown of a company like them for this publication...anything you would want to see covered or questions you would like answered?
Yes, it would be interesting to se how different scenarios may play out for them. From AI agent utilization exploding to revenue per seat compression. Difficult to know how it plays out, I guess that is why the stock has been selling off recently
The 'opacity of production' section is the crux of this whole argument and I think it's underappreciated. McKinsey's pricing power was never really about the quality of the PowerPoint... it was about the black box around how it got made. AI doesn't get that black box. The buyer can see the seams.
Great article!
Your insight about 'work' becoming hard to price aligns perfectly with what I'm exploring—the precision paradox. We articulate AI tasks with extreme clarity (detailed prompts), yet when delegating human work, we're vague. This ambiguity makes outcomes unmeasurable. Clear communication frameworks matter whether you're briefing an AI or building a pricing model for work. Check it out: https://creatism.substack.com/p/we-prompt-machines-better-than-we?r=177ve
The strongest point in this essay is that “work” becomes hard to price once production is inspectable. The next question is:
In non-binary decisions, what is the buyer actually purchasing? Output, or permission to act on output?
If quality can’t be fully contracted, then value shifts to decision infrastructure: policy constraints, exception handling, evidence traceability, and accountable sign-off when the model is wrong. That seems like the real moat, not outcome markup.
Example of that framing in practice: https://devpost.com/software/decision-kernel
Evan, I concur with your assessment of the fuzzy economics of "selling the work."
You’re right that it fails when treated as a simple pricing mask on top of transparent token costs; if the factory is visible, the margin is always negotiable.
However, the flaw in your logic is the absence of runtime economic observability: the inline capability of correlating the cost of each agent interaction and result to business value-added.
Using available technology for any of four and counting providers for runtime economic observability for each agentic outcome, your argument collapses.
This infrastructure enables the fundamental transition from "software as a tool" (priced by the seat) to "software as talent" (an autonomous labor force).
The "support exception" you noted isn't an outlier—it’s a preview. It works because those outcomes are measurable today. The failure elsewhere isn't a flaw in outcome-based logic, but the lack of economic infrastructure to govern it.
It was an epistemic mistake to conflate deterministic work with probabilistic resolution (agents at work) and the judgment—if not "sagacity"—of a McKinsey, a firm currently in the process of monetizing agent-based deliverables (ask me how).
You cannot realistically price a "seat" for an agent that performs 10,000 tasks an hour. Nor can your arbitrage tokens, API and MCP calls, reinforcing learning, etc.
We don't need a retreat to seats; we need a System of Record that provides the runtime observability required to link non-deterministic AI activity to verifiable P&L value.
Loved this analysis Evan - in a way it feels like it's going full circle back to raw value and it will become harder and harder for most, bar a few companies, to sell seat-based subscriptions