The Invoice Arrived and It Had My Name on It, (or How I Learned I Was the Product and the Customer Simultaneously)

I was reading about Salesforce's most recent earnings call when the sentence hit me. Their Chief Revenue Officer told analysts, flatly, that they had found the formula to monetize AI. Eight hundred million dollars in agent revenue in a single quarter, up from five hundred forty million the quarter before. Twenty-nine thousand deals. And a new unit of measurement, something Salesforce invented and named with the casual confidence of a company that knows it's setting the vocabulary: the Agentic Work Unit. One unit of AI work. A record updated. A workflow triggered. A decision made.¹

I work at a company that uses Salesforce. We have seats. We have licenses. We have the familiar renewal spreadsheet that procurement brings to the table every year, the one with department names and headcounts and per-user costs that everyone pretends to scrutinize but mostly just signs. And I imagine myself sitting there reading those numbers, I found myself doing something, which was imagining what that spreadsheet would look like next year.

I imagined the line items. The seats would still be there (they always are). But underneath them, or maybe beside them on a new tab nobody had been briefed on, there would be a second category. Not people. Not departments. Something else. Flex Credits consumed. Conversations handled. Actions executed. A meter running alongside the headcount, measuring work that used to belong to the people in the headcount, now being performed by something that doesn't have a name badge or a commute or a one-on-one with a manager on Tuesdays.²

And I thought: what would it feel like to be sitting in that renewal meeting, looking at a spreadsheet where your colleagues' work shows up as a line item priced by the unit?

Which is, if you stop and hold it up to the light, a question worth pausing over.

Because this is not a hypothetical for most enterprise companies anymore. It's the 2026 renewal season. Salesforce has its Flex Credits. Microsoft has layered a separate fifteen-dollar-per-user governance license on top of its existing Copilot seats, specifically for managing the agents acting on behalf of those users, with a consumption meter called Copilot Credits running underneath. ServiceNow has introduced what it calls Action Fabric, a consumption meter for automated operational work, with governed pathways that include identity, permissions, audit, and billing attached.³ Workday is positioning itself as the system of record for managing digital workers alongside human ones. SAP has drawn contractual boundaries around which AI systems are even allowed to call its APIs. Zendesk and HubSpot have moved toward outcome-based pricing, charging per automated resolution or completed prospecting task. Atlassian's Rovo credits are included in cloud subscriptions today, with overage billing metered but not yet charged, which is the softest possible way to say "the meter is on and the bill is coming."⁴

Each vendor has a different name for its meter. The underlying move is identical: every major enterprise software company is establishing a commercial control point around non-human work.

The renewal conversation is no longer "how many people need seats?" It is becoming something more uncomfortable: who or what is allowed to act through this system, how often, under whose identity, under which controls, and at what price? And embedded inside those questions is another one, the one nobody puts on the agenda: if the AI agent does the work, what happens to the person who used to?

Here is the situation, stated plainly. For about twenty years, the price of business software was attached to the number of humans who used it. Ten employees in the customer relationship management system meant ten seats purchased. Fifty support agents in the help desk meant fifty licenses. The math was readable. You could look at a renewal invoice and see your colleagues' names, or at least their department headcounts, reflected in the numbers. The human was the unit of software value.

That model has cracked. And the crack matters not because of what it means for software pricing (though it means a lot for software pricing) but because of what it reveals about how organizations think about the value of the people inside them.

The historian Theodore Porter wrote a book in 1995 called Trust in Numbers that I keep coming back to. Porter's argument, simplified somewhat unfairly, is that quantification does not naturally descend from the triumph of science.⁵ It ascends from political and bureaucratic necessity. We measure things not because measurement is inherently truthful but because measurement creates a portable kind of authority. A number can travel between departments, across organizations, up reporting chains, into boardrooms. A number does not require trust in the specific person who produced it, which is precisely why numbers become powerful in institutions where trust between individuals is thin or absent.

The per-seat software license was, in Porter's terms, a legible unit. It corresponded to something everyone understood: a person. The CFO could see the headcount. The manager could count the team. The auditor could verify the names. The commercial relationship between vendor and buyer was grounded in a unit both sides could point to and say, yes, that is a person, and that person uses this software, and therefore we pay this amount.

The new unit, the credit, the action, the agentic work unit, the automated resolution, is not grounded in a person. It is grounded in work. And work, as it turns out, is harder to make legible than a person sitting at a desk.⁶

James C. Scott, the political scientist who spent decades studying how institutions simplify complex realities to make them governable, coined a useful word for this process: legibility. His examples were forests and cities and farms, all cases where a government or institution needed to understand something complex and so imposed a grid over it, flattening local knowledge into standardized categories that could be read from the center.⁷ The grid always lost something. The Prussian foresters who replaced diverse forests with orderly rows of Norway spruce could count the trees precisely, but they killed the ecosystem that sustained the trees, and the forests eventually collapsed. The Soviet planners who collectivized agriculture could count the hectares but could not capture the local knowledge of the farmers who had been managing those hectares for generations.

Legibility has costs. The simpler the unit, the more it excludes.

A seat is legible. A person has a name, a department, a start date, a role. An agentic work unit is not legible in the same way. What did it do? Was it useful? Did the person who used to do that work agree it was done well? Did it succeed or just run? Was the output something a human would have produced, or something that only exists because the tool created the category? These are not rhetorical questions. They are the questions every finance team is about to start asking, and they do not have clean answers yet.

Here is what I think is actually happening underneath the pricing conversation, and I want to be careful how I say this because I am aware that writing about the economics of AI while being a person whose job involves AI tools carries a certain self-referential charge, the kind where you cannot quite tell whether you are observing the phenomenon or participating in it.⁸

The pricing shift is teaching organizations a new way to see their own people.

Consider what it means, practically, for a company to start tracking Agentic Work Units alongside human seat licenses. You now have two measures of work running side by side. One is counted in headcount and salary. The other is counted in credits and consumption. The AI's output is natively digital and natively countable. Every action logged, every credit burned, every resolution verified. It arrives in the system pre-measured.

Your output is not natively countable. The parts of your work that matter most are often the parts that resist quantification: the judgment call that prevented a bad decision, the relationship maintained through a difficult quarter, the question asked before the wrong thing was built, the context you carried in your head that kept three other people from going down the wrong path.⁹ These are real forms of value. They are also, in Porter's terms, not portable. They do not travel well across spreadsheets. They do not appear on renewal invoices. They are the kind of knowledge Scott would call metis, the practical wisdom that emerges from experience and refuses to be compressed into a standard unit.

Once an organization has a meter for AI work, the meter becomes a mirror. It reflects the work it can see. And the work it cannot see becomes, gradually, harder to argue for. Not because anyone decided it was worthless. Because the spreadsheet doesn't have a column for it.

This is where it gets personal, not in the abstract sense but in the Tuesday-morning sense, the sense where you are a person with a role and a salary and a set of things you do every day, and you start to notice that some of those things now have a price tag attached to them (the automated ones, the ones the AI handles) and some of them don't (the ones you do, the ones that involve judgment and context and presence).¹⁰

I work at a software company. My job, broadly, is to make sure my part of the organization has the right resources, systems, and workflows to function well. Until recently, the value of this work was measured in human terms: did the agents have what they needed, were the processes clear, did the tools reduce friction for the people using them. The unit of value was the supported human.

Now some of the work those humans did is being handled by AI agents. Not all of it. Not even most of it, yet. But enough that the conversation has shifted. And the new questions have a texture that is different from the old ones. Not just "are the people equipped?" but "which operations is the AI performing, and what are they costing per unit, and how does that compare to the cost of the person who used to do it?"¹¹ Nobody phrases it that bluntly in the meeting. But the spreadsheet phrases it that bluntly. The spreadsheet always does.

There is a concept in labor economics, less famous than it deserves to be, that the sociologist Ursula Huws has written about extensively. Huws studies what she calls the commodification of labor, the process by which specific, situated, human work gets abstracted into standardized, tradeable units.¹² A customer support interaction, which in practice involves empathy, judgment, context, institutional memory, and a hundred small decisions, becomes a "case resolved." A financial reconciliation, which requires understanding of exception patterns and regulatory context, becomes an "item processed." The abstraction is necessary for management and measurement. But the abstraction also makes the work look simple, and things that look simple look replaceable.

AI accelerates this process dramatically. Not because AI actually performs the work the same way a human does (it does not), but because AI produces outputs that can be measured in the same units. If you measure support success by cases resolved per hour, an AI agent that resolves cases can be directly compared to a human agent on that metric. The comparison is misleading, because the AI and the human are doing different things to arrive at that output, but the metric does not care. The metric sees two numbers. One is cheaper.

AI and the human are doing different things to arrive at that output, but the metric does not care. The metric sees two numbers. One is cheaper.

This is what I mean about the invoice having your name on it. Not literally. Your name is still on the seat license (for now). But the new line items, the credits, the units, the consumption meters, they represent a bet the vendor is making about where the value lives. And the bet is not on you.¹³

I want to pause here and say something that feels important: none of this is necessarily bad.

The old model, in which every piece of enterprise software charged for every human who touched it regardless of how much value that person extracted, was not particularly fair either. Companies paid for hundreds of CRM seats when only a fraction of those users logged in with any regularity. They paid full price for light users and heavy users alike. The seat was a blunt instrument, and the vendors knew it, and the buyers accepted it because the alternative, measuring actual usage, was too complicated.

The new model, in which vendors charge for work completed, has a version that is genuinely better. If Zendesk charges per automated resolution and the resolution actually solves the customer's problem, the buyer pays for outcomes, not inputs. If Salesforce charges for an AI agent that keeps records updated, the buyer can potentially reduce seats for employees who no longer need to sit in the CRM all day. If the meter is honest, if failed work doesn't count the same as completed work, if the rate card holds for the term, if the buyer can set caps, if the data exports cleanly, then this is a reasonable commercial arrangement.

IF.

That "if" is doing a lot of work.¹⁴

Because the version that should worry people is the one where the vendor captures the efficiency. Where you still pay for every human seat and you pay for every agent action on top. Where the AI reduces the labor but the software bill stays the same or grows. Where the meter is hidden until renewal. Where credits expire unused while overages bill immediately. Where the vendor's own agent is the only practical route while outside agents are treated as hostile. Where commercial lock-in arrives wearing the language of security and governance.

The question that tells you the most about which version you're dealing with is the one most vendors will try hardest to avoid at renewal: if the AI agent resolves a large share of the work, can we reduce seats? Can occasional users move to lighter access tiers? Does our cost actually go down when the work gets cheaper?

Sometimes the answer is legitimately no. Compliance requires named users. Oversight requires seats. Approvals require human licenses. But the question has to be asked. Otherwise you end up with the worst hybrid: old headcount costs plus new agent consumption, and nobody in finance can say whether the company is paying less per unit of work or just paying more for the same work measured two different ways.¹⁵

Lewis Mumford, writing about the clock in the fourteenth century, argued that the mechanical timepiece did more to transform labor than the steam engine.¹⁶ The clock did not perform work. It measured work. It created a universal, abstract unit (the hour) that could be applied to any activity, and once you had a universal unit, you could compare, optimize, and price activities that had previously been incommensurable. A cobbler's afternoon and a clerk's afternoon became the same thing: three hours. The specificity of the work, the craft knowledge, the situational judgment, was invisible inside the unit.

The credit is the new clock.

An Agentic Work Unit, a Flex Credit, a Copilot Credit, an automated resolution: each is an abstraction that makes different kinds of work commensurable. An AI that summarizes a support case and an AI that triggers an onboarding workflow both consume credits. A human who does those same tasks does not consume credits (she consumes a salary, which is measured differently and accounted for by a different department). The credit creates a new axis of comparison between human and machine work, and the axis is built for the machine, because the machine's output is natively digital and natively countable.

And so the shift we're inside is not simply a pricing change. It is a legibility change. The things that can be counted are being counted, loudly, on invoices and dashboards and quarterly earnings calls. The things that cannot are becoming, gradually, quieter.

I keep coming back to the imagined renewal meeting. The procurement lead has the spreadsheet. The seat licenses are on page one, same as always. The new tab, the one with the credits, is on page two. And the question forming in the back of someone's mind, the question that will reshape how this company thinks about its own people, is not "how many seats do we need?" It is "how much of this work can move to the meter, and what happens to the seats when it does?"¹⁷

Nobody will say it that way. They'll say "optimization" or "efficiency" or "right-sizing the license mix." They'll frame it as a procurement exercise, a software cost question, a vendor management initiative. And it is all of those things. But it is also, underneath the spreadsheet, a question about what kind of work a person is for, and whether the answer to that question changes when a machine can produce the same output for less.

C. Thi Nguyen, the philosopher who studies gamification, has written about what he calls value capture: the process by which a rich, complicated value gets replaced by a simplified metric, and then the metric becomes the thing people optimize for, and then the original value is forgotten.¹⁸ A university cares about education; it adopts rankings; it optimizes for rankings; it forgets what education was for. A hospital cares about patient wellbeing; it adopts satisfaction scores; it optimizes for scores; the scores become the reality. The danger is not that the metric is imposed from outside. The danger is that people internalize it. They start to see themselves through the metric's eyes.

A university cares about education; it adopts rankings; it optimizes for rankings; it forgets what education was for.

The same process is beginning with AI work metrics, at the level of individual workers, inside organizations that have not yet noticed it happening. The person who can describe the value of their work in meter-compatible terms, in units and outputs and resolutions, will be legible to the budget. The person whose value lives in judgment, context, institutional memory, and the kind of thinking that prevents problems rather than resolving them, will be harder to see. Not fired, necessarily. Just gradually unaccounted for.

I don't have a tidy resolution for this, and I am suspicious of people who do. The honest position is that we are early in a transition whose shape is not yet clear, and the most useful thing anyone can do right now is pay attention to what is being measured, by whom, and what the measurement leaves out. Start asking what your organization's AI tools are actually costing per unit of work. Ask whether the meter counts failures the same as successes. Ask who controls the definition of "resolved." Ask what happens to your seat licenses when agent usage scales. Ask these things before the answers are baked into a contract you'll live with for three years.

And maybe also ask yourself a stranger, more personal question, the kind you can't put on a spreadsheet but that might matter more than anything on it: what do you do that no meter could capture, and does your organization have a way of knowing it exists?¹⁹

I went back to reading the earnings call transcript after I'd sat with the imagined spreadsheet for a while. Salesforce reported 2.4 billion Agentic Work Units delivered to date. That's 2.4 billion records updated, workflows triggered, decisions made, all counted, all metered, all billed. Somewhere in the system, something is counting.

The question is whether you're counting too.

¹ I should note that "found the formula to monetize AI" is the kind of sentence that sounds like a victory if you're a Salesforce shareholder and like a weather warning if you're a Salesforce customer. The formula, in this case, means the vendor has figured out how to charge for work that used to be included in the seat price. "Found the formula" is a polite way of saying "found the meter." ↩︎

² This imagined spreadsheet is not entirely imaginary. Several companies I've read about in the past few months have described seeing exactly this kind of split emerge in their renewals: the familiar seat-based tab plus a new consumption-based tab, sometimes presented together, sometimes introduced mid-conversation by the account executive as an "add-on" or "acceleration package." The vocabulary varies. The structure doesn't. ↩︎

³ ServiceNow's framing is worth sitting with for a moment. They're positioning themselves as the place where agents trigger real operational work: provisioning access, escalating incidents, opening change requests, routing approvals, executing playbooks. Each of these is a unit of work that used to belong to a person. The consumption meter is attached to the governed pathway, which means the act of using the proper, auditable, secure route is also the act of generating a bill. Security and billing arrive together. ↩︎

⁴ Atlassian's approach is probably the most honest about the trajectory. They're saying, in effect: we're going to meter your usage now, we're not going to charge you for it yet, and we'll give you notice before we start. Which is generous. It is also how every consumption model enters the building. The feature ships included, usage gets tracked, teams build dependencies, and one renewal later the credit conversation arrives. ↩︎

⁵ Porter is a historian of science at UCLA, not a business writer, which is part of why his work hasn't traveled as far into management circles as it deserves to. The core argument of Trust in Numbers is that quantification becomes most powerful where trust is thinnest. Weak institutions lean on numbers because numbers seem to require no personal judgment. Which is, of course, its own form of judgment. ↩︎

⁶ There is an irony here that I find almost architectural. The seat license was imprecise. It charged the same for someone who logged in eight hours a day and someone who logged in once a month. But its imprecision had a kind of democratic quality: everyone was equally legible, equally priced, equally counted. The credit meter is more precise but less equal. Work performed at machine speed by an AI agent is natively countable. Work performed at human speed, with pauses for thought, questions to colleagues, and judgment calls that don't register in the telemetry, is natively uncountable. Precision rewards the countable. ↩︎

⁷ Scott's most famous example is the Prussian Normalbaum, the scientific forestry program that replaced diverse old-growth forests with monoculture plantations. The new forests were spectacularly legible: you could count every tree, predict every harvest, optimize every row. They were also ecologically dead within a generation, because the diversity they replaced was doing work the planners couldn't see. ↩︎

⁸ A friend of mine, a therapist, once told me that the hardest clients to work with are other therapists, because they know all the frameworks and can narrate their own patterns with such precision that the narration itself becomes a way of avoiding the change. I think about this in the context of people who work in technology writing about how technology is changing work. Present company very much included. ↩︎

⁹ I once spent three hours in a meeting that produced no measurable output, no ticket, no document, no artifact, but which resulted in a single question being asked that prevented my team from building the wrong thing for six months. In any metering system I'm aware of, that meeting would register as three hours of zero productivity. In reality, it was probably the most valuable three hours of my quarter. This is the kind of value that measurement systems are structurally unable to see. ↩︎

¹⁰ This split has a specific emotional texture that I think is underexplored. The automated work gets celebrated: look how many cases the AI resolved, look at the throughput, look at the deflection rate. The human work gets examined: is this role still needed at this level, could this person move to a lighter license tier, is this function a candidate for agent augmentation? Celebration for the machine. Scrutiny for the person. The asymmetry is not intentional, but it is real. ↩︎

¹¹ If you're a builder deploying agents that interact with enterprise platforms, there's a version of these questions aimed directly at you, too: which of your agent's operations consume credits, and do reads and writes cost the same? Does a failed action bill the same as a completed one? Is your agent using the vendor's endorsed pathway, or an API route that might get restricted at the next contract review? These aren't engineering questions. They're architecture questions disguised as licensing questions, and they'll determine whether your agent survives its first encounter with procurement. ↩︎

¹² Huws was writing primarily about the gig economy and digital platform labor, but her analysis of how commodification works, how specific situated work gets abstracted into interchangeable units, applies with uncomfortable precision to the current moment. The AI agent is not a gig worker. But the unit economics have a family resemblance. ↩︎

¹³ I want to be fair. Most vendors aren't scheming about how to devalue their customers' employees. They're solving a legitimate business problem: if AI agents can do things that previously required human users, and the vendor only charges for human users, the vendor's revenue falls while its compute costs rise. The credit meter is a survival mechanism. But survival mechanisms don't always care whose survival they're threatening. ↩︎

¹⁴ Nine traits tend to separate a fair meter from a captured one. The meter is visible and the unit makes sense. The customer can forecast usage. Failed work isn't billed identically to completed work. Third-party agents get a governed path rather than a blocked one. The vendor distinguishes between reading, drafting, recommending, writing, approving, and executing. The buyer can set caps. Usage data exports cleanly. The rate card holds for the term. And the model aligns with the value created. If fewer than half of these are true about your next renewal, you're looking at a captured meter dressed as innovation. ↩︎

¹⁵ This is the practical CFO risk, by the way. AI can make work cheaper while making software bills more complicated. If finance doesn't model the work, it will only see the line items. The vendor will show productivity. The business owner will show adoption. Procurement will see credits. The budget owner will see overages. Nobody will be able to say whether the company is actually paying less per unit of completed work. ↩︎

¹⁶ Mumford's argument, from Technics and Civilization (1934), is that the clock was the key machine of the industrial age, not the steam engine. The clock made it possible to think about time as something that could be divided, allocated, bought, and sold independent of the work being done. Once you could sell an hour of labor as a standardized unit, the nature of the labor inside that hour became secondary to its duration. The credit does the same thing with work units. ↩︎

¹⁷ The vendor knows this question is coming. That's why the best vendor strategy is to make the AI sticky before the buyer has data on what it costs. Get the agent into the support queue, into the finance close, into the recruiting workflow. Let the work move. Once the work has moved, the negotiating position flips. The vendor knows the work is there. The buyer knows turning it off would hurt. This isn't conspiracy. It's commerce. But knowing it's commerce doesn't make it less effective. ↩︎

¹⁸ Nguyen's work is mostly about games and how game structures infiltrate non-game contexts, but the concept of value capture has implications well beyond gaming. The relevant insight is that metrics don't just measure values; they replace them. The map becomes the territory. Once an organization starts counting AI credits alongside human work, the credit becomes the shared language, and the work that doesn't translate into credits becomes, in the organization's self-understanding, work that doesn't quite exist. ↩︎

¹⁹ This is not advice, exactly. I distrust advice from people who are living inside the same confusion they're advising about. It's more of an observation about the asymmetry of the situation: the systems that measure work are being built right now, by vendors and procurement teams and finance departments, and the people whose work will be measured are mostly not in the room where the measurement gets designed. ↩︎