Agents Are Now Priced in 'Work Units,' Not Seats — Here's How to Budget for Them Before the 2027 Cull

Sometime in the next few quarters, your finance team is going to ask why the agent line item moved from a predictable per-seat fee to a number that swings with how much work your bots actually did. That is not a bug. It is the pricing model arriving. Salesforce just reported 7 billion Agentic Work Units delivered, and it is the clearest signal yet that the unit of cost for AI is shifting from user to work. If you are still budgeting for agents like software licenses, you are about to be surprised, and the surprise is avoidable.

What is a work unit, exactly?

An Agentic Work Unit, in Salesforce's definition, is one discrete task accomplished by an AI agent. A prompt processed, a reasoning chain completed, or, the one that matters most, a tool invoked. It is the moment raw intelligence turns into something that changed a record, sent an email, or resolved a ticket.

The practical translation is even simpler: you used to pay for the person with the login, now you pay for the things the agent did. Salesforce has worked through a couple of pricing shapes to get here. It launched Agentforce at two dollars per conversation, where a conversation covered the whole back-and-forth including reasoning and tool calls, then began shifting toward action-based pricing, with enterprise rates reported at fractions of a cent per action at scale. OpenAI prices by token. Different currencies, same direction: the meter is now on work performed, not people licensed.

Why is this happening now?

Because long-horizon agents broke the seat model. A seat assumes a person, or a bot standing in for a person, who does roughly the same amount of work each month. That assumption held while agents were assistants that finished a task in one conversation. It collapses the moment an agent can pursue a goal for a week. That agent might do ten actions one day and ten thousand the next, and no per-seat fee can price both honestly.

So the vendors are doing the rational thing: they are pricing the thing that scales, which is the work, and they are reporting the totals to make it feel normal. Seven billion work units is a number designed to get CFOs comfortable with a new unit of account before anyone has to defend it in a budget meeting.

How do I budget for something I cannot see yet?

This is the part ops and finance leaders keep asking me, so here is the honest answer. You budget for agents the way you budget for cloud compute, not the way you budget for SaaS seats. You start with a baseline, you model scenarios, and you cap the downside.

Step one is a baseline. Run your agents in a quiet month and record how many work units, actions or tokens a real process actually consumes. You need a per-process number, not a per-team number, because a support agent and an outbound sales agent have wildly different work profiles.

Step two is scenarios. Budget a low, base and high case, with the high case built on a spike assumption, because agent workloads spike when a marketing campaign lands or a season hits, and a budget built on a quiet month will not survive a loud one.

Step three is the cap. Put a hard monthly ceiling on spend before you need it, not after the invoice arrives. Every serious agent platform has a way to throttle or alert on usage, and you want that configured on day one.

How do I forecast when the unit is work?

Forecast from the process up, not the platform down. The old way was to multiply headcount by a per-seat price and call it a year. The new way is to estimate the volume of work per process, multiply by a cost per unit, and sum it.

That means you need two numbers you probably do not have yet: the volume of completed tasks per process per month, and the cost per unit at your negotiated tier. Get the first from your operations data. Get the second from the vendor, and push for volume discounts early, because list prices for work units drop sharply with scale, and the rate you lock now is the rate your budget defends for the next year.

The uncomfortable truth is that your first forecast will be wrong. Treat it as a range, revisit it monthly, and do not let finance treat a consumption-based line item like a fixed contract. That mismatch is where the nasty surprises come from.

How do I measure ROI without fooling myself?

This is where the 40 percent trap lives. Gartner has forecast that a large share of AI projects will be cancelled by the end of 2026, and a big reason is that teams measure the wrong thing. When you pay per unit of work, it is tempting to report units of work as the win. Seven billion units delivered sounds great until you ask how many the business would have paid a human to do, and how many it accepted without rework.

So measure acceptance, not activity. The metric that matters is completed work that survived human review, because an agent that did a thousand tasks badly did not save you money, it created a cleanup job. Count the units that got accepted, cost them against the human time they replaced, and be ruthless about the rework rate. That is the only ROI number that survives a finance review.

What should finance do before the 2027 cull?

Three things, and I would do them in this order. First, reclassify agent spend as consumption, not software, in your budget model, so the CFO is not surprised when it moves. Second, negotiate unit pricing and caps now, while the vendors are still competing for your volume, rather than later when the model has hardened. Third, put an acceptance and rework metric on every agent in production before you scale it, so the cull, when it comes, cuts the projects that never proved value instead of the ones that did.

The takeaway

The shift from seats to work units is not something that happens to you, it is something you can plan for. Baseline per process, forecast in ranges, cap the downside, and measure acceptance instead of activity. Do that and you will be the person in the room who saw the invoice coming. Skip it and you will be the one explaining in 2027 why the agent line item tripled in a month nobody could see coming.