The Meter Is Always Running: What Consumption Pricing Does to Bad Architecture

There is no shortage of articles explaining how Agentforce Flex Credits work. They all say roughly the same thing: credits come in packs of 100,000 for $500, a standard action burns 20 of them, voice actions cost a little more, and the old flat $2-per-conversation model drew complaints for being unpredictable for smaller organizations. The rate card has been reprinted enough times to wallpaper a data center.

Almost none of them say the thing that actually matters.

When software was billed per seat, architecture quality was a quiet variable. A clumsy implementation cost the same per month as an elegant one. The license fee didn’t care whether your automation was a tidy, single-pass design or a tangle of redundant steps held together by hope. You paid for the chair, not for what the person in it did. Sloppiness was a one-time embarrassment, amortized into a fixed line item and forgotten.

Consumption pricing ends that arrangement. Salesforce has begun measuring work in Agentic Work Units, reporting 2.4 billion of them delivered across Agentforce and Slack and growing 57% quarter over quarter. Every action an agent takes is now a small, recurring withdrawal. And the uncomfortable corollary follows directly: every unnecessary action is a small, recurring withdrawal too and one that repeats for as long as the agent runs.

This is the part the rate-card explainers miss. Under metered pricing, architecture stops being an aesthetic preference and becomes a line on the invoice. Forever.

Where the bleed comes from

The credit drain rarely shows up where teams expect. It is not the headline use case that runs up the bill, it’s the design decisions underneath it.

A poorly structured knowledge base is the clearest example. A clean, well-organized knowledge source lets an agent resolve a question in a single retrieval action. A messy one forces multi-step retrieval, clarification loops, and re-queries with three or four actions to answer what should have taken one. The functional output looks identical to the user. The cost is three to four times higher, on every single interaction, indefinitely.

Prompt design behaves the same way. Bloated context, redundant variables, and vague instructions push token counts up and can multiply the cost of an action. An agent that has to be told the same thing four ways to get it right is not a working agent. It is a subscription to your own imprecision.

Then there is the orchestration layer with redundant callouts, the actions that fire when a guardrail should have stopped them, the flows that re-run logic already executed upstream. None of it breaks anything. All of it meters.

Efficiency is now a design discipline, not a cleanup task

The instinct in most organizations is to ship the agent that works and optimize later, the same way teams have always treated performance tuning. That instinct was affordable under per-seat licensing. It is expensive now, because “later” is measured in accumulated credit consumption, and the meter has been running the entire time you meant to get around to it.

The shift in question is not “can the agent complete the task.” It is “how many metered actions does the agent spend to complete the task, and is that number the lowest it can be.” Those are different questions, and the second one has to be answered at design time when the object model, the knowledge architecture, the action inventory, and the guardrails are being decided; not bolted on after the bill arrives.

This rewards a specific kind of build discipline: the habit of treating every action as if it has a price, because it does. It is, not coincidentally, the exact discipline that emerges when an organization spends years building products where every wasted operation is pure cost rather than a rounding error. Resource Interactive builds its AppExchange portfolio under precisely this constraint with apps shipped free, where there is no seat revenue to hide an inefficient architecture behind. When the meter is always running, you learn to design like it.

The takeaway nobody is selling

Consumption pricing has been framed almost entirely as a procurement question for which buying model, how many credits, what the three-year projection looks like. That framing is incomplete. The real consequence of metered AI is that it converts architecture quality from a soft preference into a hard, recurring cost.

The organizations that win on agentic CRM economics will not be the ones who negotiated the best credit rate. They will be the ones whose agents were designed, from the first whiteboard session, to do the work in the fewest possible actions. The rate card is the same for everyone.

What you spend against it is an architecture decision and now it is a permanent one.