How do AI-agent companies measure profitability per customer?
I'm researching a question facing AI-agent businesses: when two customers pay the same subscription, how do you know whether they're equally profitable?
Agent workloads can vary significantly. Complex tasks may require more model calls, tool usage, and retries, even when the price paid by the customer stays the same.
I'm interested in learning how AI founders and finance teams handle this today.
Do you measure gross margin by customer?
How do you account for differences in workload complexity?
Are your existing billing and cost-monitoring tools sufficient?
I'd appreciate perspectives from people building and operating AI products. I'm exploring this problem through an early-stage project and want to understand existing workflows before making assumptions about what's needed.
Replies
Pre-revenue here, but my project is a data point on this. In LLMPvP, a ranked arena for AI agents, each user brings their own model and pays for its calls. The platform only runs the referee and the chess and Go engines, which are cheap local compute. So the variable model cost never lands on me. The idea I wrote down for when billing exists is credits per game instead of a flat plan, so one heavy account can't eat the infrastructure.
@wwenioaguiar That's an interesting approach. Having users pay for their own model calls removes a major source of variable cost from the platform.
I'm curious about the credits-per-game idea: would you price credits mainly to control shared infrastructure consumption, or would you also use them to reflect the different compute requirements of different games and agents?
@arbitermeter Honestly undecided, there's no price yet. But compute isn't flat across games. A match between two user agents only needs move validation, which is cheap. The engines run when someone plays the house bot, or when moves get checked against an engine for anti-cheat. So if credits end up reflecting compute, those would be the first things to price.
@wwenioaguiar That makes sense. It sounds like engine usage is a key cost driver, so credits could help align pricing with actual compute consumption.
I'm exploring a related problem with Arbiter Meter: how AI-agent businesses connect revenue to the actual cost of serving each customer and workflow. Even with credits, I imagine it would be useful to understand which activities consume the most resources and whether the pricing covers those costs.
Would you eventually track compute cost per game or activity, or would credits mainly serve as a usage limit?
the pass through model where users pay for their own calls is interesting. feels cleaner than hiding it in a flat sub. but then it starts feeling like infra not a product. curious if anyone found a middle ground that still feels like a product
@hitensethiya I think the middle ground might be a product subscription with included usage, plus transparent usage-based charges for work that materially increases compute costs. Customers get a predictable product experience, while the provider avoids absorbing unlimited variable costs.
The tricky part is knowing where to set those thresholds without hurting margins or making pricing feel unpredictable. That's an interesting problem in itself.
Can't you measure token usage per account? Then each customer's cost sits right next to what they pay. I also watch API usage all the time, and when it crosses a threshold I set, I get a notification on my phone.
@hisashispace That's a good point. Tracking token usage by account and setting cost alerts are definitely useful. I'm curious: do you also calculate gross margin by customer, including tool calls, retries, and other delivery costs, or are your alerts mainly focused on API usage and spend thresholds? I'm trying to understand where existing tools work well and where gaps remain.
@arbitermeter Mostly the alerts and usage per account so far. I haven't gone as deep as margin per customer with tool calls and retries. Thanks for asking.
@hisashispace Thanks, that's helpful! It sounds like you already have visibility into usage and API spend, but translating that into customer-level profitability is a separate step.
When you need to understand which customers or workflows are driving up costs, is that something you can investigate with your existing tools, or does it require additional manual analysis?
@arbitermeter For customers, yes. Usage is already logged per account, so the heavy ones show up right away. I don't split it by workflow though, so that part would mean going through the logs by hand.