Cost per Token
Business & Governance
Cost per token is what an AI provider charges for each token (a small chunk of text or other data) a model reads or writes, usually quoted per million tokens.
Business & Governance
What Are Unit Economics in AI?
Unit economics is the revenue and cost tied to a single unit of business, such as one customer, one order, one support conversation, or one document processed. Instead of looking only at total spend, you ask what each unit costs and earns. The term comes up frequently around AI because AI costs often grow with usage, so the cost of each task, not just the monthly bill, shapes whether an AI tool is worth it.
Also called AI unit economics
Example
A dealership group with several locations is weighing an AI service that answers website and text inquiries at any hour, asks qualifying questions about budget, trade-in, and timing, and passes promising leads to salespeople. The vendor charges per conversation, plus a setup fee to connect the service to the group’s lead-tracking system.
The general manager picks one unit: a qualified lead handed to sales. Today, that cost is mostly staff time spent on calls and follow-ups. With AI, it is the conversation fees (including chats that never qualify), the setup fee spread over a year of expected volume, and the minutes staff still spend reviewing each handoff. Using made-up round numbers, if today’s cost is $20 per qualified lead and the AI version comes to $12, the service looks promising, as long as lead quality holds up.
Why unit economics matters more with AI
Much business software is sold as a flat subscription, often priced per user (sometimes called per seat). Within the plan’s limits, an extra report or an extra customer email usually costs nothing more. Many AI tools work differently. Each request runs the model. Providers that sell direct access to AI models typically bill by usage, measured in tokens, and many AI products built on those models pass that cost on through per-conversation, per-task, or credit-based pricing. How token pricing works is covered under cost per token.
The result is that AI costs tend to grow with volume. On usage-based pricing, doubling the number of conversations an AI assistant handles can roughly double what you pay. Some tasks also cost far more than others: a long contract review, a multi-step agent task, or a request that gets retried several times uses much more processing than a short reply. That is why the useful question shifts from “What does this tool cost per month?” to “What does it cost to complete one task, and is that less than the task costs us now?”
Even AI products sold on a flat plan often come with usage limits, credits, or overage charges, so volume still matters.
How to compare cost per task
Start by choosing a unit that matches the work, such as one invoice processed, one ticket resolved, or one qualified lead. Then work out two numbers.
- Cost per unit today. Staff time per task multiplied by the full hourly cost of the people doing it (pay plus benefits and overhead), plus any tools or outside services the task uses.
- Cost per unit with AI. A month of usage, subscription, and maintenance fees, plus a monthly share of the one-time setup and integration cost, all divided by the tasks completed that month. Then add the time people still spend checking, correcting, or finishing each task.
As a rough formula:
cost per task with AI ≈ (monthly fees and upkeep + monthly share of setup) ÷ tasks per month + review hours per task × hourly staff cost
Do not skip the review time. Many business uses keep some form of human review, and those minutes are a real cost that is easy to leave out. Finally, compare quality, not just cost: a cheaper unit that produces more errors or weaker outcomes may not be cheaper at all.
Common mistakes to avoid
- Counting only the usage fee. Model charges can be small next to setup, integration, maintenance, and review time.
- Assuming optimistic volume. Setup costs spread over a large forecast look small. If real volume comes in lower, the cost per task rises.
- Measuring the wrong unit. Cost per conversation can look low while cost per useful outcome, such as a resolved ticket or a booked appointment, is much higher, because not every conversation succeeds.
- Comparing unequal pictures. Putting full staff costs (pay, benefits, and overhead) on one side and only software fees on the other makes AI look better than it is.
- Ignoring rework. Errors that slip through and need fixing later belong in the cost per task.
- Forgetting the revenue side. Unit economics weighs what a unit earns against what it costs. If AI changes the value of each unit, such as how often a lead turns into a sale, count that too.
Why people debate AI companies’ margins
The same idea applies to AI providers themselves, and it sits behind much of the commentary you may see in the news or on LinkedIn. For a traditional software company, serving one more customer usually costs little, which is a big reason software businesses have long been known for high gross margins. For AI companies, every request requires inference, the work a model does to produce each answer, and that work runs on expensive computing hardware. So each customer, and each heavy user in particular, carries a real cost to serve.
Views differ on how this plays out. Some argue that falling computing costs and more efficient models will improve margins over time. Others point out that competition keeps pushing prices down, that heavy users on flat-price plans can cost more to serve than they pay, and that it is unclear how durable any company’s advantage, or moat, will be. The question is still open.
For a business buying AI, the practical takeaway is that vendor pricing may shift as providers adjust their own economics, through price changes, usage caps, or new plan tiers. It is sensible to build that possibility into your cost-per-task math.
Further reading
- AI ImplementationRelevant if you are redesigning a process, connecting systems, and deploying AI, where cost per task becomes a practical question.
Before you buy
Decide where AI fits before you commit.
SN3 Advisory helps leadership teams decide where AI will create the most leverage, before resources are committed to technology.