The price on an AI model’s pricing page is the cost of one attempt. It is not the cost of the work. A cheaper model that fails more often, needs retries or sends more output to a human reviewer can cost more per finished task than a model with a higher list price. Most AI business cases skip that step, which is why so many show savings on paper and a larger bill in practice.
This page brings together Circuit & Capital’s coverage of AI costs for the people who approve the spend. The guides explain how to tell AI spend from AI return, how to put a cost on one workflow before scaling it, and how to write a one-page rule for who may use which tools. The briefings look at the market around that spend: Gartner’s forecasts for AI spending and infrastructure impairment, vendor speed claims and whose number they are, and the unit economics of companies built on models they rent.
Two free calculators do the arithmetic. AI model cost per finished task compares models on attempt cost, retries, success rate and review time, and finds the daily token volume at which self-hosting costs less than an API. Token cost to gross margin starts from what you charge per task and shows how much token and review cost leave as contribution. Both now come with a free spreadsheet download with the formulas visible, so finance can check the logic and plug in its own numbers.
Agents add a second kind of cost: what they can touch. An agent that inherits a person’s access, has no spend cap and no named owner is a liability before it is a saving. The agent inventory and agent access limit sheet record what each agent may read, write, call and spend, and who answers for it. Regulators have noticed. The SEC’s fiscal 2027 examination priorities say examiners will review how registrants monitor the AI tools they use.
The rule across all of it is simple. Use your own volumes and prices. Treat a vendor benchmark as the vendor’s claim until you have run the test yourself. And count human review time, because it is often the largest line.
Who this page is for, and how to use it
This page is for the people who approve AI spending: finance leads, operations owners and founders. If you are new to the subject, read the spend-versus-return guide first, then pick one workflow and price it with the 90-day method before scaling anything. Use the calculators with your own volumes and the vendor’s current price page, not our examples.
Guides
- How to tell AI spend from AI return
- Put an AI cost on one workflow
- A one-page AI use rule
- Limit what an agent can touch
- Extra questions for an AI bill of materials
- What Article 50 asks a buyer
Tools and free downloads
- AI model cost per finished task · free Excel calculator
- Token cost to gross margin · free Excel calculator
- AI agent inventory and access map · free CSV template
- Agent access limit sheet · free CSV template
- Vendor claim checker · free checklist
Briefings
- Gartner says the AI build-out may not pay back · October 21, 2026
- Gartner revised the AI bill. Most of it is not your budget.
- Astra Ultrafast is a speed claim. The bill is still yours.
- AI spend is not the same as AI return
- The new unit economics of AI-native startups
- The AI productivity paradox at mid-market companies
- NVIDIA put the fence outside the agent
- The SEC’s 2027 exam list names deepfakes and vendor oversight
Other topics: Renewals and SaaS spend · Security and vulnerabilities · All topics