- Services
- Agency
- About us
Platform
Mustang
HR
Live
The Workforce
Procurement
Live
The Controller
Coming Soon
In Build
Growth & Customer Experience
Marketing
How to Calculate AI Workflow Cost Before You Scale It
Share:
How to Calculate AI Workflow Cost Before You Scale It
Publish date
Share:
Publish date
Share:
Everyone is doing some version of the same exercise right now. We are putting AI into workflows, automating work that used to sit with people, and sooner or later somebody asks the question that decides whether the whole thing keeps moving: is this actually worth it?
On paper, the economics can look almost absurdly good. A model call costs cents. A task that used to take a person an hour can run in minutes. The spreadsheet practically sells the project for you.
Then production gets involved and the number starts moving. Context gets loaded again. A run repeats. A browser or business system adds its own cost. The awkward case still lands on a person. Different teams start looking at different versions of what the workflow costs.
We have been on both sides of that conversation, building the workflow and then explaining why the savings on the slide are not showing up exactly where expected. We have built the spreadsheet, then watched production rewrite it.
That scar changed how we do the math. We start with the unit of work the operation already understands: one support case closed, one vendor review completed, one incident resolved, one onboarding case completed, one QA cycle closed. Then we ask what it cost to get that unit all the way through successfully.
The model underneath can be GPT, Claude, Gemini, or whatever gets released next month. The unit cost should survive the swap.
And small misses stop being small at volume. In the worked example below, a $0.60 human exception line becomes $15,000 a month at 25,000 cases.
Why Model Spend Understates the Cost of a Production AI Workflow
The model invoice covers the AI call. The operation pays for more.
Those 4 costs rarely live in the same place. That is where the ambiguity comes from.
Use the unit your operation already counts, then assign every cost in the workflow to that unit. The math gets much easier once everybody is talking about the same denominator.
Worked Example: What It Costs to Resolve “I Was Charged Twice” at Scale
Take a billing support case almost anyone has seen. A customer writes in because the same purchase appears twice.
The workflow reads the message and account history, checks recent charges, checks the refund rules, and either resolves the case under those rules or sends the exception to a person.
To make the model line concrete, give that case 8,000 input tokens and 1,000 output tokens. At current standard rates on September 24, 2026, both GPT 6 Sol and Claude Sonnet 5 price that token volume at $0.026. We keep both names here on purpose. Swap the provider and only this line of the worksheet should change.
Now add the operating reality. If 10% of cases need 6 minutes from a person at an internal cost of $60 an hour, the expected human exception cost is $0.60 per case.
At 25,000 cases a month, the model line is about $650. The expected human exception line is $15,000. Together they are already $15,650 before any paid tool or retry cost.
The model is around 4% of that combined number.
This is why a cheap model call can produce a bad budget. The question is not whether $0.026 is cheap. It is whether you built the business case around $0.026 while the workflow was quietly carrying another $0.60.
The 5 Numbers We Track for Every Production AI Workflow
Once the unit cost is visible, the next question is where it is coming from.
These 5 numbers give us enough signal to decide what to inspect next without turning cost management into a finance project.
They also stop a common mistake: cutting the easiest visible line while the expensive part sits somewhere else.
What Each Cost Pattern Tells You to Change
The same monthly spend can come from very different problems. The fix depends on which number is moving.
A high retry rate deserves a reliability investigation. High human minutes point to the exception path. High model and context spend points to what gets sent through the model and which model handles each step.
This is where cost stops being reporting and becomes an architecture decision.
How We Reduce AI Cost Without Reducing Quality
A few rules have survived contact with production for us.
- Keep deterministic work in code. We do not pay a model to do arithmetic, numbering, assembly, or another task software can perform exactly.
- Give each task only the context it needs. Repeated context is paid for repeatedly.
- Keep the model interchangeable. GPT, Claude, Gemini, and the next release can sit underneath the workflow while company context and operating rules stay stable.
- Optimize against the outcome. A cheaper model that creates more retries or more human review can make the workflow more expensive.
The model catalog will change. These rules survive the release cycle.
How Mustang Tracks Cost Without Tying the Workflow to One Model
This is one reason Mustang sits above the model layer.
Mustang can run on one model or several, from any provider. The Grip keeps company knowledge outside the model. Kernel loads the knowledge a task needs, routes work to a cleared worker, checks the result, and tracks cost per run, execution success, and output quality.
Our current FinOps implementation also reads the provider’s own reported usage instead of guessing it. A spend ceiling can hold a run for a named person rather than letting the cost surface later in a monthly report.
Mustang is currently measured at 35%+ lower cost to run AI in production. We care about that number because production is where context, checks, retries, tools, and people all show up.
A 15 Minute Cost Audit You Can Run on One Workflow
Pick one workflow and use a recent batch of real runs.
Add machine spend. Add human minutes. Count retries and failed runs. Divide the total by successful outcomes.
Then circle the biggest line. That is your first optimization target.
That is where we would start too. Give us one workflow, a recent batch of runs, and the systems it touches. We will help you put a real unit cost around it, find where the money is going, and decide which architecture change is worth making first.
Sources
- OpenAI API pricing and GPT 6 Sol model page, accessed September 24, 2026. Standard short context pricing: $2 per 1M input tokens and $10 per 1M output tokens.
- Anthropic, Claude Sonnet 5, pricing updated August 10, 2026. $2 per 1M input tokens and $10 per 1M output tokens.
Share:
Related Insights
AI Agents, Workers, and Orchestrators: Who Does What?
Everyone is doing some version of the same exercise right now. We are putting AI into workflows, automating work that used to sit with people, and sooner or later somebody asks the question that decides whether the whole thing keeps moving: is this actually worth it?
Preparing Your Team for AI: Our Six-Stage Guide
Everyone is doing some version of the same exercise right now. We are putting AI into workflows, automating work that used to sit with people, and sooner or later somebody asks the question that decides whether the whole thing keeps moving: is this actually worth it?
The Elastic Workforce. How AI Agents Are Complementing Your Best People.
Every operations leader knows the exact moment their infrastructure breaks. The volume of incoming work exceeds the physical capacity of the team.
Who Pays When Your AI Goes Rogue?
Picture this. A grieving passenger asks an airline chatbot for a bereavement discount. The AI, trying to be helpful, fabricates a retroactive refund policy that simply does not exist. The passenger buys the full-price ticket, asks for the refund later, and gets denied.
Context Snapshotting: The Missing Layer in Your AI Debugging Stack
You can't debug AI "Ghost Bugs" without freezing the data. Learn how to architect "Context Snapshotting" and "Time-Travel Debugging" for RAG pipelines.
Working on something similar?
We’ve helped teams ship smarter in AI, DevOps, product, and more. Let’s talk.