Buyer's guide
What It Costs to Run AI Agents in Your Own Azure Tenant
The cost of an AI agent is more than the model bill. Here are the four cost buckets, what drives each one when agents run in your own Azure tenant on Microsoft Foundry, the costs teams tend to miss, and how to estimate one workflow before you commit.
Running an AI agent costs four things: the build, the model usage, the Azure platform around it, and the people and process that operate it. When the agent runs in your own Azure tenant on Microsoft Foundry, the agent runtime itself carries no separate charge; you pay for model tokens, the tools the agent calls, any hosting compute, storage and search, plus Microsoft licensing for governance. Volume, steps per task and model choice drive most of the running cost, and ongoing operation often outweighs the model bill.
The short answer
The four cost buckets of an AI agent
| Bucket | What it covers | What drives it |
|---|---|---|
| Build | Workflow mapping, integrations, prompts and rules, testing on real work | How many systems the agent touches and how messy the data is |
| Model usage | Tokens the agent reads and writes on every task | Task volume, steps per task, document size and the model chosen |
| Azure platform | Tools the agent calls, hosting compute, storage, search and monitoring in your tenant | Which tools are used, how much data is indexed, how long logs are kept |
| Operation | Monitoring, exception handling, updates, governance and reporting | How many exceptions need a person and how often the process changes |
Model usage
What drives the model bill
Models on Microsoft Foundry are billed mainly per token, with separate rates for input and output. An agent that reads a long contract and writes a short summary is driven by input; one that drafts long replies is driven by output.
The biggest levers are the number of tasks, the number of model calls each task makes, and which model handles each step. Routing simple steps to a smaller model and saving the strongest model for hard judgment calls is usually the largest saving available.
Microsoft also offers provisioned throughput, an hourly reservation that trades flexibility for predictable cost at steady high volume, and discounted batch processing for work that can wait up to a day. Cached input, where an agent reuses the same instructions or documents, is billed at its own rate. Check Microsoft's current Foundry Models and Azure OpenAI pricing pages for the numbers, because they change often.
Azure platform
Platform costs in your own tenant
Agent runtime
Microsoft states there is no separate charge for creating or running Foundry-native agents built with prompts and workflows. Agents built on other frameworks run as hosted agents, billed on the compute they use.
Built-in tools
File search is billed on vector storage, code interpreter per session, and web search per block of transactions. Connectors such as Logic Apps or Azure AI Search carry their own charges.
Data and search
Storage for documents, indexes the agent searches, and the databases it reads from. Index size and refresh frequency set the cost.
Monitoring and logs
Tracing, evaluation runs and log retention. Keeping a full audit trail is worth paying for; keeping it forever usually is not.
Governance licensing
Managing agents with Microsoft Agent 365 and giving each one an Entra Agent ID depends on your Microsoft 365 agreement. Confirm what your current licenses include before you budget.
Network and security
Private endpoints, key management and security monitoring keep data inside your tenant. Many firms already pay for these, so check before adding them twice.
Estimate one workflow
Five questions to size the cost of one agent
- Question one
How many tasks a month?
Count the invoices, tickets, requests or documents the workflow handles today, and how that number moves through the year.
- Question two
How many steps per task?
List what a person does for one task: read, look up, compare, decide, write, update. Each step is roughly one model call or tool call.
- Question three
How big are the inputs?
A one-page form and a long contract cost very differently to read. Note the typical size of what the agent has to take in.
- Question four
Which systems does it touch?
Each system with a ready connector is cheap to reach. Each one without is a build and maintenance cost.
- Question five
What happens on an exception?
Estimate how often a person has to step in and how long each case takes. That is often the largest line in the total.
How Tekscape prices it
How this maps to what Tekscape charges
Tekscape builds and runs agents in your own Azure tenant, so the Azure and Microsoft usage stays on your own bill where you can see it. Our fee covers the build as a fixed-scope pilot, then a monthly fee for operating each agent only after it passes its acceptance test.
The details of that model, and what to define before any agent is priced, are in our guide to outcome-based pricing for AI agents.
Frequently asked questions
Sources
- Microsoft: Foundry Agent Service pricingNo separate charge for Foundry-native agents; billing for models, built-in tools, third-party tools and hosted agents.
- Microsoft: Azure OpenAI pricingPer-token standard pricing, provisioned throughput, discounted batch and separate cached-input rates.
Size one workflow with us
Bring one workflow to a free AI outcomes session. We will walk through the five questions with you and show what a pilot would cover.
Book a free AI outcomes session