Guide · AI agents and costs

How much does an AI agent cost, and what drives the price

Assaf Semoun, Founder & CEO of Zooz-TechUpdated 19.11.20266 min read
In short

There is no single price for an AI agent. The cost has two parts: a one-time build and a running cost. The build is driven mainly by the number of systems the agent connects to, the channels, the state of the knowledge sources, the permissions and what the agent is allowed to write. The running cost is driven by model consumption, search and storage, hosting and ongoing support. You can control both at the design stage, and a precise estimate comes after a short scoping phase or a POC on your data.

How much does an AI agent cost?

In short: it depends on what the agent needs to do, not on the technology. An agent that answers questions from well-organized documents in one channel is a completely different project from an agent that identifies a customer, reads from an ERP, writes to it and works in WhatsApp and by voice.

Zooz-Tech does not publish a price list, because a generic range does not help you make a decision. What helps is understanding what moves the price. Every agent has two cost components:

  • One-time build: specification, integrations, knowledge, permissions, security and testing.
  • Running cost: model consumption, search and storage, hosting, monitoring and improvement.

What drives the build cost?

In short: how many systems, how many channels, the state of the knowledge, and what the agent is allowed to do.

Systems to connect

Every system is an integration to build, authorize and test: SAP, Priority ERP, Dynamics 365 or Salesforce. A cloud system with a well-documented API is simpler. A system that runs on-premises needs a gateway or an MCP layer that exposes only defined entities and actions to the agent. When several agents need the same systems, one MCP layer serves all of them, instead of a separate integration for each agent.

Channels

Website, WhatsApp, voice or Teams. Each channel has its own adjustments: answer length, buttons, voice messages, authentication and a communications provider. One brain across several channels saves building twice, but each channel needs its own testing.

Knowledge sources and their state

Digital, well-organized documents go into the index quickly. Scanned files, duplicate versions and legacy systems require text extraction, cleanup and indexing before the agent can rely on them.

Permissions and identity

An agent that answers only general questions is simpler. An agent that shows personal information needs to identify the user, for example with a one-time code checked against the CRM, and to respect the permissions that already exist in the organization.

Language

Hebrew requires handling prefixes, synonyms and everyday language. Every additional language, and voice in particular, adds testing.

Read-only or write actions too

An agent that only reads data is simpler. An agent that creates records, opens cases or posts documents needs controls: a draft instead of a final posting, human approval and a log of every action.

Security and privacy

An agent that handles personal information needs a design that complies with Amendment 13 to Israel's Privacy Protection Law: where the data is stored, what is sent to the model and who sees what. In an organization that works according to ISO 27001, the agent also goes through the organization's security review and documentation.

Testing and evaluation

An AI agent needs more than manual testing. You build a set of test questions with expected answers (Evals), run it on every version, and also check content safety and permissions. That is how you know the agent has not quietly gotten worse.

What do you pay every month after go-live?

In short: usage. The running cost grows with the number of conversations, their length and what the agent does in each one.

Model consumption

There are two main billing models. In Copilot Studio, you pay in usage units (Copilot Credits), bought in packs or pay-as-you-go, and a generative answer or an agent action consumes more than a fixed answer. In Azure OpenAI, through Azure AI Foundry, you pay by tokens: the text that goes into the model and the text that comes out of it. A long prompt, a long conversation history and whole documents in the context increase consumption on every question. Check the prices themselves with Microsoft, because they change.

Search and storage

Azure AI Search is priced by service tier and capacity, not per query. The number of documents and the size of the index, including the vectors, determine which tier you need. Add to that file storage, and sometimes a database for conversations.

Hosting and channels

The services the agent runs on, such as Azure App Service or Azure Functions, and channel costs, such as a WhatsApp provider or telephony.

Monitoring, maintenance and improvement

An agent in production needs tracking of answer quality, errors, consumption and costs (AgentOps), knowledge updates, and improvement based on the questions the agent could not answer. This is not an extra: without it, quality goes down and cost goes up.

Which factors raise the cost, and how do you control them?

Cost driverWhat raises itHow to keep it under control
Connected systemsMore systems, on-premises systems and undocumented interfacesStart with the system that delivers the most value, and use one MCP layer for several agents
ChannelsEach channel built separately, voice and telephonyOne brain for all channels, and add a channel once the first one works
Knowledge sourcesScanned files, duplicates and outdated documentsStart with approved, well-organized sources and expand gradually
Write actionsWriting directly to core systemsA draft and human approval, and only defined actions
Security and privacyRequirements that surface only right before go-liveA security and privacy review already in the specification phase
Model consumptionLong prompts, a large model for every task and calculations inside the modelFocused retrieval, the right model size and deterministic logic outside the model
Search and storageIndexing everything there is, without filteringIndex only what the agent needs, without duplicate versions
MaintenanceNo monitoring, so users are the ones who find the problemsEvals on every version and ongoing monitoring of quality and cost

How do you keep the running cost low?

In short: give the model only what only a model can do. Everything else goes to deterministic logic, search or a cache.

  • Deterministic logic outside the model. Calculations, rules and dates do not need a language model. A flow in Power Automate or in code gives the same result every time, and saves tokens.
  • Retrieval instead of long prompts. Instead of putting the whole manual in the prompt, search retrieves only the passages relevant to the question (RAG). Less text in every call.
  • Caching. Recurring questions and answers that do not change do not need to go through the model every time.
  • The right model size. Classification, routing and field extraction work well on a small model. Keep the large model for tasks that really need it.
  • Focused scope. An agent that does one process well costs less to build and run than an agent that tries to answer everything.

Gil Car · Invoice agent with SAP on-premises: the Copilot Studio agent reads the invoice and checks it against SAP to prevent duplicates. Four dates are calculated in Power Automate, with deterministic logic, not by the agent. As a result, the outcome is consistent on every invoice, and the model consumes fewer tokens. The draft is created in JournalVouchers through an MCP layer, and an authorized user approves it inside SAP.

Copilot Studio or Azure AI Foundry: which costs less?

In short: there is no blanket answer. It depends on usage volume and complexity.

In Copilot Studio, building is faster, the connection to Microsoft 365 and Power Platform is built in, and the running cost is measured in usage units. It suits internal agents and low to medium volume. In Azure AI Foundry, you pay for consumption: tokens, search and infrastructure. Development and maintenance take more, but at high volume you can choose a model for each task and control the cost better. The full breakdown is in the comparison of Copilot Studio and Azure AI Foundry.

What drove the cost in our projects?

Ashtrom Group · Operational agent in WhatsApp: the agent identifies the customer and the order, pulls the truck data from Priority ERP and calculates the arrival time based on traffic. Here the cost is driven by the WhatsApp channel with voice messages, the live connection to Priority, the maps service and hosting on Azure.

KESHER · Resident service in every channel: the same brain on the website, in WhatsApp and by voice, on Azure AI Foundry and Azure AI Search. Here the cost is driven mainly by indexing the local authority's content, the channels and the connection to the CRM.

How does Zooz-Tech estimate the cost?

In short: start small, on real data, with a scope and end date set in advance.

  1. Discovery. A working session with the process owners, the users and IT. We map processes, systems and data, and agree on a success metric.
  2. POC on real data. A working version of one process, on the organization's data and systems. At the end, we decide together: Go / No-Go. Even if you stop, you keep a grounded estimate of the cost, benefit and timeline of the full project.
  3. Specification and phases. An approved specification with milestones and acceptance criteria. Development moves forward in sprints, and scope changes are decided together, with their impact on time and cost.
  4. After go-live. Monitoring of usage, quality and costs, so the running cost is measured, not guessed.

The full process is described on the How we work: Delivery & POC page. On agents in general: AI & agents and Copilot Studio agents.

Frequently asked questions

How much does an AI agent cost?

It depends on the number of systems the agent connects to, the channels, the state of the knowledge sources and the actions it performs. Zooz-Tech gives a precise estimate after a short scoping phase or a POC on the organization's data.

What is the difference between the build cost and the running cost?

The build is one-time work: integrations, knowledge, permissions, security and testing. The running cost is paid by usage: model consumption, search and storage, hosting, monitoring and improvement.

How is model consumption calculated?

In Copilot Studio, by usage units (Copilot Credits) consumed with each message or action. In Azure OpenAI, by tokens of the text that goes into the model and comes out of it.

How do you reduce the running cost?

Move deterministic logic out of the model, retrieve only the relevant passages instead of long prompts, cache recurring answers and choose a model size that fits the task.

Why start with a POC?

Because it shows the solution working on your data and gives a grounded estimate of the cost, benefit and timeline of the full project, before you commit.

Assaf Semoun
Assaf Semoun

Founder & CEO of Zooz-Tech, solutions architect, Microsoft 365 Architect Expert.

Want to know what will drive the cost of your agent?

Book a call