Azure AI FoundryThe custom part of your agent, inside your Azure.
Microsoft Foundry, formerly Azure AI Foundry, is where agents that need more than Copilot Studio run: your own code, your choice of model, private networking and an identity of their own. QServices builds Foundry agents that read documents, check the ERP and stop at a draft until a named person approves.
Microsoft Foundry is Microsoft's Azure platform for building AI apps and agents. It was called Azure AI Studio in 2023 and Azure AI Foundry from November 2024, and took its current name in November 2025. It brings over 10,000 models from Microsoft, OpenAI, Anthropic, Meta and others together with Foundry Agent Service, evaluation, safety controls and one set of Azure permissions.
Prompt agents are configuration; hosted agents are your own code. Most ERP agents end up hosted, because they need custom checks.
| Prompt agent | Hosted agent | |
|---|---|---|
| What it is | Instructions, a model and tools, run by Foundry | Your code (Agent Framework, LangGraph, OpenAI or Anthropic SDKs) in a sandbox per session |
| Platform charge | None: you pay model tokens and tools | Container time: $0.0994 per vCPU-hour and $0.0118 per GiB-hour (East US 2 list) |
| Networking | Private networking available | Bring your own virtual network, set when the Foundry resource is created |
| Identity | Project or agent identity | Its own Entra agent identity, created at deploy |
| Fits | Q&A and simple tool use | ERP checks, document pipelines, multi-step agents |
Foundry's visual workflows retire on 1 December 2026; Microsoft points to Agent Framework or Logic Apps instead. From Microsoft Learn and the Azure Retail Prices API, October 2026.
Azure Policy can block Global deployments to keep processing in a Data Zone.
MCP tools need approval on every call unless you change it.
Every agent identity in the tenant is listed in the Entra admin centre.
Tool-call guardrails are still in preview.
Azure OpenAI has no hard spending cap, so budgets and alerts do that job.
| Copilot Studio | Microsoft Foundry | |
|---|---|---|
| Who builds | Business builders and our team, low-code | Developers, in code |
| Best for | Agents in Teams and Outlook, approvals, flows | ERP checks, private networking, custom models and evaluation |
| Billing | Copilot Credits | Model tokens, tools and hosted-agent compute |
| Together | Can call a Foundry agent as a connected agent (preview) | Can publish agents to Teams and Microsoft 365 Copilot |
Microsoft's own advice: check for a ready-made agent first, then Copilot Studio, then Foundry. See our [Copilot Studio page](/platforms/copilot-studio-consultant).
Yes. Azure AI Studio became Azure AI Foundry in November 2024 and Microsoft Foundry in November 2025.
Prompt agents have no platform charge: you pay for model tokens and tools. Hosted agents add container time, about $0.0994 per vCPU-hour and $0.0118 per GiB-hour at East US 2 list prices, only while sessions run.
Prompt agents are configuration that Foundry runs. Hosted agents run your own code in an isolated sandbox per session. Visual workflows retire on 1 December 2026.
Microsoft states prompts and completions aren't available to other customers or model makers and aren't used to train their models. Stored agent data sits in your resource's geography, and Data Zone deployments keep processing in the US, EU or APAC.
Yes: through Business Central's MCP server, Logic Apps connectors, Dataverse or the ERP's own API. Business Central and Dataverse each allow 6,000 requests per user per 5 minutes, so we batch calls.
MCP tool calls in Foundry need approval by default. We keep that on for anything that writes, so the agent drafts and a named person approves.
Facts on this page checked on 23 October 2026.
Founder and CEO
Chief Sales Officer
Join our live webinar on Sept 24 and see five
ways to automate meetings, approvals, and
document search with tools already in your
tenant.
