Azure AI Foundry vs Copilot Studio — the Australian decision guide for enterprise agents

Every Australian organisation that has crossed the line from "interested in agents" to "building one" is asking the same question this quarter. Should the first production agent live in Microsoft Copilot Studio, or in Azure AI Foundry? It is the right question to ask early. The two surfaces feel similar at the demo level and diverge sharply at the operating-cost and governance level. Picking the wrong one for the use case in front of a team is one of the easier ways to spend six months building the second version of an agent that should have been the first.

Frontrow has been through this decision with clients enough times this year to publish the working version of the guide. Microsoft itself has been clear that the two platforms are designed to interoperate, not compete, but the framing inside Australian organisations is almost always "which one first." That is the question this guide answers.

The one-line distinction worth remembering

Copilot Studio is the low-code, Microsoft 365-grounded surface for business teams to compose conversational agents that live inside Teams, SharePoint and Microsoft 365 Copilot. Azure AI Foundry is the pro-code, Azure-native platform for engineering teams to build, evaluate and operate agents at scale, with full model choice, tracing and integration into custom systems. Same Microsoft AI stack underneath, very different operating models on top.

When Copilot Studio is the right answer

Copilot Studio is the right surface when the agent is conversational, lives close to Microsoft 365 data, needs to ship inside weeks rather than months, and is going to be owned and iterated by people who are not full-time engineers. Frontrow's working list of Copilot Studio fits is short and consistent.

• The grounding data lives in SharePoint, OneDrive, Teams, Dataverse or a small set of named systems with off-the-shelf connectors.
• The knowledge base is under a few hundred documents, and document refresh is measured in days not minutes.
• The intended audience reaches the agent through Microsoft Teams, the M365 Copilot side panel, or an embedded chat surface on a SharePoint site.
• A subject-matter expert in the business should be able to update the agent's instructions, prompts and content without filing a ticket.
• The unit economics work at $0.01 per Copilot Credit, with the $200-per-month 25,000-credit pack as the natural starter shape.

The classic Copilot Studio win Frontrow sees in Australian mid-market is the policy and process agent. HR builds it. The agent reads a SharePoint library of policies, leave entitlements, approval workflows. It answers in Teams. It refers complex cases to a human. The build is two to four weeks. The operating cost is small and predictable. The owner is HR, not IT.

When Azure AI Foundry is the right answer

Foundry is the right surface when the agent needs to be engineered, instrumented and operated like a production application, when model choice and evaluation matter, and when the integration footprint reaches well beyond Microsoft 365.

• The use case requires fine-tuned, custom, or non-Microsoft models, including Anthropic Claude and Meta Llama variants now hosted in Foundry.
• The grounding data is large, fast-moving, or sits in core line-of-business systems (the data warehouse, the case management system, the asset register, a regulated data store).
• The agent will be deployed into channels beyond Microsoft 365 — a customer-facing web app, a contact-centre workflow, a mobile experience, a partner portal.
• The team needs structured evaluations, tracing, A/B testing, prompt versioning and the rest of the production AI engineering surface.
• The expected throughput will run thousands of interactions per day, where the Foundry consumption model lands materially cheaper than Copilot Credits at the same volume.

The classic Foundry win Frontrow sees is the regulated workflow agent — claims triage at an insurer, advice scaffolding at a wealth manager, frontline coaching at a contact centre. Engineering owns it. The model is chosen for accuracy and audit trail. The agent runs against a vector index over a hundred-thousand-document corpus. Evaluations gate every release. The cost model is consumption, the governance model is rigorous, and the operating discipline is software engineering.

The cost reality on each path

Copilot Studio prices in Copilot Credits. The headline rate is $0.01 per credit, with the $200-per-month pack including 25,000 credits. Microsoft 365 Copilot users get some interactive usage included, but autonomous and scheduled actions always consume credits. The cost is predictable, easy to budget, and forgiving for the first few hundred users on a single agent. It scales linearly with usage.

Foundry prices in Azure consumption across several axes. Foundry-native agents using prompts and workflows have no platform fee, and the cost is in the model token consumption, the connectors, and the tools the agent calls. Hosted agents add compute at $0.0994 per vCPU-hour and $0.0118 per GiB-hour. From 1 June 2026, Foundry agent memory becomes a separately billed line at $0.25 per 1,000 events for short-term memory, $0.25 per 1,000 memories per month for long-term, and $0.50 per 1,000 retrievals. Foundry's economics are excellent at scale and unforgiving at low volume, where the engineering overhead can outweigh the consumption itself.

The cost crossover for Australian mid-market sits roughly around 2,000 to 5,000 daily interactions on a single agent. Below that, Copilot Studio usually wins on total cost of ownership including the engineering hours saved. Above it, Foundry usually wins on consumption cost and on the engineering controls that come with it.

The third answer — both

The pattern Frontrow sees most often in mature deployments is composed. Foundry hosts the model, the evaluation, the vector index and the deeper tool calls. Copilot Studio composes the user-facing agent that lives in Teams or M365 Copilot, and reaches into Foundry for the heavy lifting. Business users own the conversational layer. Engineering owns the model and grounding. This is how Microsoft itself shapes its larger reference architectures, and it is the pattern that survives a year of iteration without rewrite.

What Frontrow recommends for an Australian mid-market first agent

Default to Copilot Studio for the first production agent unless the use case clearly fails one of the Copilot Studio fit checks above. The reason is operating discipline, not capability. Building the first agent in Copilot Studio forces clean grounding data, named ownership outside of IT, and a real adoption signal before the engineering investment lands. The lessons from the first agent shape the second one. The second agent is often a Foundry build, sometimes a composed build. The third one almost always is.

Australian organisations that start with Foundry on a use case that should have been a Copilot Studio build tend to land six months later with a beautifully engineered agent and a flat adoption curve, because the business never owned the conversational layer. Australian organisations that start with Copilot Studio on a use case that should have been Foundry tend to bump into the cost ceiling or the model-choice ceiling and reach for Foundry second, which is the right shape of mistake to make.

Frontrow runs the platform decision conversation as a 30-minute call against the actual use case in front of a team. Most decisions land in that hour. The build follows in weeks if the call is held early, in months if it is held after the first prototype has already taken a side. Book a time through the contact page.