Microsoft now offers three distinct platforms for building AI agents. Each serves a different audience and maturity level - but they're converging. The gap between low-code and pro-code is narrowing with every release, and capabilities that required full Azure development twelve months ago are now available in Copilot Studio's visual builder.
For most organisations, this is good news. It means you can start building real, production-grade agents today without a development team, and you won't be left behind as the platforms evolve.
The short version: Copilot Studio delivers the majority of what most organisations need at a fraction of the complexity. It's fast to build in, constantly improving, and doesn't require dedicated developers. Azure AI Foundry is the right choice for enterprises with engineering teams who need absolute control - but that's a smaller group than you might think.
Three platforms, three personas
M365 Agent Builder
For individuals building personal productivity agents within Teams, Word, or Outlook. Conversational setup, minimal configuration. A great starting point for learning what's possible.
Copilot Studio
For professional makers building agents for teams, divisions, or the whole organisation. Rich connectors, MCP support, multi-agent orchestration, and built-in evaluations - without writing code.
Microsoft Foundry
For full-stack developers who need maximum control over models, memory, deployment, and lifecycle management. Complete flexibility, complete responsibility.
A word on Agent Builder specifically: it's best understood as a tool for end users to create small, personal agents that do simple things - answer questions from a SharePoint site, summarise a meeting, draft an email. It does technically support connectors, but the model choice is heavily restricted and you have no control over how the AI reasons. In practice, that means it's more prone to hallucination and less reliable on anything nuanced. There's a productivity ceiling, and most teams hit it quickly. Once you need multi-step logic, proper grounding, or any kind of quality assurance, you've outgrown Agent Builder. That's not a criticism - it's by design. It's the on-ramp, not the motorway.
Different people in your organisation will use different tools. That's by design. Someone in sales might build a personal agent in Agent Builder to answer quick questions. Your operations team might use Copilot Studio to automate a multi-step approval process. Your engineering team might use Foundry for a customer-facing product. The key is matching the tool to the persona, the audience, and the requirements.
How capabilities compare
| Dimension | Agent Builder | Copilot Studio | Foundry |
|---|---|---|---|
| Model selection | None (decided for you) | Curated choice | Full catalogue (11,500+) |
| Deployment | M365 only | M365 + web + channels | Anywhere |
| Connectors & tooling | Some connectors, limited control | 1,400+ connectors, MCP, REST APIs | All of the above + OpenAPI |
| Multi-agent | No | Yes | Yes + cross-platform orchestration |
| Evaluations | None | Test sets, quality scoring, groundedness | Full suite + adversarial red teaming |
| Lifecycle (CI/CD) | None | Environments, versioning | Full Git + pipelines + native versioning |
| Observability | Basic usage stats | Built-in analytics | Full OpenTelemetry + App Insights |
| Memory | None | Not yet (expected) | Memory as a service |
| Content safety | Fixed | Configurable | Highly configurable + standalone endpoint |
| Audience | Internal only | Internal + external | Anyone |
| Cost model | Included in M365 | Copilot credits | Azure consumption + API gateway |
The maturity path
There's a natural progression most organisations follow. You start by building agents for yourself - learning how AI works, building confidence in the outputs, understanding what's possible. Then you share with your team. Then your division. And eventually, once governance and evaluations are well established, you expose agents to customers.
Copilot Studio sits comfortably across most of that path. It scales from simple team assistants to complex, multi-agent business process automation without requiring a platform change. Agent Builder is your entry point for individual exploration. Foundry is where you go when you've outgrown what's possible in low-code - or when you're building for external customers with strict requirements around safety, observability, and cost control.
When does Foundry make sense?
Foundry is the right answer when you genuinely need fine-tuned models, long-term memory services, full Git-based lifecycle management, adversarial red-teaming evaluations, or OpenTelemetry-grade observability. It's also the natural home for customer-facing agents requiring granular cost control via API gateway policies.
These are real requirements - but they're enterprise-grade requirements. If you're asking "do we need this?" the answer is usually "not yet." And the Microsoft ecosystem is designed so that Copilot Studio agents can call Foundry services selectively when you need just one advanced capability without rebuilding your entire solution.
The convergence
The distance between Copilot Studio and Foundry is shrinking with every platform release. Evaluations, MCP support, multi-agent orchestration - features that were pro-code-only twelve months ago - are now available in Copilot Studio's visual builder.
Microsoft has been clear: the ingredients will converge. The difference between the platforms will increasingly be about how you assemble them, not what you can build. For most organisations, starting with Copilot Studio means building on a platform that meets you where you are today and grows to meet you where you'll be tomorrow.
Not sure where to start? We help organisations identify the right platform for each use case, build their first agents in Copilot Studio, and establish the governance foundations that make scaling possible. Get in touch to talk through your situation.
Based on John Savill's technical overview of Microsoft agent platform decision points. Published May 2026.