DE·EN·FR Contact

AI & Copilot · Custom Agents

Copilot Studio Agents — custom agents that do more than smalltalk.

Microsoft 365 Copilot is good for generic knowledge work. But when your answers sit in a CRM, an ERP, a wiki, or a custom application — you need a custom agent. We build them on Microsoft Copilot Studio, with Power Platform connections to Microsoft Dataverse, Microsoft Dynamics 365, and third-party systems.

Book 60-min demo View agent types
Microsoft Copilot Studio Power Platform · Apps, Automate, Pages Microsoft Dataverse Power Platform connectors · 1,500+ standard, custom APIs

When a custom agent is the right answer

Standard Copilot or custom agent — a question of three criteria.

We're often asked whether a custom agent is "better" than Microsoft 365 Copilot. The question is wrongly framed — both have their domain. Three criteria help make the right choice.

Does the knowledge sit outside Microsoft 365? Microsoft 365 Copilot naturally accesses SharePoint, OneDrive, Outlook, Teams. But when the right answers sit in Microsoft Dynamics 365, an Atlassian wiki, an Asana project, or a custom database — a custom agent on Copilot Studio is the natural lever. Via Power Platform connectors we wire up the source, define the permission logic, and shape the answer rules.

Should a workflow be triggered? Standard Copilot generates text. A custom agent can trigger actions — open a ticket in Microsoft Dynamics 365, start a Power Automate flow, send a mail via Outlook, book a calendar slot in Microsoft Bookings. As soon as you need not just information but action, you can hardly avoid Copilot Studio.

Should the agent be role-specific? Microsoft 365 Copilot is the same for every person with a license. A custom agent can be configured differently per role, department, or location — with clear data boundaries, dedicated personas, specific use cases. That's the most common organizational driver we hear in initial conversations.

Three agent types we build

From knowledge bot to data agent with write access.

We split custom agents into three archetypes. The types differ in complexity, license logic, and risk profile. Which type fits your concern we clarify in the discovery conversation.

Type 1

Q&A agent

The agent answers questions from a scoped knowledge base — documentation, FAQ, internal wikis, compliance handbook. Classic RAG pattern. Lowest complexity, fastest time-to-value (4–6 weeks). Examples: HR bot for employee questions, service bot for product documentation, IT bot for software inquiries.

  • RAG architecture on SharePoint, Dataverse, external sources
  • Source references in every answer
  • Fallback for uncovered questions
  • Deployment in Teams, Microsoft 365 Copilot, Power Pages
Type 2

Process agent

The agent executes workflows — starts Power Automate flows, writes to Microsoft Dynamics 365 or Dataverse, calls third-party APIs. Medium complexity (8–14 weeks). Examples: onboarding agent that guides a new hire through all steps; service agent that escalates a ticket; sales agent that creates proposals.

  • Power Automate connection as action layer
  • Microsoft Entra ID–based permission logic
  • Confirmation steps before irreversible actions
  • Logging and audit trail of every action
Type 3

Data agent

The agent creates reports, dashboards, and analyses from live data — typically over Microsoft Dataverse or Microsoft Fabric. Higher complexity, because data model, permission logic, and performance must come together. Examples: sales-pipeline agent for management, service-KPI agent for the operations lead.

  • Live connection to Dataverse, Microsoft Fabric, Power BI
  • Predefined standard reports plus open questions
  • Row-level security at the source
  • Caching and performance optimization

Examples from practice

Three anonymized custom agents we delivered in 2025.

Service knowledge bot — machinery manufacturer, ~ 800 staff

Q&A agent for service staff. Accesses 12,000 documents in SharePoint and the Microsoft Dynamics 365 knowledge base. Answers questions like "Which screw sizes for model X 2018?". Time-to-value: 5 weeks. Current usage: about 40 queries per day, about 70% hit rate.

Onboarding agent — mid-sized service provider, ~ 250 staff

Process agent that guides new staff through onboarding. Starts Power Automate flows for equipment request, IT accounts, training booking. Distributes reminders, collects confirmations. Time-to-value: 11 weeks. Effect: HR effort per onboarding reduced by an estimated 35%.

Sales-pipeline agent — B2B software vendor

Data agent for management. Answers questions like "What does our DACH pipeline for Q4 look like?" or "Which deals have been in stage 3 for more than 60 days?". Live access to Dataverse, with row-level security at management level. Available in Teams as a quick action.

What we don't build

No "AI does everything" promise. We don't build agents that make complex decisions without a knowledge basis — for example, application agents that autonomously reject people. Such use cases violate the EU AI Act and are ethically problematic. We say so honestly, even when the inquiry would be attractive.

Honest carve-out

When Copilot Studio isn't the right tool.

We are a Microsoft Partner — and still say when Copilot Studio doesn't fit. Three situations:

You want to build a product with an AI feature that goes to the outside. Copilot Studio targets employee applications. If your use case is a customer-facing product, scaling beyond Microsoft 365 tenants, with individual branding and other LLMs — then our sub-service Open AI Integrations is the better lever.

You need full control over the model. Copilot Studio uses Microsoft's model stack (Azure OpenAI). If you want to use a specific model (Anthropic, Mistral, local Llama), need fine-tuning, or must host on-premises — then Copilot Studio is too constrained.

Your use case is high-risk under the EU AI Act. Assessment agents in HR, credit-scoring agents, biometric identification — we don't classify such use cases as Copilot Studio projects at all. We refer to specialized advisors and don't take the engagement.

FAQ

What management wants to know before the first conversation.

When is a custom agent worth it versus standard Microsoft 365 Copilot?

Three triggers: first, when the knowledge sources sit outside Microsoft 365 (CRM, ERP, wiki, Filemaker, third-party systems). Second, when a workflow should be triggered, not just an answer generated. Third, when the agent should be available to specific roles, with clearly defined data and action boundaries.

How long does development of a Copilot Studio agent take?

A simple Q&A agent with a clearly scoped knowledge base is in production in 4 to 6 weeks. A process agent with workflow connection to Microsoft Dynamics 365 or third-party systems takes 8 to 14 weeks. Data agents with Dataverse write access are an architecture topic and need according preparation.

What does a Copilot Studio agent cost in licensing?

Microsoft licenses via two models: per-user (Copilot Studio User License) or via message-based packages. Which model is cheaper depends on user count and interaction frequency. We work this out as part of our License Cost Calculator advisory.

Can Copilot Studio agents access SAP, NAV, or other third-party systems?

Yes — via Power Platform connectors (standard, premium, custom) or custom API integration. We have experience with SAP, Oxaion, NAV, and several in-house systems. Authentication is key: service principal with least privilege, no pseudo user.

How do we prevent the agent from hallucinating?

Three technical levers: first, grounding on concrete knowledge sources (RAG pattern), not on the open model. Second, explicit fallback rules (rather "I don't know" than a made-up answer). Third, logging of all answers with source trail so hallucinations are identifiable after the fact. Plus organizationally: a clear correction process for wrong answers.

Do Copilot Studio agents also work in Microsoft Teams?

Yes, that's one of the most common deployment paths. The agent can be published as a Teams app, with its own bot profile and permission control via Microsoft Entra ID. Alternatively as an embedded agent in Microsoft 365 Copilot, in a Power App, or on a Power Pages site.

What about data protection and the EU AI Act?

Custom agents are often classified as "limited risk" under the EU AI Act, sometimes as high-risk — depending on the use case. We classify the planned agent in our AI governance program and build the configuration so that training duty (Art. 4), transparency, and audit trail are met.

60-min demo · concrete use cases · no obligation

See a Copilot Studio agent in action.

We show you one of our reference agents live. Afterwards we discuss whether and how your concern works as a custom agent — Q&A, process, or data.

Book 60-min demo Or 30-min initial conversation

Accompanying services

What typically runs alongside this engineering work.

Engineering projects rarely stand alone — license logic, architecture clarification, quality gates, knowledge transfer, and follow-on operations usually run in parallel. Here are the most common accompanying services we add to Discovery Spikes, sprint fixed-price engagements, or Application Care contracts.

Up front · architecture

Advisory & Architecture

Before implementation: tenant structure, data model, security concept, integration mapping. The result is an architecture document any engineering team can pick up — including one other than us.

Read more →

Up front · CSP

License Advisory & CSP

Which license bundles for which users, which add-on SKUs are needed, where you are over- or under-licensed. Procured via Microsoft Licensing Partner — with the option to use CSP purely as a control mechanism without margin maximization.

Read more →

During · quality gate

Project Assurance

Independent second opinion during a running implementation project — whether we are delivering it or another partner. CMMI-based quality gates, risk reviews, fixed price per gate.

During · adoption

Training & learning program

Not the classic two-day workshop that's forgotten after a week — but a dynamic learning program over 4–6 weeks with kickoff training, application phases, and advanced sessions. Training matrix for roles and topics.

Read more →

After · operations

Application Care

After go-live: a predictable Application Care contract with monthly flat rate, SLA-based. Includes releases, hotfixes, extensions, tenant hardening — and continuous support instead of merely reacting to tickets.

Read more →

After · knowledge

Knowledge Recovery

When the original developers are gone, the previous partner is no longer reachable, or the documentation is outdated — reverse engineering of the existing solution with a documented result: code map, data model, customization inventory.

Read more →