Copilot Studio Agent Build Playbook
Copilot Studio is the low-code platform for building custom AI agents. Done well, it produces measurable ROI + integrates cleanly with your AI governance. Done poorly, it produces expensive shadow AI that fails audit. This is the EPC Group build playbook.
Frequently Asked Questions
What is Copilot Studio and when should we build a custom agent?
Copilot Studio (formerly Microsoft Copilot Studio, formerly Power Virtual Agents) is Microsoft's low-code platform for building custom AI agents that either extend Microsoft 365 Copilot (declarative agents inside M365) or run as standalone conversational experiences (Teams bots, web chatbots, embedded assistants). Build a custom agent when: (1) you need domain-specific knowledge Copilot doesn't have (product catalog, internal process docs, vertical-specific procedures), (2) you need agent-driven actions (create Salesforce lead, escalate ServiceNow ticket, approve invoice), (3) you need brand-controlled UX (call center-facing agent with your logo and voice), or (4) you need auditability + versioning on the AI behavior (regulated verticals).
What are the design patterns?
Two structural choices intersect with two behavior choices. STRUCTURE: (a) Declarative agent — Copilot Studio agent installed into M365 Copilot; users invoke it from within M365 Copilot chat with "@AgentName". (b) Standalone — deployed to Teams / web / mobile / call center as its own experience. BEHAVIOR: (i) Topic-authored — deterministic conversation flows built via graphical topic authoring; predictable, testable, verbose to build. (ii) Generative — LLM-driven with grounding on connected data sources; flexible, faster to build, harder to audit. Best practice is hybrid: topic-authored for regulated actions (compliance-critical workflows), generative for everything else.
What data sources can Copilot Studio agents use?
Six categories. (1) SharePoint sites — grounding on document libraries via managed identity. (2) Microsoft Dataverse tables — structured data with row-level security. (3) External APIs via Custom Connectors (Salesforce, ServiceNow, SAP, custom REST APIs). (4) Uploaded files (PDFs, Word docs) into agent knowledge. (5) Public websites via crawler. (6) Enterprise Data Integration — Fabric, Azure SQL, Snowflake through Dataverse virtual tables or Fabric shortcuts. Every source needs a documented data classification + access control + refresh schedule as part of the agent design.
How do we handle testing + ALM?
Copilot Studio has native dev/test/prod environments through Power Platform environments. Best practice: separate Dataverse environments per stage, solution-based packaging for portability, test topics with automated dialog test suites, generative agents tested via a curated prompt/response evaluation set. Deployment via solution import into the target environment. Governance-critical: the same solution should not skip the test environment — enforce this via Power Platform Governance policy. EPC Group's engagement includes the ALM configuration + test suite scaffolding + Power Platform Governance policy setup.
How do we control cost?
Copilot Studio pricing is message-based via Message Capacity add-on packs. Cost lever 1: cap message capacity per environment to prevent runaway. Lever 2: rate-limit per-agent at the tenant level. Lever 3: route generative queries to cheaper models where possible (Azure OpenAI PTU-backed for high-volume, GPT-4o for premium cases). Lever 4: measure cost per resolved interaction against the human-alternative cost per resolved interaction — this is the metric that justifies the agent existing. EPC Group builds Power BI dashboards showing cost per resolved interaction by agent + BU.
What does an EPC Group Copilot Studio engagement produce?
Six deliverables (4-8 week fixed-fee engagement per agent, depending on complexity). (1) Use case brief with cost/benefit vs alternatives (human, existing chatbot, off-the-shelf tool). (2) Agent design document — declarative vs standalone, topic/generative hybrid map, data source classification, access controls. (3) Built agent + dev/test/prod environments + ALM configuration. (4) Test suite (topic + generative eval set) with pass/fail gates. (5) Cost + capacity monitoring dashboard (Power BI). (6) Governance policy integration — the agent is registered in the AI Portfolio Map (see /answers/what-is-a-vcaio) and covered by the vendor risk policy (see /answers/copilot-governance-policy-library).
Talk to a senior architect
Email contact@epcgroup.net or call 888-381-9725.
North America's oldest continuous Microsoft Gold Partner (2000 until Microsoft retired the program in 2022) — today holding all six Microsoft Solutions Partner Designations.