How to create an internal policy review assistant in Microsoft Copilot

João Barros 15 de September de 2026 4 min read

This tutorial shows how to create an assistant in Microsoft Copilot for internal policy review: identify inconsistencies, update legal references and suggest clearer language — useful for teams that want to keep policies consistent and auditable.

Prerequisites

Account with access to Microsoft Copilot and Copilot Studio (administrative permissions to create agents).
Set of internal policy documents in Word or PDF format stored in OneDrive or on SharePoint.
Basic knowledge of prompts and instruction structure (e.g., examples of correct/incorrect policies).

Step 1: Define the objective and review criteria

Before creating the agent, clearly write what you want the assistant to do. Define criteria such as: check for non-inclusive language, update references to laws, point out contradictory clauses and suggest simplifications. This helps build prompts and training examples.

// Exemplo de critérios mínimos (texto de apoio para o agente)
// - Identificar termos ambíguos
// - Sugerir linguagem inclusiva
// - Verificar datas e referências legais
// - Marcar conflitos entre secções

Step 2: Prepare a corpus of examples

Gather examples of correct and flawed policies. Include before/after pairs to show how to correct them. These examples serve as few-shot learning within the Copilot Studio prompt or as training content.

Exemplo (antes): "O funcionário deverá obter aprovação do seu superior." 
Exemplo (depois): "O/a trabalhador/a deve obter aprovação do respetivo superior hierárquico."

Step 3: Create a new agent in Copilot Studio

Open Copilot Studio and choose to create a new agent. Give it a descriptive name like "Revisor de Políticas Internas" and paste the description with the criteria defined in Step 1. Select the appropriate base model (for example GPT‑4o, if available) and configure memory/context content to connect to the document repository.

// Configurações recomendadas
tag: RevisorPolíticas
model: GPT-4o
context_window: 8k
memory: ligado (para guardar historial de revisões)

Step 4: Connect the agent to the document repository

Grant the agent access to the OneDrive/SharePoint where the policies are stored. Configure read and versioning permissions so the assistant can consult the current text and annotate suggestions without modifying the original file.

// Exemplo conceptual de passos:
// 1. Conectar SharePoint site
// 2. Seleccionar a pasta "Políticas"
// 3. Dar permissões de leitura ao agente

Step 5: Build prompts and response templates

Create prompts that request structural analysis and editable suggestions. Use an output template with clear sections: Summary, Issues found, Suggested text and References for human review.

Prompt template:
"Analisa o documento X. Fornece:
1) Sumário das alterações sugeridas
2) Lista de problemas (linha/parágrafo)
3) Sugestão de texto revisado (apenas o segmento)
4) Risco/urgência e referência legal, se aplicável"

Step 6: Test the agent with a pilot document

Submit a test document and ask the agent to perform the review. Evaluate the quality of suggestions, the accuracy of references and whether the agent detects conflicts between sections. Adjust prompts and examples as needed.

// Fluxo de teste
1. Carregar documento-piloto.docx
2. Executar comando: "Rever documento-piloto.docx"
3. Receber output estruturado e marcar feedback

Step 7: Iterate and adjust confidence/temperature

Adjust model parameters such as temperature (lower to get more deterministic responses) and length limit. Add more before/after examples if the agent makes incorrect or overly creative suggestions.

// Sugestão de parâmetros
temperature: 0.2  // respostas mais fiáveis
max_tokens: 800

Verify the result

Confirm that the agent produces a report per document with: a clear summary, a list of issues with locations (line/paragraph), alternative suggested texts and indication of legal references. Ask two human reviewers to evaluate 10 pilot documents and calculate the acceptance rate of suggestions (>80% is a good initial target).

Conclusion

After validating the policy review assistant in Microsoft Copilot, you can expand it to other document collections, automate review requests and integrate with approval workflows. Try lowering the temperature further for greater consistency and add organization‑specific rules. Tip: want a ready-to-paste prompt example for Copilot Studio?