Document Agents Data Flow​

This article describes the data flow for Document Agents.

Overview

Templafy’s Document Agents function as a governed orchestration layer that transforms user intent into branded, compliant documents by coordinating AI models, enterprise knowledge, and administrative controls. From a functional perspective, the system operates across interconnected layers that manage user interaction, governance, model execution, data grounding, and analytics. 

How it Works

1. The process begins in the conversational user experience in PowerPoint in the Templafy add-in, under Find Template --> Create tab. A user provides a natural language prompt, such as a request to create a proposal or report, and may optionally upload files (e.g. PDF, Word, PowerPoint, Excel) to provide additional knowledge and context to the prompt.
2. Within the agent configuration layer, policy and governance controls shape how the document will be produced. Administrators define pre-configured instructions based on enterprise best practices, manage approved templates, enforce tone-of-voice guidelines, and define required structure and formatting for specific document types. These controls ensure that every generated document adheres to organizational standards for branding, compliance, and consistency.
3. At the core of the system is the synthesis engine, which orchestrates the full data flow. The synthesis layer combines user intent, governance instructions, structural requirements, and contextual knowledge into a structured, model-ready request. It acts as a controlled intermediary between enterprise systems and AI models, ensuring that models operate within clearly defined boundaries.
   • To enrich generation with relevant information, Custom Agents can retrieve contextual knowledge from enterprise and third-party sources. This may include content from SharePoint, OneDrive, external connectors, or applications such as Glean. In addition, Document Agents leverage admin-controlled assets and static content libraries managed within Templafy to ensure approved messaging and brand elements are incorporated. 
4. Once the prompts are assembled, they are routed to Large Language Model (LLM). Templafy uses LLM's from OpenAI, hosted on Azure, and also supports customers self-hosting their own OpenAI model. Moreover, it is possible to connect any self-hosted LLM for knowledge/retrieval augmented generation of text. This model-agnostic design allows organizations to align generation with their AI strategy while maintaining consistent governance controls. The model performs the generative reasoning and returns structured content to the synthesis layer, where formatting rules, templates, and brand assets are applied to produce the final PowerPoint document. It is a branded, compliant, and structurally consistent document generated within enterprise-defined constraints and enriched with trusted contextual knowledge.
5. In parallel, the system captures operational metadata to provide analytics and insights for administrators. These insights support governance oversight, adoption tracking, and measurement of business impact without exposing underlying document content.

Functionally, Templafy’s Document Agents help organizations to maintain control and transparency. The conversational interface captures intent, the governance layer defines structure and policy, the synthesis engine orchestrates data and model interaction, AI models generate content within constraints, enterprise systems provide contextual grounding, and analytics deliver observability. Together, these components create a controlled AI document generation environment that embeds compliance, brand integrity, and enterprise knowledge directly into the document creation workflow.