🎯 Why this matters

Your customer just said: “Can we build one agent that answers questions across all our projects — even when those projects share nothing in common?” They have design docs in SharePoint, decision records in repos, and meeting notes from Issue #006’s pipeline. Each data source is a silo. No single agent connects them today.

This is the build. A Copilot Studio agent backed by Azure AI Search with a custom skillset that stamps every document with cross-project metadata. The agent itself is trivially simple — generative answers pointed at the index. All the intelligence lives in the enrichment pipeline. Build the hard part once, get a working agent on Day 1.

Type 🔧 Practical Build Build by Tyson Dowd Stack Copilot Studio + AI Search + Power Automate Paired concept Issue #007 · Holographic Memory Published May 20, 2026

🧭 The four components, in build order

Each numbered step below maps to a single thing you build. The order is intentional: the index schema comes first because everything else depends on it, but the Power Automate flow is what you spend the most time tuning.

1️⃣ Azure AI Search IndexCustom schema with project_name, decisions, related_artifacts, summary_context.
2️⃣ Power Automate FlowSharePoint trigger → Azure OpenAI enrichment → push to AI Search REST API.
3️⃣ Copilot Studio AgentGenerative answers + custom cross-project system prompt. Zero custom code.
4️⃣ Plugin Action“Index this conversation” — turns chat exports into searchable knowledge.
🟦 Source Artifacts BREADTH OVER PRECISION
📄
SharePoint Sites
5+ project sites — design docs, decision records, architecture specs.
📦
Repo Artifacts
READMEs, ADRs, PR descriptions across project repositories.
🎙️
Meeting Transcripts
From Issue #006 pipeline — approved meeting summaries.
💬
Chat Exports
Manually pasted via the plugin action (Step 4).
Trigger on create/modify
2 ⚡ Power Automate Flow ENRICHMENT
SharePoint Trigger
Fires automatically when a file is created or modified in any monitored site.
📝
Extract Text
SharePoint/Word connector pulls content. Strip formatting; keep semantic structure.
🧬
Azure OpenAI Skill
Single prompt extracts project, decisions, related artifacts, summary context.
📤
Push to AI Search
POST enriched document to Azure AI Search REST API (/docs/index).
Enriched documents (JSON)
1 🟦 Azure AI Search Index HYBRID INDEX
🗂️
Custom Index Schema
Fields beyond standard content: project, author, decisions, related artifacts, summary context.
🛠️
Custom Skillset (alternative)
Web API skill calls Azure OpenAI during indexing — alternative to the Power Automate path.
🔎
Hybrid Retrieval
Keyword + vector + metadata filters. search_filter scopes by project on demand.
Top-N results + metadata
3 🟪 Copilot Studio Agent GENERATIVE
🪄
Build Steps
Create agent → add AI Search knowledge → enable generative answers → paste prompt.
📝
Custom System Prompt
Instructs the agent to synthesize across projects and cite both sources.
⚙️
Zero Custom Code
No code inside Copilot Studio. The agent is a thin conversational layer.
Cross-project answer
🟩 Output USER VALUE
💬
Copilot Chat
User asks a cross-project question and gets a synthesized answer.
📎
Source Citations
Every answer cites documents from both projects — building reader trust.
+ User-fed knowledge loop
4 🧩 Plugin Action — “Index This Conversation” USER-FED KNOWLEDGE
🧑‍💻
User Workflow
User exports a transcript, pastes it, invokes the plugin action.
Power Automate Cloud Flow
Parses conversation, runs Step 2 enrichment, indexes into the same instance.
🔁
Closes the Loop
Users actively feed the system — memory grows even without formal documents.

🏆 Where this pattern lands

The short answer: any team that has 3+ concurrent projects with implicit cross-project decisions

This is the build behind Issue #007’s Holographic Memory pattern. It hits hardest in organizations where institutional knowledge is high-value, project-bounded, and currently invisible to AI tools. Lead with the “impossible question” demo — it’s the moment the room goes quiet.

🧑‍💼 Professional services & consulting

“What lessons from the Acme engagement apply to our new Baker project?” Cross-engagement knowledge reuse without senior partner availability.

⚙️ Engineering & R&D

“Why did Team A choose OAuth PKCE and does that affect Team B’s compliance approach?” Architecture decision reuse across squads.

🏛️ Regulated industries

“Show me all decisions across all projects that affect our Q4 audit scope.” Compliance traceability without manual cross-referencing.

🤝 M&A and integration

Rapidly connecting knowledge bases from two merging organizations where nothing was linked before.

🎓 New hire onboarding

“Catch me up on everything that happened on this project before I joined.” The agent synthesizes from meeting notes, design docs, and decision records simultaneously.

What Copilot Studio can’t do (yet). The agent can’t follow related_artifacts links hop-by-hop like a graph database. It retrieves the top-N search results and synthesizes from those. For deeper traversal, add a pre-query Power Automate flow that expands related documents before passing them to generative answers. Also: the search_filter parameter in the Azure AI Search connector lets you scope queries by project or metadata — use this to build topic-specific variants of the same agent.

🛠️ Build the demo in one day

Get a working agent on the customer call this week

  1. Pick two SharePoint sites from two different projects. 10–15 documents per site — enough to have cross-project overlap in people, technologies, or decisions. Don’t over-engineer the selection — any two real projects will have implicit connections.
  2. Manually enrich 20 documents for the demo. For each doc, add a JSON metadata file with project_name, decisions, related_artifacts, summary_context. Yes, this is manual for the demo — the Power Automate pipeline automates it in production. ~2 minutes per doc with Copilot Chat helping you write the metadata.
  3. Create the index and the agent. Index the 20 enriched docs. Create a Copilot Studio agent, add Azure AI Search as a knowledge source, enable generative answers, paste in the cross-project system prompt. Test with: “What decisions from Project Alpha affect Project Beta?”
  4. Demo the “impossible” question. Ask a question that no single document answers — one that requires connecting a decision in Project A with a constraint in Project B. If the agent synthesizes an answer citing both projects, you’ve demonstrated the pattern. If not, the related_artifacts metadata needs richer cross-references.
💡 The line that opens doors: “What if a new hire could ask one question and get an answer that connects decisions across every project your team has ever run — with sources cited from documents that were never explicitly linked?”
Build pattern by Tyson Dowd, architected from a 1:1 discussion on March 24, 2026 about the “Windows 7 phase of AI” — where the differentiator is no longer raw model capability but how systems carry memory across tools, projects, and time. This is the practical build for Issue #007 — Holographic Memory. It also pairs with Issue #002 — Scoped Multi-Source Search (retrieval pattern) and Issue #006 — Meeting-to-Knowledge Pipeline (meeting transcript source).