Copilot Studio agents

Questions like “who do I ask about this?” went back and forth in chats and email. Procedures were spread across SharePoint, and the answer depended on who you asked.

I kept behavior separate from knowledge. The instructions hold guardrails, citation rules and routing. The facts stay in the connected document library, so the answers change when the documents change.

Every answer cites a source. The agents only give guidance: they don't approve anything, change records or submit financial data.

A test question asking who managed a specific site came back wrong. There were three causes:

• The instructions had grown past the platform's 8,000-character limit without any warning
• Content inside Excel files wasn't being indexed reliably
• There was no routing or alias rule, so a site's short code and its city name looked like different things

I cut the instructions from 19,437 characters to 7,725, added document-first routing and alias rules, and the lookup returned the right answer.

I wrote 18 structured test cases, 24 demo questions and a 50-question bank covering assignments, deadlines, accruals and cash flow. Not all 50 pass yet. The bank is there so the team can keep measuring after my internship.

• A PowerPoint agent that turns one Excel workbook into a branded 21-slide deck, following a versioned brand and build spec
• A VS Code agent that builds a new Power BI report or rebrands an existing one from a plain-English request, generating the model, a date table and dozens of measures
• Written brand standards for Power BI, PowerPoint and HTML, so people and agents follow the same rules