Copilot Studio
Copilot Studio grounding: the three mistakes I see most
Most custom agents fail not because of bad topic design, but because of how their knowledge sources are structured. Here's what to fix first.
Grounding is the agent
Topic design gets the headlines, but in practice the quality of a Copilot Studio agent is dominated by the quality of its grounding. Three mistakes show up over and over.
Mistake 1: Pointing at the whole SharePoint
A "kitchen sink" knowledge source dilutes retrieval. Be ruthless: a single, curated document library beats a tenant-wide site every time.
Mistake 2: PDFs with no text layer
Scanned PDFs without OCR are invisible to the agent. Run them through a document intelligence step first, or convert to native Word/Markdown.
Mistake 3: No freshness signal
Knowledge sources don't tell users "this is from 2022." Build a small Power Automate flow to date-stamp documents and surface the date in the agent's response.
A simple rule
If you can't explain in one sentence which document your agent will pull from for a given question, your grounding isn't ready.
