Copilot Studio Technical Guide: Fixing SharePoint Regressions & "Get Items" Bugs

If your Copilot Studio bot recently broke after months of stability, you are experiencing the same regression issues as many other developers.

Users report that workflows running perfectly for 5+ months have suddenly failed. The consensus in the community is that a backend model update (likely forcing a shift from GPT-4.0 to GPT-4.1) has degraded the AI's ability to handle structured tasks.

Here is a breakdown of exactly what is broken and the technical workarounds to fix it.

The 3 Critical Issues

1. "Dynamic Chain" Regression

The core issue lies in the "Dynamic Chain/Fill" feature. Previously, the AI could reliably take a user’s input and generate a correct OData filter for SharePoint.

• The Bug: Now, the AI frequently fails to generate the filter string, passing a blank query to SharePoint.

• The Result: The bot returns nothing, even when data exists.

2. Native Action Limits & Hallucinations

Relying on out-of-the-box Copilot actions is dangerous for production.

• The 100-Row Wall: Native "Get items" actions often default to scanning only the top 100 rows.

• Forced Hallucination: When the bot cannot find data in those 100 rows, it refuses to say "I don't know." Instead, it often makes up (hallucinates) fake file names or data to satisfy the user.

3. Broken "Post in Teams" & Marketing Hype

Users feel like "paid beta testers" because basic advertised features do not work.

• Example: The native "Post in a Teams chat" action is currently reporting failures where it drops input parameters completely, sending empty messages.

Technical Survival Guide: Move from AI to "Hard Logic"

To fix this, you must stop relying on Copilot's "Knowledge Sources" and AI guessing. You need deterministic logic using Power Automate.

1. Architecture: Topics + Cloud Flows (Not Knowledge Sources)

Stop using "Knowledge Sources" to connect to SharePoint. They are currently too unreliable and uncontrollable.

• The Fix: Build your bot using specific Topics that trigger Agent Cloud Flows.

• Why: This gives you full control over the inputs and outputs, bypassing the buggy "Dynamic Chain" AI.

2. Data Retrieval: Enable Pagination

If you use Power Automate to fetch data, you must manually bypass the default limits.

• Step: Inside your Flow, go to the "Get items" action settings.

• Action: Turn on Pagination.

• Value: Set the threshold to 5000 (or your required limit).

• Result: This prevents the "missing data" errors that cause AI hallucinations.

3. Input Handling: System.Activity.Text & Regex

Do not let the AI interpret variables. It is too prone to errors with the current model regression.

• Method A (Raw Input): Use the system variable System.Activity.Text combined with a Parse Value node. This captures exactly what the user typed, bypassing the AI's intent extraction layer.

• Method B (Regex): For specific IDs or codes, use Regular Expressions in your Topic node to force a strict match (e.g., extracting an Order ID).

4. The AI Builder Workaround

If you simply must use AI to generate flexible queries (e.g., complex natural language searches), do not use Copilot's native processing.

• The Fix: Use the "Create text with GPT using a prompt" (AI Builder) action inside your Cloud Flow.

• Prompt Engineering Tip: You must be explicit to avoid errors. In your prompt, instruct the model: "Use OData Version 4.0 format" and explicitly define the column names.

• Workflow: Pass the string generated by AI Builder into the "Filter Query" field of your SharePoint "Get items" step.

Summary

The "Magic" features of Copilot Studio are currently unstable for structured data.

The Solution:

1. Abandon Native "Get Items" and "Knowledge Sources."

2. Adopt Power Automate Flows with Pagination enabled.

3. Enforce input capture using System.Activity.Text and Regex.