After updating to the latest cloud release, AI-driven case routing in our case management system is failing for multi-geo deployments. Cases created in our EMEA and APAC regions are being incorrectly routed or not routed at all, causing significant SLA breaches.
Our setup uses AI Builder models trained on historical case data to predict the best team assignment. The routing worked flawlessly in our North America environment but started failing after we expanded to multiple geographies. Cases from non-US regions either get routed to the wrong teams or sit in an unassigned state.
We’ve verified that:
- AI models are published and active in all regions
- Regional teams are configured in the case management module
- Locale settings match each region’s requirements
The error logs show:
AI Prediction Error: Model training data mismatch
Region: EMEA, Locale: en-GB
Expected locale: en-US
It appears the AI model expects US-formatted data but receives region-specific formats for dates, currencies, and text fields. Has anyone successfully configured AI-based routing for multi-geo deployments? What’s the proper approach for handling region-specific training data and locale configurations?
This is a known challenge with AI Builder models in multi-region deployments. The AI model training is locale-specific, and a model trained on en-US data won’t perform well with en-GB or other locale inputs. You need to either train separate models for each region or implement a data normalization layer that converts all inputs to the training locale before prediction.
I’d recommend the normalization approach since maintaining multiple models is operationally complex. Create a Power Automate flow that preprocesses case data before sending it to the AI model.
For your volume distribution, a single unified model is feasible but requires careful data preparation. When combining regional data for training, you must normalize all locale-specific fields to a common format and add a ‘region’ feature to the model inputs. This allows the model to learn region-specific patterns while maintaining a single deployment.
Critically, ensure your training dataset is balanced across regions. If North America is overrepresented, the model will be biased toward NA patterns. Use stratified sampling to create a balanced training set even if actual case volumes are skewed. Also, retrain the model quarterly with fresh data from all regions to maintain accuracy as patterns evolve.
Thanks Maria. We’re considering the normalization approach, but I’m concerned about the latency impact. Our SLA requires case routing within 30 seconds of case creation. If we add a preprocessing flow, will that delay the routing decision? Also, how do we handle text fields where locale differences aren’t just formatting (like different terminology or languages)?
For latency concerns, implement the normalization as a synchronous plugin rather than an asynchronous flow. Register it on the case create event before the AI routing step. This adds minimal overhead (typically <200ms).
For text fields with language differences, you’ll need to either use multilingual AI models or implement translation as part of the preprocessing. Azure Cognitive Services Translator can be called inline and adds about 300-500ms depending on text length. Just be aware of the cost implications if you’re processing high volumes of cases.