Hi Kuldeep — taking these in order.
1. Bring your own AI gateway
We currently support Bring Your Own Key, but there are no plans to support customers bringing their own AI gateway.
2. Why upgrading a Now Assist plugin breaks things
The cause is structural, not something you're doing wrong. Now Assist doesn't ship with the platform — it's a set of Store apps on their own release cycle. A platform patch upgrades the core release and leaves those Store apps behind. That version drift is what breaks things, and because the symptoms show up as UI failures and missing artifacts, it reads like a configuration problem rather than a version mismatch.
Support's "upgrade everything" advice is directionally right but blunt. The more precise move is to update at the product level, not the plugin level. Request and update the parent app your instance is entitled to — Now Assist for ITSM, for HRSD, for Creator — and let it pull its dependencies and entitlements down with it. Picking individual child plugins off the list is how you end up with a half-matched dependency graph, which is usually worse than not updating at all.
The habit change: treat Now Assist Store app updates as a mandatory follow-up to every platform patch, not an optional one. Sync Now in Application Manager, update the parent product app, then check Upgrade History for skipped changes before you call it done. Sub-prod first. That sequence prevents most of what you're describing.
3. Predictive Intelligence / Task Intelligence
No, PI and TI are not being deprecated, and yes, you should keep using them. They solve a different problem than Now Assist does. PI is classic supervised and unsupervised machine learning — classification, similarity, prediction from your historical data. Now Assist is generative: content creation, summarization, conversation, reasoning. Those are complements, not substitutes.
The deprecation chatter you may have seen is about specific surfaces rather than the capability. The Predictive Intelligence Workbench has been deprecated since Yokohama, and the Regression Framework was deprecated with Washington, with engineering rebuilding the regression pipeline for better accuracy and re-releasing it first through Task Intelligence. Neither of those means the ML layer is going away.
The architectural argument for keeping PI in the mix: where you have volume and labeled history, a trained classifier is cheaper, faster, and far more testable than routing every categorization decision through an LLM call. The strongest designs I see put the deterministic classification and routing work in PI, and reserve LLM reasoning for the genuinely ambiguous branches. You get better unit economics and a much easier system to evaluate.
4. Moveworks integration
Otto is ServiceNow's new offering, built off Moveworks infrastructure and completely integrated with the ServiceNow AI Platform. It can be first experienced in ServiceNow EmployeeWorks and the AI Control Tower, and will roll out across all products in the year ahead.
