One platform default used to power everything, whether it suited the job or not. Now you choose.

Set it once, or fuss over every AOP. Each AOP has a new model field that quietly follows your Master Orchestrator unless you tell it otherwise.

Effort decides how hard the model thinks before it opens its mouth. Available everywhere except GPT 4.1.

Priority processing is a fast lane for when latency matters. OpenAI GPT 5.5 only, for now.

Receipts. The model that actually ran is stamped onto every run, so Run History, the Debugging console and evaluation runs stay honest even after you change the setting.

How model configuration works →

Consent screens irritate a lot of people. There is now an Enterprise-Managed Authorization section under Settings → MCP server & A2A. Switch it on and Leena trusts tokens issued by your own identity provider, so users are signed in silently.

You point us at your JWKS URL, which checks token signatures, and your Issuer, so nothing else gets through. Optional claim mappings for subject, email and employee ID. Switch it off and everyone goes back to the standard sign-in screen, with your configuration saved for later. Needs MCP and A2A enabled.

Set it up →

• Find & Replace in the AOP instructions editor behaves itself now: highlighting, overlapping dialogs, scroll-to-match, jump-to-top
• Tools scoped to an AI Colleague appear in its Tools tab even before you have used them in an AOP
• MCP actions come with descriptions, plus a searchable View all side sheet
• Slack-triggered runs show the channel conversation the AI Colleague actually picked up, summarised and expandable, in both the Debugging console and Run History

An expired token used to take a sync down without a word, and you would find out days later.

Set an expiry date, pick a warning window between 1 and 90 days, and choose who gets nagged. From that point it is a daily email until someone deals with it. Previously only Confluence and SharePoint offered this; it is now available across the connector library.

Turn on expiry reminders →

"Unknown error" was doing a lot of heavy lifting.

Click a Failed status now and you get a real explanation: whether the problem is your credentials, the source system, or us, written in plain language, with the next step attached. Even empty syncs and data from legacy connectors have something useful to say for themselves.

See what the failure panel shows →

Once an article had been used in a test run it vanished from the picker permanently, which made "fix the content, then re-test it" quietly impossible. Prior use is now just an icon and an optional filter.

So: re-test after a fix, re-run a benchmark set after a model change, build overlapping sets. Revolutionary stuff.

Smart Testing →

KM General Settings now asks a simpler question: fast and reliable, or slower and highly accurate? That is the whole decision.

The WorkLM toggle is gone from the interface, and your existing configuration is untouched. The model underneath both options also got an upgrade, so answer quality improved across every bot without anyone touching a setting.

Model selection →

New default OCR engine, and a long list of things that stopped going wrong:

• Content no longer arrives out of order or goes missing, and low-confidence blocks are kept rather than binned
• Scanned PDFs, rotated pages, right-to-left languages and complex tables all behave
• Special characters survive in non-English content, and hidden HTML stays out of your articles
• Password-protected PDFs say so, instead of silently producing an empty draft
• Images in answers are better chosen, with exact caption matches winning

Known limitations and how to structure articles →

HRSD-style scripted criteria are now honoured. ServiceNow evaluates its own rules, we enforce the outcome at query time at both knowledge base and article level, and deny always wins. We never read or store your script contents.

• Connector setup only warns about unsaved changes when there are, in fact, unsaved changes
• UAT and Staging bots show production-inherited settings as read-only rather than throwing an error
• Article expiry dates stopped showing a day early in US time zones
• One enormous sync no longer blocks every other bot's sync, and several stuck-in-syncing cases are cleared
• Source document URLs are captured across every connector, so View source and View Policy show up regardless of origin
• Webview highlighting scrolls the cited passage into view, and tables highlight only the rows that matter

The old NLP service missed a lot and needed constant keyword-mapping to stay useful. It is gone.

Accuracy up to 90%, on both email tickets and live chat.

Tell it about your business. Free-text instructions per department steer predictions towards how you actually work. It layers on top of the standard logic, so you are nudging behaviour, not rebuilding it.

Stop maintaining keyword maps. That whole chore largely goes away.

There is a safety net underneath. Prediction can only choose from your department's active category tree, so it cannot invent a category you never created. And if it fails or comes back empty, the ticket is still raised with the department's fallback category. Nobody gets stuck behind a classifier having a bad day.

English, Arabic, Hindi, Spanish, Portuguese, French, Korean, Japanese, Chinese, Dutch, German, Hungarian, Italian, Polish, Russian, Thai, Vietnamese, Tagalog, Somali, Haitian Creole, Czech, Romanian, Slovak, Slovenian and Tamil.

Language configurations →

A new Inactive User Report lives under Main Settings → Dashboard Settings → Inactive User Report. It surfaces users who are missing from the Employee Master, or sitting in it marked as Terminated.

One list, instead of finding out when a ticket sits unassigned for a week because its owner left in March.

Inactive User Report →

Deactivating a user now requires clearing their dependencies first: active tickets, SLA policies, reports, and any other Helpdesk configuration pointing at them.

Yes, this makes deactivation slower than one click. That is the point. The alternative was discovering the mess a fortnight later, in the form of a breached SLA.

Inactive User Management →

When an admin approved someone's unavailability and picked a replacement, only new tickets went to the replacement. Everything already open stayed with the person on leave. You can now optionally hand over their existing active tickets too.

One catch: tickets do not automatically come back when the unavailability period ends.

Comments now save themselves as you type. Navigate to another ticket, leave Helpdesk, or have your laptop die mid-sentence, and the draft is still waiting for you.

• Pending-on-user nudges are recorded in the ticket timeline, so you can see who was chased and when
• Reports got faster
• Watcher email details are now visible in Bot View

Query Clusters is a new tab inside the Handled and Not handled drill-downs. Rather than scrolling through thousands of individual requests, you get themes, and how much volume each one drives.

Each row is a cluster with an auto-generated name, the knowledge base articles the NLU reached for, the tools and skills it invoked, and the volume behind it. Click through for every individual query inside that cluster, with user and timestamp, searchable and exportable.

Start with the Not handled tab. It is a ranked list of everything your AI Colleagues are failing at, ordered by how much it is costing you. That is a roadmap written by your own employees. The Handled tab is the mirror image: which themes are being resolved, and what is doing the resolving.

Under the hood it groups by meaning rather than keywords, so "how do I apply for leave" and "leave application process" land together. Clusters name themselves. Anything that does not fit goes to a Miscellaneous bucket instead of being forced somewhere wrong, and once enough similar strays pile up they graduate into a cluster of their own. Greetings, filler and test messages are filtered out so your volumes are not flattered. Everything syncs hourly.

Query Clusters →

Big exports used to be a hostage situation. Keep the tab open. Do not navigate away. Hope it does not time out at minute nine.

Now it runs in the background and emails you when it is done.

Before you submit, pick your date range and up to 10 recipients. The dialog checks whether that range actually contains anything, so you find out now rather than after opening an empty file.

When it is ready, recipients get a secure download link that stays live for 7 days.

And it remembers. Every export sits in one place with its status, searchable and filterable, refreshing itself while something is still cooking. Failed ones can be retried, finished ones re-run, both pre-filled from the original request.

Google Drive, Gmail, Slack, HubSpot and Rovo, each connecting in a few clicks. Because they are MCP-based, the agent discovers every tool the app offers the moment you link it. No hand-building actions one at a time, and anything those apps add later shows up on its own.

A connector for the privileged-access and identity platform. IT and security flows can now handle privileged-account and access tasks through the agent instead of bouncing someone into a separate console. A real unlock for automations where access has to be governed and audited.

HR, payroll and workforce data from Ceridian Dayforce, available to your AI Colleagues and workflows. Employees can self-serve the everyday questions about pay, time and personal details inside the assistant, and automations can act on Dayforce data without anyone exporting a spreadsheet.

If the AI needed a data asset that did not exist, you had to abandon the flow, go build it in the Workflows Admin console, and come back to start again.

It can now create DOA tables, Master Tables and Data Models mid-generation, either because you asked ("create a DOA for expense approvals by department and amount") or because it worked out the workflow needs one. Everything it creates is put in front of you for review before it saves, and duplicate names are blocked, so nothing appears behind your back or collides with what you already have.

Generating workflows with AI →

You can now stop a generation you already know is going nowhere, instead of watching it finish out of politeness. It cuts off immediately, with no broken interface and nothing orphaned behind it. Re-running means pasting the prompt back in, which is a small price. Works in both Edit via AI and full generation.

Unpublished workflow apps and AI-generated chats can now be deleted. They used to just accumulate. Deeply unglamorous, long overdue, and genuinely nice if you iterate fast.

Workvivo is now a channel, set up from Global Settings like any other. Setup guide →

• Slack — Notifications arrive with their attachments in the same thread, and you can see the bot thinking while it works
• WhatsApp — Messages arrive in the right order with formatting intact, and Tags are visible and filterable in logs
• Google Chat — Tables in agent responses render properly
• SMS — Delivery status in analytics now matches what actually reached the handset
• Links that open desktop apps are clickable again

• Users can rate a voice call
• Internal system messages stopped appearing in transcripts
• Email addresses render correctly in transcripts, and enabling voice on an autonomous agent no longer errors out

• Studio Assist degrades gracefully on long tasks instead of going silent, plus a round of interface improvements
• Quick replies show up reliably. Most were being generated and then dropped before anyone saw them
• Run explanations are based on the complete picture rather than partial results
• Duplicate responses that were blocking people mid-conversation are fixed
• Streaming and multi-chat are on by default for new bots, small talk frequency is respected, and run counts are labelled correctly in charts

• Exports show which channel each notification went out on
• Unread counts stopped counting things you had already read

• No more being logged out after stepping away
• The unified dashboard works on mobile, and user names are editable
• Audit log filters load on first open, and invalid date ranges are rejected
• Name sorting is no longer case-sensitive, and column sorting is consistent
• Scrolling and pagination fixed across log tables, and duplicate SendGrid domains blocked
• Dashboard login fixed on iOS, and dashboard users work across regions

Apps — Login stopped offering bots that cannot sign you in, first load is faster, and an iOS layout issue is fixed.

Reliability — A daily crash caused by oversized report attachments, fixed. A redirect loop that was taking parts of the platform down, fixed. Integrations reconnect on their own when a connection drops mid-session. Scheduled jobs stopped running twice, and small file uploads work again.

Accessibility — Another round of fixes across the dashboard.