Product Newsletter — October 2026

October was about building by conversation and testing without the legwork. Forge, the assistant you knew as Studio Assist, now builds AI Colleagues, AOPs and tools from a description, and Code tools cover the logic nothing else could. Workflows Studio can test itself. Voice 2.0 speaks your employees' language and switches mid-call. Knowledge admins can bring their own test questions and finally tell when an article is actually live. And there are three dates worth putting in your calendar.

The short version


📅 Three dates for your calendar

Nothing here breaks on its own. Each one goes more smoothly if you act early.

WhenWhat changesWhat to do
30 November 2026GPT 4.1, Sonnet 4.6 and the Opus 4 family (4.5, 4.6 and 4.7) retire. Anything still on them moves automatically: GPT 4.1 to GPT 5.6 Terra, Sonnet 4.6 to Sonnet 5.5 and the Opus models to Opus 5.5.Nothing, if the replacement suits you. Prefer a different model? Switch before the 30th. Retiring models →
January 2027Voice 1.0 retires and voice 2.0 becomes the default for everyone.Switch your UAT assistant to voice 2.0 now and test your real use cases. Switch to voice 2.0 →
February 2027Slack retires its Assistant experience, the one with separate Chat and History tabs.Book a change window with your CSM before the end of January. Plan the Slack move →

🧠 AI Colleagues Studio

Studio Assist learned to build, and got a new name to go with it.

Forge: describe it, approve it, done

In July, Studio Assist could look but not touch. It investigated runs and explained what went wrong, and that was it. Forge is the same assistant with its hands untied. Tell it what you need and it builds AI Colleagues, AOPs, helper AOPs and tools, plans bigger requests step by step, and reuses whatever already exists on your bot.

Open it from the Forge icon in the top bar, from Create via Forge on the AOPs page, or with Cmd + J (Ctrl + J on Windows).

You sayForge
"Build an onboarding procedure for the HR colleague, and create any tools it needs"Plans the AOP and its tools, waits for your go-ahead, then builds them in order
"Why did this run fail?"Reads the run's full evidence trail and explains the root cause, with a timeline
"How many runs failed in the last 7 days?"Answers with charts and tables instead of raw logs
"Run the onboarding procedure every Monday at 9am"Sets up the Workbench schedule for you
"Is Benefits enrollment ready to move to prod?"Runs readiness checks and opens the migration screen beside the chat
"Did last night's Confluence sync work?"Checks your knowledge connectors, and can start a sync or open the connector form
"Is the PII guardrail enabled?"Tells you, and opens the right settings page so you can change it
👍

You stay in charge

Forge shows a plan before it builds anything and asks before sensitive steps, such as inviting a user. Settings are always changed by you, on the settings page. It works only on the bot you are in and only with the permissions you already have, and every change it makes is recorded in Audit Logs under your name.

It also answers "how do I" questions from Leena AI documentation, keeps separate chats with full history, and takes attachments, so you can hand it the SOP instead of retyping it. Forge is in beta.

Meet Forge →

Code tools, for logic that doesn't fit a form

Some jobs don't map neatly onto an API call or a workflow: combining data from two systems, reshaping a response, running a calculation. Until now, that usually meant building a workflow just to host a few lines of logic.

A Code tool is a small, purpose-built tool for exactly those cases. Describe it to Forge, and Forge writes it, tests it in a sandbox and publishes it to your Tools list. From then on it behaves like any other tool: map it to AI Colleagues, reference it in AOP instructions, and control where it runs.

  • Connections stay swappable. Point each system the tool uses at a different connection without touching the tool itself.
  • Settings live outside the tool. URLs and other values sit in shared variable groups under Integrations, so one change updates every tool that reads them.
  • Moving to production is guided. You map each connection and variable to its production value, and sensitive values are never shown.

Code tools are created through Forge only. Editing them through Forge is coming soon.

How Code tools work →

Five new models, and four on their way out

ModelProviderBest for
GPT 6 AstraOpenAI, Azure OpenAIFrontier reasoning for the hardest agentic work
GPT 6.1 SolOpenAI, Azure OpenAIAstra-class reasoning for everyday production work, at a fraction of the cost
GPT 5.6 TerraOpenAI, Azure OpenAIBalanced, lower-cost everyday production workloads
Opus 5.5Anthropic, Amazon BedrockThe latest Opus, stronger on agentic and knowledge work
Sonnet 5.5Anthropic, Amazon BedrockCapable everyday agent work at Sonnet pricing

New models carry a New badge in the picker. Not sure which to choose? Compare models, in the picker header, opens benchmarks on cost, speed and accuracy.

Retiring on 30 November 2026: GPT 4.1 moves to GPT 5.6 Terra, Sonnet 4.6 to Sonnet 5.5 and Opus 4.5, 4.6 and 4.7 move to Opus 5.5. They already show an Available until 30 November 2026 badge, and hovering over it tells you the replacement. AI Colleagues Studio admins and builders will get an email too.

📘

Nothing breaks on the 30th

Any Master Orchestrator or AOP still set to a retiring model simply runs on its replacement. If you'd rather choose a different model yourself, switch before the date.

Model configuration → · Compare the benchmarks →


⚙️ Workflows Studio

The end of "proxy in, click through, check the report, repeat".

Workflows that test themselves

Proxy in as a bot user. Trigger the workflow. Fill in every input and approve every approval by hand. Open the report. Do it all again for the next scenario, and again after every fix. Coverage stayed low, go-lives stayed slow, and issues turned up late.

1. Build a test case. For each node, choose Actual, which runs exactly as it would in production, or Mock, which uses values you define, with no integration calls and no notifications.

NodeWhat you can mock
Triggers and inputsThe acting user and the form data
ApprovalsThe approver and their decision
Actions and functionsThe status and the output
Async callbacksThe callback payload
DelaysSkip them entirely

2. Decide what "pass" means. The expected item status, expectations on specific nodes (skipped, executed, approved, rejected and so on), or custom conditions on data-centre values.

3. Or let AI write them. Click Auto generate, optionally add a prompt or upload a BRD or SOW, and AI drafts test cases aimed at the scenarios you haven't covered yet.

4. Run and review. Run one case, a selection, or every Ready case. Each run gets its own ID and a passed, failed, error and in-progress summary. Reports look like regular workflow reports, with a Mock tag on mocked nodes and an expected-versus-actual breakdown for every condition. If a case is waiting on a real input or approval, submit it on someone's behalf, reassign it or nudge them, right from the report. Test runs never touch production data.

5. Keep them current. Test cases belong to a workflow version and carry over to new ones. Change the workflow and its test cases move to Error, so nothing stale slips through. Your mocks are kept: review, re-save, and you're back to Ready to Test.

🚧

Actual means actual

Nodes left on Actual really call your integrations and really notify people. Mock anything you don't want touching a live system. Testing isn't available for utility workflows or List Views.

Find it under Workflows Studio → open a workflow → Proceed to testing.

Automated testing → · Create test cases → · Read the reports → · Test cases across versions →


🎙️ Voice

Voice 2.0 is live, and it's multilingual.

Voice 2.0: it speaks their language

Voice 2.0 runs calls on a native speech-to-speech model. Instead of turning speech into text, working out an answer, and turning it back into speech, it listens and replies in one step. You hear the difference in how natural it sounds and in how it copes with being interrupted. Any assistant that turns voice on now starts on voice 2.0.

Voice 1.0Voice 2.0
LanguagesEnglishThe caller's language, switching mid-call if they do
ConversationSpeech converted to text and backNative speech, lifelike voices, natural interruptions
VoicesExisting voice library10 new voices
New settingsNoneVoice model instructions and a speech detection threshold
Call historyNot availableIn AI Colleagues Studio
StatusRetiring by January 2027The default for every new voice setup

Call history lists every voice 2.0 call, on the web and on the phone: who called, how long it lasted, how it ended, and which AI Colleague runs it triggered. Search by caller, and filter by status or by source.

Making the switch. Go to Admin Console → Orchestrator → Voice → ⋮ → Switch to voice 2.0. The switch is one-way, so start with your UAT assistant and test your real use cases. Happy with the results? Switch production straight away. January is a deadline, not a waiting period.

🚧

Two things to check first

Phone calls. Voice 2.0 phone calls currently need a Leena AI Twilio number. If you use your own Twilio account, your phone calls stay on voice 1.0 for now. Your SPOC can help you plan around it.

Your assistant's setup. Voice 2.0 works with assistants running on AI Colleagues. If yours is still on the older autonomous agent setup, it needs to move to AI Colleagues first.

Voice settings and UAT checklist → · Voice 1.0 and 2.0 side by side → · Phone calls on each version → · Call history →


📚 Knowledge Management

Fewer "why isn't it answering?" tickets, and a UAT environment that's actually yours.

Bring your own test questions

Most teams arrive with a list of questions their knowledge base must answer, or a set of failed cases they want to re-run until they pass. Smart Testing only ran questions that AI generated from selected articles, so checking your own list meant proxying in as an employee and ticking answers off in a spreadsheet.

Now you can run your own list:

  1. Download the template.
  2. Fill it in with the test user, the question and the expected answer.
  3. Upload it as a CSV, XLS or XLSX file.
  4. Check the rows. You see exactly which rows are valid before anything runs.
  5. Get the report. The same accuracy report you get for AI-generated tests.

Close UAT against your own acceptance criteria, in a single pass.

Run your own test cases →

Know when an article is actually live

"Published" used to promise more than it delivered. It meant the article had been downloaded and parsed, not that the assistant could answer from it yet. So admins would publish, test straight away, get nothing back, and raise a ticket.

Published articles now show a Processing icon while they're being prepared, and a Processing Failed icon if that step goes wrong. No icon means ready, which keeps the board clean. A system-managed Live tag lets you filter for articles the assistant can actually use.

See the new article states →

Expiry reminders that don't give up

A single "X days before expiry" email was easy to miss, and one missed email meant an article quietly expired. It is now a recurring schedule in Dashboard Settings. Choose when the first reminder goes out, how often it repeats, and the most it should ever send. Each article's own expiry date drives its schedule, and reminders stop as soon as the owner updates the date or the cap is reached. Your existing setting carries over automatically.

UAT is yours now

As announced earlier, Production, UAT and Staging now run independently, with no automatic syncing back from Production. Test in UAT without your changes being overwritten.

Moving a setup up a level got clearer too. Clone settings is now called Pull, and it tells you upfront that both connection settings and parsing configuration will be copied. Connections you've already pulled stop cluttering the "add new" list, and each one gets its own option to re-pull.

Environments in Knowledge Management →

Conflict Analysis: a summary, and a Stop button

Open a conflict group and a summary now shows the total conflicts, the articles affected, and the split across Critical, High, Medium and Low. Filter by severity, conflict type and date, search by document name, and your filters carry into exports. Started a run by mistake, or one that's far too big? Stop it from the dashboard instead of waiting it out.

Conflict Analysis →

Connectors, and everything else in Knowledge Management
What changedWhy you care
ServiceNow custom templatesKnowledge articles built on custom templates are picked up automatically, with no per-template handling
SFTP with SSH keysSFTP connections now support SSH-key sign-in, the standard many enterprise servers require, and stored credentials are encrypted
Connections through ForgeAsk Forge to add or edit a knowledge connection, and it opens the setup wizard right beside the chat
Partial parses get a second goIf an image or table in a document fails to process, the article is flagged and re-parsed automatically. The gap used to stay, silently
Images in answersIn-line images in answers are now on by default for bots using the GPT 5.4 model
Clearer sync errorsArticle sync errors that used to say "unknown error" now fall into clearer categories, so you can tell whether an issue is on our side or yours

💬 Channels & Employee Web App

Slack is moving, WhatsApp got friendlier, and the Web App has a new front door.

Slack: time to move to the Agent experience

Slack is retiring its Assistant messaging experience in February 2027. If your employees see Leena AI in Slack with separate Chat and History tabs, your app needs to move to Slack's new Agent experience before then. Leena AI has supported it since the 6 October release, so there is nothing to rebuild.

What employees get. Leena AI moves into the regular Slack Messages tab, like any other direct message. Each conversation runs in its own thread, personalised suggested prompts sit at the top, and a stop button appears while Leena AI is answering. One refresh of Slack and they're there.

How the switch works.

  1. Book a change window with your Leena AI CSM.
  2. Your Slack admin updates the app in Slack's app settings: one button, then a small change to the events the app listens for. About 10 minutes.
  3. Leena AI switches your bot in the same window, because the change isn't picked up automatically.
  4. You test it together with a quick conversation.
🚧

One-way, with a prerequisite

Slack doesn't allow switching back. The Agent experience also needs Generative AI (AI Colleagues) responses and response streaming turned on for your bot. Plan your window before the end of January 2027.

The step-by-step guide →

WhatsApp that feels like WhatsApp

  • Read receipts and typing. Employees see blue ticks and "typing…" as soon as their message arrives, and the indicator stays up while an AI Colleague works through a longer request.
  • Picks up after sign-in. On assistants that need sign-in, the employee's message is answered once they've signed in, instead of the conversation starting over.
  • Re-engagement in their language. Prompts go out in the employee's language when the template has been approved in that language.

The WhatsApp experience →

A new sign-in for the Web App

The Web App login has been refreshed in your company's brand colours. The verification code is now a single field that accepts paste and autofill and submits on its own. The wording is clearer throughout, and translated into 49 languages. And the phone country code is now pre-selected from the employee's location, instead of always defaulting to India.

Readable on any brand colour

Text on brand-coloured areas, like an employee's chat bubble or a button, now switches between black and white automatically, picking whichever is easier to read against your colour. There's nothing to configure.

If your brand colour is on the lighter side, you may notice black text where it used to be white, because white on a light colour is hard to read. Darker brand colours stay white.

The result: the Web App's accessibility score is now 99.98 out of 100, up from 93.82 in May, with zero critical issues. Worth knowing if WCAG reviews are part of your procurement process. This applies to both the current and the new Web App layout.

Goodbye, Outlook and Telegram

The Outlook and Telegram channels are retired. They can no longer be set up or chosen in channel settings, the Flow Dashboard or notifications, and new messages on them aren't processed. Past conversations and analytics stay available.


🔔 Notifications

One flow, one preview, and a dress rehearsal before the real thing.

A guided notification builder

Creating a notification now follows the same three steps for Bot, Email, SMS and WhatsApp: Add details → Set schedule → Review & test, with a live preview of what employees will receive.

What's newWhy you care
Test sendSend it to a single employee from the review step, before it goes to everyone
Duplicate anythingCopy any notification, whatever its status, and reuse it
Audiences without losing your placeThe audience page opens in a new tab, and the new audience is selected automatically once you save it
Attachments on long descriptionsAdd files to longer messages
Voice notesRecord one inline, right in the details step
WhatsApp placeholdersMap them inline, and the mapping is saved on the template for next time
Send nowEmail and SMS can go out immediately, not only on a schedule

Create a notification →


🏛️ Admin Console & Dashboard

Safer sign-in for admins, and fewer surprises once they're in.

Dashboard SSO, one identity provider per domain

The dashboard used to support a single SSO app. You can now connect several identity providers and route dashboard users to the right one by their email domain, with a default app for everyone else. Handy after an acquisition, or anywhere contractors sign in differently from employees.

Setting up SSO is safer too. Save an app as a draft, test the sign-in before it goes live, then turn SSO on as a separate, confirmed step, so you know it works before anyone depends on it.

The Authentication page has been redesigned, with clearer password policy and IP restriction settings. IP changes apply only when you click Save, and typos or duplicate ranges are caught before you do. And opening a workspace with a shorter session length now applies that limit straight away.

Dashboard authentication →

No more surprise sign-outs

In the last hour before your dashboard session ends, a banner warns you, with a Re-authenticate option that keeps your work intact. Whatever you were in the middle of stays put.

How the expiry warning works →

A friendlier front door to the dashboard

Sign-in, password reset and account activation have all been redesigned. When someone sets a password, a live checklist shows which rules they've met as they type, so nobody has to guess what "strong enough" means.

Employee sync knows who your HRs are

Automated employee syncs have a new Automatically identify HRs setting. Turn it on, and anyone assigned as the HR of an active employee is flagged as HR on their profile after every successful sync. When HR assignments change in your source system, the flag follows, with nobody maintaining it by hand.

A safety check holds the update back if too few HR assignments in your data match a real employee, so a bad feed can't wipe everyone's HR flag. Because the sync owns the flag, it can't be edited on an individual profile.

Employee sync setup →

Audit logs, on for everyone

Audit logs are now on by default for every bot, under Admin console → Logs, with no separate enablement. They're visible to System Admins and Audit & compliance managers. And they now cover AI Colleagues Studio too: actions by builders and admins are recorded, including changes made through Forge.

Audit logs →

Smaller changes worth knowing
  • The dashboard remembers where you were. It reopens the last module you used, for each user, and the modules have been reordered around AI Colleagues.
  • Maintenance banners can wait. They now have a Remind me later option, plus a few UI fixes.
  • Redesigned authentication pages for Leena channels, with a new layout and better flow: Web authentication, Apps authentication, Other auth settings and SAML apps.
  • Uploads that delete themselves. The Leena file upload service can now remove a document automatically after a set duration. Ask your SPOC if you'd like to use it.

🔢 October, by the numbers

99.98

Web App accessibility score out of 100, up from 93.82 in May. Zero critical issues.

8

Jobs Forge can do, from building AOPs to debugging runs to moving them to production.

5

New models in the picker. Four older ones retire on 15 November.

10

New voices for voice 2.0.

49

Languages the new Web App sign-in speaks.

~10 min

Of your Slack admin's time to move to the Agent experience.


👋 That's October

Want a walkthrough of Forge, help switching to voice 2.0, or a Slack change window booked? Your Leena SPOC is the person to ask: your Implementation PM if you are still onboarding, your CSM once you are live. Leena AI support is always there too.