Every release of the IQ Chat platform, in the order it reached customers. Dated from the release notes, written for the people who use and run the workspace.
Administrators can give a new account a password directly, with no invite email, and can set a first password for someone whose invite is still pending. The old invite link then stops working.
Accounts whose sign-in belongs to an identity provider or to SCIM keep their password there, and the console offers no password action for them.
The model picker marks the workspace default. People who never chose a model of their own now follow the default when an administrator changes it, and a deliberate choice of another model is still kept.
When web search is not available on the chosen response tier, the assistant names the tier that has it instead of asking for a setting that does not exist.
On the public demo, every suggested prompt works as clicked, the web-search prompts search the web, and guests no longer see controls a free account does not have.
Use a whole linked SharePoint folder for one message instead of attaching its files ten at a time. A file that is already in the chat is not attached a second time.
The assistant reads an email attachment whole, from Outlook or Gmail, through the person's own connected account, and can copy an email's attachments into a project.
In a project, the assistant can run an analysis over every copied file at once, in the background for up to an hour, and deliver the files the analysis wrote to the person who started it.
Chat ratings (early access): now and then the chat asks how a conversation is going, at most once a week per person. Administrators see the answers, and no message content is stored with a rating.
Agents as their own Microsoft Teams apps (early access): an agent gets its own name and icon in Teams with no Microsoft 365 license, published by an administrator or sent to the organization's Teams administrator for approval.
The public demo opens with prompts a guest can run and previews of what an account adds, and a guest's chats carry over into a new free account.
Reading a long CSV file whole covers every row, where it used to stop at the first 1,000.
Data protection (early access): sensitivity labels on chats, documents, generated files and projects; detection of sensitive information such as card numbers, keys and passwords; and policies that audit, warn, mask, block or limit which AI providers may receive a message. Every policy starts in test mode and acts on nothing until an administrator enforces it.
Microsoft Purview sensitivity labels on OneDrive and SharePoint files can be mapped onto workspace labels (early access, with Data protection).
Scanned PDFs are read. Pages with no text layer go through optical character recognition, tables come back as tables, and each page lists the readings it was unsure of so figures can be checked.
The assistant reads linked SharePoint folders itself, with the viewer's own Microsoft 365 access, and can copy a folder into a project to list, search and read the copies. Every answer keeps the list of files it read, with their versions.
Microsoft 365 search follows the conversation: a follow-up searches for what was asked about earlier, a pasted SharePoint link limits the search to that folder, and Teams search reads further.
The assistant says what it searched and what came back, and no longer guesses at causes such as indexing or permissions it cannot see.
Projects, simplified: the sidebar names the project, a band over the chat shows who is in it, and Instructions, Files and Memory open from one place.
Projects live in the sidebar. Opening one makes the whole window the project: its chats, a panel for instructions, files and memory beside the chat, and a color the project's editors choose.
Shared projects (early access): every chat in a shared project is the team's, each person's question uses their own connected apps, project files open and download for everyone in the project, and a link that no longer opens says why.
Long chats cost less. History is trimmed in occasional steps instead of on every turn, so the provider's prompt cache keeps working between them.
Move a private chat into the team in place, mention colleagues with @, write messages to the team that the assistant does not answer, and see who is typing.
Add from SharePoint lists every site a person's Microsoft 365 account can reach, not only the sites they follow.
Every agent has one page (overview, activity, approvals, routines, memory and setup), and a new agent is created in a wizard that drafts each step from a description.
Shared projects (early access): share a project with colleagues as Can edit or Can chat, with shared instructions, files and project memory, an activity list, and a history of instruction versions.
Agents that work on their own (early access) can do something later or on a schedule when their owner asks, and the files they make while answering in Teams are attached to the reply.
Administrators can switch the personal assistant and other working agents on or off separately.
Agents in Teams pick up a message within seconds, turn a colleague's @name into a real mention, show tables as tables, and send a long answer as several messages.
Chats that hold many documents have a context ceiling: older documents become excerpts that stay searchable, and the assistant can reopen one whole when a question needs it.
A connected app that never answers no longer freezes an agent's run: each tool call has a deadline, and a run is stopped at its time budget.
In a Teams chat with other people in it, an agent answers only with what colleagues' questions may use, never its owner's mail, calendar, files or memory.
SharePoint is its own entry in Documents. Link a SharePoint folder and Documents shows it live and copies nothing: each person browses it with their own Microsoft 365 account, so SharePoint's permissions apply as they are.
The workspace's light follows the time of day, with a greeting to match. Preferences > Appearance > Time of day keeps it steady instead.
Long chats with many PDFs no longer fail with a generic error, and a chat upload over 23 MB is refused with a message that names the limit.
Skip on the first-run tour puts the tour away until the next sign-in, as its text promises.
The personal assistant's page shows everything that waits for you, including your other agents' approvals, with Approve and Reject in place, and opens with its morning brief.
Tell the assistant what to always report and what never to bother you with, in your own words. Its looks and its morning brief follow those rules.
Ask an agent about its own work: its recent runs, its routines, the approvals waiting for you, and the tools its settings keep off. A run that failed or was cancelled can be retried.
The light theme has the depth the dark theme has, with the same palette.
The assistant writes in its owner's language and gives times in the owner's time zone.
The personal assistant (early access): one per person. On a schedule you set, it looks through the apps you connected, reports what needs you, and writes a morning brief. A look only reads; it never sends or changes anything.
Agents that work on their own (early access) are set up in one step: describe the job, and the draft includes routines on real schedules, the tools, and the approval rule.
A new product tour shows only what the workspace offers, and offers people who already took it a short tour of the steps added since.
Custom apps and automations can use every connector the workspace offers them, and MCP connectors show their server's own icon.
Every custom app is served with the workspace's own design kit, so apps look like part of the product. Apps built before the kit keep their own look.
An app's color changes in one tap with no rebuild, from six tones. This replaces the earlier "Try another look".
Admin > Insights shows which features and screens people use and the path each person took through the app. The data stays in the workspace's own database, never includes what people type, their chats or their files, is kept 180 days, and administrators can turn it off.
The Help Center explains how feature use is measured, in all eight languages.
About seventy kinds of broken or malicious files (Office zip bombs, images that are not images, truncated photos, mixed encodings, mislabeled formats) are now refused with a plain reason instead of failing the turn. Password-protected PDF, Word and Excel files say so.
Large spreadsheets and CSV files load in a fraction of the time and memory. A 40 MB CSV went from over two minutes to two seconds. Arabic in Windows-1256 and UTF-16 text without a byte order mark decode correctly.
A capability matrix records what every model the product can call accepts, re-checked nightly, so a model change can no longer silently break a feature.
Structured deck creation failed for a week on one newer model that rejects a forced tool call. The planner now adapts to each model, and the failure is logged with its real cause.
Exporting a long conversation to Word is four times faster. Documents search and the admin conversation view no longer slow down as the library and the conversation grow.
Asking for a tool by name, in any shipped language, reaches that tool. Decks written with code get the company template. Chat attachments are accepted up to 23 MB from a device; larger files go through Documents.
Regional Arabic interface languages for Saudi Arabia, the United Arab Emirates and Egypt. All 67 Help Center guides and the demo content were rewritten in professional Modern Standard Arabic.
Production workspaces can run on the client's own domain, with the platform address kept as a fallback until the domain is verified.
Conversation search no longer times out on very common words. Images sent to a model are fitted to every provider's limits. A model round that loses its connection before any text arrives is sent again once.
Administrators can limit Google sign-in to their own domains. Public sign-up and guest chat can require a Cloudflare Turnstile check.
Word and PDF downloads of a conversation or a reply embed the real images, and structured decks can use pictures from the chat.
Edit an uploaded photo by asking: fix, restyle or clean up an image, or change one small region of a larger picture.
Escape takes a message back: it stops the reply, removes it from the screen and puts your prompt back in the message box.
The admin panel was reorganized by topic, and the Help Center follows the new layout in every language. Child workspaces can switch web search, image generation and dictation on or off for themselves.
Demo workspaces open with realistic documents, templates, Skills, scheduled tasks and automations for their industry, translated into seven languages. Nothing seeded ever runs by itself.
Connector failures name the connector, the real cause and what to do, instead of a generic error. One-hour prompt-cache writes are priced at the provider's real rate.
A directory of the people in the workspace: typo-tolerant search, teams, locations, expertise, reporting lines, local time, and thanks between colleagues. Each person edits their own page or lets the assistant draft it from their own words.
Fill in a profile from Microsoft 365, and sync the directory from Microsoft Entra ID or Google Workspace. Birthdays are opt-in, day and month only.
Chat can answer who someone is, who their manager is, where they work and who knows a topic, for members who may see the directory.
System and test accounts are left out of People by default, and every directory action is audited without content.
Six permissions per folder (view, download, upload, edit, delete, manage), granted to everyone, a role or a person, and inherited down the tree.
Browse OneDrive, SharePoint and Google Drive from Documents and copy files into the library. Members can share their own folders, and administrators review every member share.
The audit log records which documents the assistant read for an answer, by id and title only, and who opened each shared link.
An agent can work on its own: routines on a schedule or on new mail, background tasks, memory, and approvals for anything that reaches people outside the organization. An agent can have its own Microsoft 365 account and answer in Teams.
Chat widgets: an answer can carry a small live card with figures, charts, tables, checklists and controls. Sliders and toggles recompute in the browser with no model call.
Internal passes such as the tool router and chat titles can run on the newest small models where a parity check showed the same decisions at a tenth of the price.
The App Builder takes a sketch, screenshot or mockup and a CSV, Excel or JSON file. Pick a layout, try another look in a second, and point at an element to edit just that.
Public share links, off until a workspace administrator allows them. Continue a shared chat as your own. Save an answer to Documents as Word, PDF or Markdown.
Formulas are typeset in answers. Images from other websites load only when you ask, so that site cannot see who reads a conversation.
App builds moved to a faster default model: every benchmark app passed its first check in under a minute against a median of four and a half minutes before. Builds are metered from real token usage.
A reply that falls into a loop is stopped. A model's raw reasoning never appears as the answer. A stopped reply records its provider and tokens.
Long conversations read their history from the provider's prompt cache instead of paying to send it again on every turn.
Changes administrators make to models, providers and settings reach every instance within about two seconds. The admin panel reads a fixed number of queries whatever the workspace size.
A chat reopens on the model it was last used with. Read aloud and uploads finish within one deadline instead of timing out at the edge.
59 task guides in eight languages, shown to each reader according to their role, and a chat tool that answers "how do I" questions from them.
Every member gets a personal space in Documents with automatic Uploaded and Generated folders. Files attached to chats and files the assistant makes land there.
Administrators can rename users. Accounts provisioned by SCIM keep SCIM as the owner of their username and email.
Administrators publish the organization's documents (up to 50 MB each, 2,000 per workspace) in folders. Members browse, search full text, and chat cites numbered passages with the document and folder.
Before replying, the model looks at pages of the file it generated. Every Word and Excel deliverable passes a quality check (heading styles, list markers, formula errors, numbers stored as text) with one repair round. Generated files name the requesting user as author.
A workspace default language. Every outgoing email, in all eight languages, is sent in the recipient's language.
Full-text conversation search matches words in any order. Replies cut off at the output limit get one recovery round. PDF exports keep Cyrillic, Polish, German and Arabic text.
A native Anthropic, OpenAI or Perplexity model with no explicit backup fails over to its exact twin on OpenRouter, so an outage at one provider does not stop the conversation.
Frontier models get a 32k output budget per round, with one continuation when a reply is cut off. Stop interrupts a quiet generation within about a second.
Sign out ends this device only. Preferences > Security lists signed-in devices with "Sign out of all devices".
Administrators are notified of new self-service accounts. Signing in to an account whose email was never confirmed says so and offers to resend the link.
Every log line is redacted of credentials, keys and email addresses, and no user search terms are logged. Every audit row links to its request trace.
Chat turns start about half a second sooner, and the chat shell opens in a third of the time under load. The newest OpenAI GPT-6 models are supported.
An OpenAI-compatible API with personal keys and prepaid credit. Requests follow the user's role and model allow-list, with token and cost reporting per request.
Sign-ins and refused attempts, and what people do (chats, shares, exports, deletions, uploads, agents, apps, connectors), are in the workspace audit log, with categories and filters. Platform operators' changes are logged in the platform's own log, not the customer's.
Feature switches for child workspaces: a platform default each workspace can override.
A release gate runs about 8,100 tests on every build before it can deploy. Code-generated files can use numpy, pandas and matplotlib. Company PowerPoint decks follow the template's layouts and keep-out areas.
CSV attachments reach the model exactly as written. GPT models create files again after a strict-schema change at the provider.
The side-panel editor and the standalone document editor were retired. Generated files are now revised by asking in the conversation, which keeps every version in one place.
Automations, off by default: a workflow builder and an unattended executor with activation review. Scheduled tasks can run on selected days at a set time.
Every chat tool goes through the per-turn router, so Microsoft 365 and Google Workspace searches run only when a question needs them. Clarifying questions are offered on every response tier.
A request too large for the model no longer breaks the chat. Cold starts no longer stall requests for two minutes.
Scheduled tasks that run as the assistant and post a completion notification. Two-way voice mode with interruption.
Hostname-aware single sign-on, passkeys, enforced two-factor authentication, delegated administration, a public guest principal with its own limits, and self-service registration with bot protection.
A third presentation mode, template by default and brand-guided on request. Deck planning recovers from oversized sources and invisible limits.
Centralized model packages. Job-function roles with allow-lists. An automated data-retention engine. Project knowledge with retrieval. Structured data views. Text to speech. Native PDF processing with Anthropic models.
The App Builder runner. Reusable Skills. Streaming that resumes after a dropped connection, with background-generation notifications. The billing and plan catalog. Server-side session revocation.
The multi-tenant architecture with one database per workspace. The PowerPoint deck engine. The isolated Python runner. Plan-driven Microsoft 365 search. Cryptographic invite tokens.
The MCP runner and tool catalog. Persistent chat artifacts. SCIM 2.0 provisioning. Provider failover. Legacy Office document uploads.
Microsoft 365 and Google Workspace connectors with a query planner. Token usage telemetry and the pricing engine behind usage and cost reporting.
Adapters for the frontier model providers, support-ticket dispatch, and keep-alive heartbeats on long streams.
The core platform API reached general availability, four weeks after the first build.
A guided pilot runs IQ Chat on one workflow your team already repeats, with your models, your context and your controls.