What shipped.

Connect your Slack workspace under Settings → Integrations and Iris brings the work to you. “Page the team” posts every handoff and SLA breach to the channels you choose — the customer’s message, why it escalated, a deep link straight to the conversation, and an @mention of whoever it’s assigned to, plus any specific teammates you pick to notify on that channel. “Reply from Slack” lets a teammate answer the customer without leaving the thread: type “@Iris reply ‹your message›” and it goes to the customer as you, through the normal reply path — ordinary Slack chatter in the thread never reaches them, so your team can talk freely and only a deliberate reply is sent. Each feature shows whether it’s actually running, connecting takes one click with no app-store review, and only teammates whose Slack account is linked can be @mentioned.

While writing an internal note, type “@slack” and pick a teammate — Iris DMs that person on Slack with your note and a link back to the conversation. It’s a separate, deliberate gesture from the normal @name mention (which still just notifies them inside Iris), and the composer shows exactly who the note is being sent to before you send it. Turn it on per workspace under Settings → Integrations → Slack; it’s free, off by default, and built to work the same way for the tools we add next.

A creator asked why their product was flagged; the AI never looked it up and invented a plausible-sounding reason, while the real one — written by the review team, in plain words — sat on the product the whole time. Now, when a customer asks about the state of THEIR account or content, the AI checks the actual record first and its reply must cite exactly what it read: the send is allowed only after the system verifies every stated fact — the right item, the right kind of fact — against what the lookup actually returned in that same exchange. A fact the record doesn’t carry can’t be asserted, one item’s record never licenses claims about another, and an empty result licenses only “there are none” — never a diagnosis. Every verified send (and every blocked attempt, with the suppressed text) is visible to your team in the thread’s notes.

A reply slipped out claiming “I have scheduled a callback for you today at 4:30 PM” — nothing in the system schedules calls, and the old keyword safety-net missed the phrasing entirely. Every reply the AI composes is now read by a model before it sends, and its claims are receipt-checked: “I’ve told the team” only survives if the conversation was genuinely assigned; “I’ve done it” only with a real completed action behind it; a promised call, a scheduled appointment, or a predicted outcome the system can’t guarantee is replaced with an honest line and the thread lands in a teammate’s queue. Statements about a customer’s account state send only from capabilities that actually read the account, and timeline promises (“refunds take 5–7 days”) only when they come from your own approved knowledge. If the claim check itself is ever unavailable, the reply is held for the team rather than risked.

When Iris said “I’ve shared this with the team”, the thread was only assigned if a routing rule matched — otherwise it sat in nobody’s queue while the customer waited on a promise. The assignment now happens BEFORE the line is sent, every time: your routing rules decide first; otherwise the least-busy teammate on shift gets it; and if nobody’s on shift, the first teammate back owns it — with the customer told the truth: “they’re offline right now, you’ll hear back after 10:00 tomorrow”, computed from your team’s actual hours, time off and holidays. Customer messages never name individual teammates.

The Team screen opens with a Now strip: who’s on shift (and until when), who’s away, who’s on leave, who’s back tomorrow — with Iris itself pinned first, always on. Each member’s panel sets their working hours (follow workspace hours, custom weekly windows — split shifts and overnights included — or always available), specific-date overrides (“this Saturday 11–15, covering”), and time off — adding a leave that starts now offers to hand open conversations to a teammate so nothing sits unanswered. Anyone can set themselves away for an hour, the day, or until a date (14-day cap); recurring hours are owner/admin. Workspace holidays live once in Settings → Business hours → Closed dates. A Coverage-this-week card shows humans-on-shift per hour and flags the gaps before they happen — the same availability Iris reads when it hands off, so what you see is what it does.

The Reports teammate table now opens with an Iris AI row: conversations handled, closed, resolution rate, p90 first reply, replies per close, CX from AI-handled rated chats, and the same 28-day activity texture — computed from the same definitions as the “resolved on its own” headline above it, so the two can never disagree. It’s a measurement row, not a team member: it can’t be assigned conversations and has no hours to audit. Seeing “p90 first reply: <1m” next to human rows is the point — it shows exactly where the AI carries the load.

A customer who fires three messages in a row used to risk three stacked replies, each answering one fragment. Iris now waits a few seconds for the burst to land and answers once, as the newest message, with everything they said in context — including a reply that was already being composed when another message arrived: it steps back and lets the fuller answer through. Your own handover rules still check every message in the burst, and nothing about single messages changes.

Several small gates used to decide things about a customer’s message by matching English keywords: is this a greeting, a thanks, a payment question, a follow-up? Those checks ate typos (“how do ayouts work?” sailed past a filter its correctly-spelled twin tripped), went blind in Hindi or any other language, and sometimes chose silence. All of them now share one comprehension step: a fast model reads each inbound message in the context of the conversation — what kind of message is this, is it about this customer’s own account or a general how-does-it-work, does it touch money or legal, what language is it in — and the rest of the system acts on that reading. Meaning survives typos, works in every language your customers write in, and a question about how something works is answered from your knowledge instead of being mistaken for a sensitive account matter. The old keyword checks remain only as a fallback for the rare moment the reading is unavailable.

Iris has always routed a customer’s own payment problems and legal matters to a person, even when it knew the answer — a safety default that was invisible and unchangeable. It’s now a Safety rules card on the Replies screen: “Payment & legal questions go to a person” can be turned off by an owner or admin if you’d rather Iris answer what your knowledge covers. General questions like “how do payouts work” always answer either way — the hold only ever applied to a customer’s own money or a legal matter — and every flip is recorded in the workspace audit log. The identity rule (Iris never claims to be human) sits on the same card and is not configurable.

When a customer’s question had no answer in your knowledge, Iris said so once — but a NEW unanswerable question later in the same chat could be met with silence, because the “once per conversation” guard treated it as a duplicate. Every substantive customer message now gets either an answer or an honest expectation line; if several questions in a row have no coverage, Iris says once more that it’s one for the team, then stays quiet rather than repeating itself.

When Iris answered a customer on its own, the thread stayed bold in the inbox — unread only cleared on a human reply, so the list kept demanding attention for work that was already done. An AI answer now clears the unread marker exactly like a teammate’s reply — and only when it answered the latest customer message, so a customer who writes again at that same moment keeps their bold row.

The article editor’s image button used to ask for a URL — fine for engineers, useless for everyone else. It now opens a normal file picker and uploads the image for you, and pasting a screenshot straight into the body works too. Images land on your workspace’s storage and appear inline as you write.

Trusted mode was drafting correct answers to broad questions like “tell me about AutoDM” — the sending rule only trusted a reply whose single top source was a brain page you’d confirmed, so your own published help articles never counted, and one unproven draft mixed into the retrieval blocked everything beneath it. The rule is now the honest one: when every excerpt the AI actually answered from re-checks, at that exact moment, as your live knowledge — a published article or a confirmed/proven brain page, at the exact revision the AI read — the answer sends in Trusted. If anything unproven, imported, stale, or delivery-suspended is in the mix, it still drafts for your team. Tool-assisted replies and proposed actions keep drafting, exactly as before.

A customer asking “tell me more” or “and how many can I add?” after an answer used to be treated as a brand-new question — the words alone matched nothing, so the AI went quiet or reached for account tools it didn’t need. Short follow-ups are now read in the context of what the customer just asked: when the bare words find nothing, the AI re-searches your knowledge with the previous question attached and answers from the page that actually covers it — and only keeps that contextual result when it genuinely grounds better. Account tools are reserved for account-specific state, not a detour for questions your pages already answer.

Yesterday’s safety fix (which stopped tool-assisted replies from borrowing a page’s authority) was quietly over-broad: because account tools are enabled, EVERY reply took the tool path — so even plain answers from pages you’d confirmed were arriving as drafts, and customers got “a teammate will reply” for questions the AI knew cold. Now a reply that used no tools and no proposed action, grounded on a page you confirmed, sends on its own again — while anything that actually touched a tool still drafts until you approve. The safety fix stays; only its overreach is gone.

New knowledge used to appear days late: a draft was only written after a close AND a three-day observation window. Learning is now split from judging — the draft appears the moment the conversation closes (marked “Unverified answer”), and the trust verdict still lands after the same careful three-day watch: confirmed resolutions upgrade the badge and earn credit toward going live; if the customer came back unhappy, an untouched draft quietly retracts itself — anything you edited is never auto-deleted, it gets a review note instead. Reopened-and-reclosed conversations can finally learn again too (an old bookkeeping flag used to block them forever).

A chat where the team or AI answered and the customer stayed silent used to sit “open” for two days before the cleanup swept it. Bubble chats ending on our reply now close after ~12 quiet hours — the rating card reaches people the same day, the open list stops counting finished work, and a new message still reopens instantly. Email keeps the longer window; assigned chats, handoffs and anything pending are never auto-closed.

The agent now checks meaning before writing a new draft from a conversation: a re-statement of something it already learned updates that draft instead of minting a look-alike page, and a page’s name — including the names of pages merged into it — is unique in your workspace, enforced at the database, not by good intentions. For pages that already overlap, a meaning-based sweep nominates likely duplicate pairs for you: a “possible duplicate” chip in the Brain list, a count in Needs your eye, and a review block on the page with two honest actions — “Keep that one” merges this page into the survivor you pick (its name becomes an alias so the old spelling can never come back as a twin; the survivor’s content is untouched), “Not a duplicate” dismisses the pairing. When several pages chain together, the drawer names the strongest survivor across the group — and says when it’s only seen part of a large group instead of pretending. Nothing ever merges on a score alone; every merge is your call.

The three attention cards now carry a full-width texture of their daily movement under the number. Charts start where your data starts — a workspace that went live mid-period no longer shows weeks of empty axis with everything squeezed into a corner (the legend says “since” when trimmed). “Why customers are unhappy” fits in half the height: five one-line quotes per column that expand in place, and “View all ratings” opens a side drawer with its own scroll instead of stretching the page. The Daily Brief can be hidden for the day with one tap — tomorrow’s comes back on its own.

When the AI answers and the customer replies with a bare “thanks”, “ok” or 🙏, the conversation now closes right there and the rating card appears while they’re still in the chat — the way a person wraps up — instead of a cleanup pass closing it two days later when nobody’s around to rate. It never closes this way if a teammate is involved, the AI handed the thread to the team, money or legal came up, or anything is still pending — and it’s chat-only, not email. Saying thanks again after the close just lands quietly; a real follow-up reopens the thread instantly.

An audit caught the brief mislabeling its own data: it called customer messages “conversations”, counted a conversation as “closed today” if it was merely touched that day, narrated seven-day totals as the day’s work, and — due to a type bug — reported customer satisfaction as permanently empty while real ratings existed. Every number now carries its true name and window: new conversations, customer messages, closes stamped at closing time, the day’s own draft/sent/edited counts, satisfaction from the day’s actual ratings, and open gaps described as the running total they are. The brief also stopped drawing dramatic conclusions from measures that are only days old — it says “counting since” instead.

A reply the AI composed around an account lookup could go out on its own in Trusted mode — and claim it was “from” a confirmed Brain page — when that page merely happened to be in the retrieved context, not in the answer. Sentences a model writes around tool output now follow the same rule as everything else it composes: they draft in Trusted mode until you approve, and the line under a tool-assisted reply says “using your account tools” instead of naming an unrelated page. Page attribution is reserved for answers actually built from a page.

A rare race could shuffle a conversation: if the same message was delivered twice over the live connection, it jumped to the bottom — so your just-sent reply could appear above the customer’s older messages. Threads now keep strict time order no matter how updates arrive.

Opening a customer memory used to show knowledge-page chrome that made no sense for a person: a “not yet proven” trust badge, grading zeros, and a Confirm button for something that is never asserted as fact. A memory now opens as what it is — “used only for this person” — with its recognition line and topics readable, and just Edit and Retire. Editing a memory keeps it a memory; confirming or pinning one is refused everywhere, not just hidden.

The Reports screen is rebuilt around action. It opens with your three levers — how much work Iris actually takes off the team, the long-wait tail (the p90 truth behind the fast-looking median, with a count of customers who waited over an hour), and CSAT against the 4.0 bar — and each card jumps to its section. Then, in order: what Iris closed on its own and what stopped it (the stakes floor, handoffs and reopens side by side), which topics cost the team the most repeat work — ranked so the next thing to teach Iris is the top row — where open threads are aging, why customers are unhappy (worst-rated topics plus their words, every rating one tap from its conversation), and each teammate’s quality against their load — who needs coaching, whose workflow to copy. Every number states how it’s counted, small samples show their n, and counts that depend on the outcome events shipped 12 July say so instead of pretending.

The Brain screen has a search bar now — and it understands what you mean. “Withdrawing earnings” finds “How to Withdraw Earnings”; “take out my money” finds your settlement-timing pages. It searches every page in every state, retired ones included, so a page you shelved last month is one search away. Open any result to confirm, edit, retire — or restore a retired page right from its drawer. The Trainer searches the same way now, so asking it “do we have a page about payouts?” finds the pages a word-match used to miss.

Point the Trainer at a page — “retire the page about the old pricing”, “bring back that shipping page”, “mark this as correct” — and instead of walking you through menu steps, it lines the action up on a card you approve with one tap. It finds the page first (by meaning, in any state), shows you exactly what it’s about to do and to which page, and tells you plainly when something can’t be done — a page that’s already retired, or one that changed since you looked at it, is called out instead of silently skipped. Troubleshooting a conversation now also shows which Brain pages the AI leaned on there, so “retire the page this chat used” works too.

The per-topic Track record table is gone. It tried to say “who replies on each topic”, but a topic could show confirmed-by-you while its replies drafted, and the general row mixed everything unrelated into one meaningless number — it confused more than it explained. The rule it hid is now the whole screen: in Trusted mode, Iris sends on its own only when an answer you approved carries it — an exact pinned answer or a reply grounded in a page you confirmed. Everything it composes on its own arrives as a draft, unless you deliberately choose “Everything it’s sure of”. Two safety nets replace the old behind-the-scenes topic benching, and both are visible instead of hidden: a confirmed page whose sent replies keep needing corrections is suspended from replying on its own (it stays yours — it shows up in Needs your eye with the reason, and your edit or re-confirmation restores it), and if live replies get corrected a lot across the board, Iris’s own compositions pause to drafts for a few hours — shown on the Replies screen with the reason. Payment and legal chats still always go to a person.

A new registry under AI → Skills. Register the account facts your answers hinge on — which plan someone is on, how they pay, whether a payout is on hold — and bind each to one field of a look-up action you already have, with a declared type (yes/no, a number, one of a fixed list). Exactly what it does today, no more: it stores the typed binding, lets you Test any fact right from the screen (a one-off check against a customer id you type — used once, never stored), and — for facts you opt into Sampling — collects status-only telemetry after AI conversations. Sampling is the governed path: it resolves only for identity-verified conversation contacts via their verified sign-in, type-checks every value (never guessed), and every provider call gets an audited receipt. Replies do not consult or quote these facts yet; that’s the next step, and it will stand on this verified plumbing. Nothing is shown to customers unless you later allow it.

The Brain now has a Pinned tab showing every word-for-word answer in one place. And pinning is no longer a special door: open any page that carries an exact question and answer and choose “Make pinned” — from then on, that exact question gets that exact reply, instantly, with no AI in between. Unpin any time; the page keeps its history either way.

A pinned reply is sent word-for-word — it can’t check which plan someone is on or how they signed up. So answers that give different instructions for different customers (“if you’re on X, do this; if you’re on Y…”) can no longer be pinned: you’ll get a clear explanation and the better path — keep it as a normal page and the AI picks the right branch per customer. Existing pins like this keep working as knowledge, just not as blind word-for-word sends, and they’re flagged in Needs your eye so you can decide.

When a payment or legal question hands a conversation to your team, the AI now stays out of the follow-ups too. Before, a customer’s “ok” or “any update?” a minute later could pull the bot back in to answer the very topic it had just promised a human for. Now, once it says a teammate will reply, it means it — for the whole conversation, the expectation line is sent once and everything after waits for a person.

When a customer replies “ok”, “thanks”, “theek hai”, or just 🙏 after the conversation has been handled, a human doesn’t answer with three more sentences — and now your AI doesn’t either. Pure acknowledgements get silence and the seen indicator. Anything that could actually be an answer to a question (“sure”, “yes”, “done”) still gets full attention.

A customer who opens a second thread minutes after the first is the same person — and the AI finally treats them that way. It sees their other recent conversation for context, it won’t repeat the identical “I’ve shared this with the team” line across their threads (a varied line once, then quiet, while the thread is still assigned to your team), and a promise made in one thread carries into the next instead of being forgotten.

A brand-new visitor saying “hi” gets one short, human greeting. And that’s the only time — the AI no longer re-introduces itself mid-conversation, greets nobody twice, and never answers a greeting with a wall of text.

Two pages that say the same thing now merge into one — the survivor keeps both pages’ track records and history, so no earned trust is lost. Pages that disagree with each other get flagged side-by-side in Needs your eye with a one-tap resolution: keep the right one, retire the other. All of this holds up under heavy simultaneous learning, so a busy day can’t corrupt what your agent knows.

Learning used to wait for the cleanest closes only. Now every resolved conversation gets distilled into a draft page — and each draft wears an honest birth badge showing how its conversation actually ended, so you can tell a proven lesson from a plausible one at a glance. Drafts still never speak to customers as fact; they wait for real outcomes or your word. The practical difference: your review queue fills with useful material much faster, and so does your agent.

The private memory pages your agent keeps about returning customers were showing up in the Brain approve queue as if they were knowledge drafts — noise you had to skip past. They now live in their own Customer memories tab: browsable, never mixed into approvals, and never part of the shared knowledge your agent asserts to anyone else.

Until now, every change action — cancel, pause, resend — waited for a teammate to approve a card, every single time, forever. Each change action now has a dial: “Asks first” (the default, exactly as before) or “Run automatically”. Automatic runs are deliberately narrow: only for customers your site has identity-verified through your integration, only for actions whose every input is locked to that verified sign-in (nothing typed by the visitor), capped per day at a limit you choose, and every run is logged in the conversation thread and the audit list. Over the cap — or for any unverified visitor — it quietly falls back to the approval card. Switch back anytime.

The Reports CX card used to show comment snippets you could read but not act on. Now every reason row is clickable — it opens the conversation behind it, with who rated and whether the AI or a teammate handled it. And a new “View all ratings” view lists EVERY rating in the range, not just the ones with comments — filter by score (1–2 / 3 / 4–5), newest first, and page back as far as you like. Built to answer one question fast: who did we let down, and why?

The promise was always “a topic that proves itself starts replying on its own” — but quietly, going live still waited on an approval almost nobody saw. Fixed properly: when a topic clears its proving bar (enough drafts your team sent unchanged, at your workspace’s own thresholds), it promotes itself — recorded in Track record and the audit trail. To bench a topic back to drafts, ask your agent in the Trainer and approve the card. Demotion was always automatic when a topic slipped; promotion now keeps its side of the bargain.

Two teaching doors were still filing into the old lists: approving a “Ready to approve” card from the question queue, and applying a weekly-review suggestion. Both now create real Brain pages — exact Q→A answers delivered word-for-word, marked Confirmed by you, visible at Brain and searchable when you ask your agent “what do you know about X?”. One brain, no side drawers.

Replies produced through your agent’s tools now carry the same source chips as knowledge answers, and any Brain page that grounded a sent reply gets “used” credit on its grading record — so the Brain screen’s numbers reflect what your knowledge actually does in production. A page you confirmed carries your word on this path too: answers grounded in it go out live in Trusted mode, exactly as promised.

Opening a long page from the Brain list could land you at its bottom — title out of reach, the page hiding behind the top edge and refusing to scroll back. The drawer now always opens at the title, keeps the title and close button pinned while you scroll, and no longer inherits a scroll fling from the list behind it.

The old Deployment screen — rollout percentages, trust maps, thresholds — is gone. In its place, one rule: answers you’ve confirmed go live instantly; everything Iris came up with on its own stays a draft until it proves itself. One switch with three positions (Drafts only · Trusted answers · Everything it’s sure of), and a Track record beneath it showing who replies on each topic, with honest counts — “41 sent · 0 corrections”, never a progress bar. Payment and legal chats always go to a person, shown right there as a pinned row instead of a hidden rule. And the moment you confirm a page in the Brain, it tells you what that means: “Live — Iris now answers this itself.” Drafts never consume your paid answers — practice is free.

A new “+ Add knowledge” button on the Brain screen. Write or paste anything — a fact, a policy, a whole SOP doc — and it’s distilled into clean pages, one per topic (a long paste becomes several, never one blob). You review every page before it saves, and what you save becomes “Confirmed by you”. For whole websites or big files, Knowledge sources still do that job.

Whenever you add or edit knowledge, the brain checks it against everything it already knows and tells you what it found: “same topic — update that page instead of creating a twin”, “already covered”, or “this contradicts an older page — retire it?”. Each finding is a choice you tick, never something done behind your back. Nothing writes without your say-so, and retiring an old page is always your explicit call.

Tell your agent “we take 5% per transaction” in the Trainer and it becomes a real brain page: distilled, overlap-checked, shown as a card you approve — exactly like adding it from the Brain screen. Behavior asks (“never promise refunds yourself”) still become behavior rules; the trainer knows which is which, and splits a message that contains both. You can also ask “what do you know about refunds?” — it searches your actual brain and answers honestly, including what’s missing.

Pinned answers (word-for-word replies to exact questions) and snippets (private internal facts) have moved into the Brain as normal pages — pinned ones keep their exact-answer behavior, and both finally get what they never had: outcome grading, edit history, and one place to see everything your agent knows. Your existing ones keep working untouched.

The Brain list now pages through your whole wiki (it previously showed only the newest 50), and a Select mode lets you confirm-to-canon or retire many pages in one sweep — useful when reviewing hundreds of seeded drafts.

The inbox has always kept a view of every conversation where a teammate @mentioned you in a note — but it was buried where nobody found it. It’s now a visible @ chip right in the inbox filter row (and in the assignee menu): one tap shows the threads that need you, newest mention first. Tap again to go back.

This is the big one. Everything your agent knows now lives as a self-maintained wiki you can browse at Brain (in the left rail) — real pages for facts, policies, procedures and per-customer memories, not a flat list of articles. Each page carries a STATE that says exactly how much to trust it: a draft may quietly inform the agent’s tone or a hypothesis but is NEVER stated to a customer as fact; a proven page has earned the right to be stated; a page you confirm becomes top authority. Pages are written by the agent as it resolves conversations — you don’t hand-maintain a wiki, you curate one.

When a conversation genuinely resolves, your agent distills the reusable lesson — “how to fix X”, “our policy on Y” — into a draft brain page, generalized and stripped of any customer’s personal details. The writer is sandboxed: a bad lesson can at most create a draft that only informs, never a fact the agent will assert. The same lesson learned across several conversations merges into one page instead of piling up duplicates.

The heart of it, and what makes this different from a normal knowledge base. A draft becomes assertable only when it has actually resolved real customers cleanly — the customer’s issue closed and they didn’t come back on the same problem (a true-close, not a reopen). Frequency of use moves nothing; the agent’s confidence in itself moves nothing. Only real outcomes — or your explicit confirmation — promote a page. If a promoted answer later causes reopens, it loses trust automatically.

No page ever pretends to be more proven than it is. Each wears an honest badge: “Seeded · not yet proven live” for a page drafted from your history that hasn’t earned it yet, “Proven across N” for one that resolved N real customers, and “Confirmed by you” for one you blessed. A seeded page is a starting point, never a claim.

You don’t start empty. We replayed your past support conversations through the same learning loop, so your brain opens already full of pages distilled from what your team actually resolved — plus a memory page for every returning customer. Everything from history is labelled honestly as seeded (informs, not yet proven) until it earns its place live. The moment your AI is on, those pages start proving themselves on real customers.

Open any page in Brain and you can Edit it. Write or paste what the agent should know — and here’s the nice part: a longer entry is automatically distilled into one clean, structured page you confirm before it saves, so pasting a hundred lines never leaves a sprawling raw blob in your brain. A small one-line fix just saves as-is. A page you edit becomes “Confirmed by you”. You’re the authority on your business; the agent handles the structure.

For every customer who’s talked to you more than once, your agent now keeps a private memory page — the topics they’ve raised, past resolutions, when you first and last spoke. So a returning customer is recognized, and the agent personalizes with context instead of starting cold — while still confirming specifics rather than assuming.

Platform mechanics — how payouts, KYC and automations actually work on SuperProfile — live in a shared base brain that every workspace inherits read-only. So your agent can speak to those the day you turn it on, without you teaching it, while your own learnings stay isolated to you.

A background heartbeat now runs the whole loop on its own: it grades conversations once their outcome has settled, distills the true-closes, promotes what proved out, retires what didn’t, and keeps the brain tidy — all without you lifting a finger. It only spends effort where your AI is on; where it’s off, it simply rests.

Nightly, your agent maintains its brain: knowledge that stops getting used slowly loses standing and drops back to proving, retired outage notes are cleared, and dead links are swept. Knowledge earns its keep continuously instead of accumulating forever.

When a new workspace is created, your existing help articles and contact history are drafted into the brain automatically on day one — so the agent has a running start from whatever you already have, and grows the rest from real conversations.

Before acting on a request that touches money, an account, or anything sensitive — a refund, resending access, changing account details — your agent now confirms identity and authority first, or hands to a human. A how-to question (“how do refunds work?”) is answered freely; a do-it request is gated. It never moves money or makes an irreversible change itself.

On every substantive turn your agent now either attempts an answer or makes a visible handoff to your team — it can’t just leave a customer hanging. If it’s genuinely unsure, it says so and brings in a human, once, cleanly, rather than stalling.

Your agent will never tell a customer “done — I’ve processed that” unless there’s a verifiable record of it. If it can’t point to a real receipt, it doesn’t claim the action — it proposes it for a human to complete. No hollow confirmations.

When a customer sends a screenshot — an error, a payment page — your agent now actually looks at it and factors what it sees into the answer, instead of ignoring the attachment. It describes the image once and reuses that, so it’s cheap and fast.

When a chat closes, your agent quietly notes the resolution — the root cause, what was done, and where (even “sorted over a call”) — so the hard cases that resolve off-channel still leave something learnable behind, instead of a close with no answer in the transcript.

Two rough edges on the new Brain screen, smoothed: the page could flicker as it loaded, and on the frosted-glass appearance the detail panel showed the list bleeding through behind the text. Both fixed — the screen is steady, and the panel reads cleanly on every appearance.

When Iris fully answers a customer and there’s nothing left to do, it now wraps the conversation up on its own — closing it right after the reply so the customer immediately sees the “How did we do?” rating prompt, exactly as if a teammate had closed it. It only does this on a genuinely complete answer: never when it asked the customer a question, promised a follow-up, proposed an action for a human to approve, or handed off to your team. If the customer writes again, the chat reopens automatically. The close shows as “Closed by Iris (resolved)” in the conversation’s Activity, and these don’t count toward any teammate’s resolution rate.

The Insights → Knowledge gaps list is meant to be the FAQ you should write next. It had filled up with noise — acknowledgements (“ok”, “Thanku”), half-typed fragments (“Kya huaa sir”), phone numbers and stray links — because it logged whatever a customer typed whenever the AI couldn’t answer. Now every gap passes a quality check first: greetings, thank-yous, fragments and personal info (emails/phones/links) are filtered out and never listed, and the genuine questions are rewritten into one clean, self-contained question and de-duplicated, so ten ways of asking “how do I delete my account?” count as one. Gaps are ranked by how many DISTINCT people asked (one person retyping no longer looks like ten). And account-specific problems no article can fix — a payment not received, KYC stuck, a connection failing for that user — are split into a separate “Recurring issues” list so they don’t crowd out the FAQs. Existing lists get tidier as new questions come in.

The trust map used to lump everything under one “general” bucket. Now Iris quietly learns your real topics — payments, access, refunds, whatever your customers actually ask — with no setup and no fixed list. It groups similar questions as they come in and names a topic once a theme clearly recurs, so “Rollout & the trust map” splits into the topics that matter to you and your agent can earn (or lose) trust on each one independently. Topics appear on their own over the coming days as conversations flow in.

Press ⌘K (or Ctrl+K) from any screen — or click Search in the sidebar — and the whole product becomes a search box. Jump to any screen (Reports, Rollout & the trust map, Billing, Team…), run an action (turn AI on, run the weekly review, switch theme, sign out), or open any conversation by name, email or #number. Keyboard-first (↑↓, Enter, Esc), remembers where you go. It’s powered by the same navigation map your agent uses, so it always knows every screen.

At the top of Reports there’s now a once-a-day report the AI writes about ITSELF — an honest morning standup on yesterday: what worked, what your team had to rewrite, and the deeper pattern it sees re-reading its OWN past briefs (a topic it keeps getting wrong, a fix it proposed that never shipped, whether its clean-rate is trending). Every line is grounded in the day’s real numbers — it never invents a figure. Owner-only and off by default; turn it on in AI settings and, if you like, get the same report by email each morning. An unread brief shows a gentle dot on the agent orb; open any past day from the strip to re-read it. It’s reflection, not control — anything it suggests still goes through your approval.

The weekly review — where your team’s edits become concrete suggestions you can approve — no longer makes you wait for its cycle. Ask your agent to run it (“turn my team’s edits into suggestions”, “what can I approve to raise the clean rate?”) and it mines the last two weeks right then and shows you what it found. There’s also a “Run now” button on the review card. If there’s nothing to learn from yet, it says so honestly instead of pretending it’s blocked.

On a busy conversation toolbar (combined history, long names), the contact’s avatar could spill over the “Open/Closed” status pill. The header now holds each piece in its lane at every width.

When the AI misses a question and a teammate answers it in that same conversation, Iris now notices. Instead of asking you to retype what your team already wrote, the question arrives as a “Ready to approve” card — the teammate’s answer, generalized and attributed by name, with a link to the source conversation. One tap makes it a pinned answer plus a permanent test; you can edit it first, or choose to be asked instead. Questions nobody answered still come to you as before. Facts only ever come from you or your team — never invented.

The question queue used to refresh on its own every ~6 hours. A new “Check for new questions” button on the Trainer runs the sweep on demand and tells you what it found — or just ask your agent “any pending questions?” and it sweeps mid-chat and lists what’s waiting, including the ones your team already answered.

Your agent can now see your knowledge base — how many articles, what’s published, what’s draft — and answer from the real list instead of "I can’t read your help articles". Under "What your agent may operate", a new Help articles capability lets it propose new drafts, edits, publishing or deletion — every change as a diff you approve, with the diff saying plainly when something goes live on your public help center. One-click revert, like everything else it does.

Ask it how trust is earned, what counts as an edited draft, or how the question queue works, and it now answers with the real mechanics — the exact 90% similarity threshold for "sent unchanged", the 25-drafts-at-85%+ promotion bar, how unanswered questions become its questions to you, and how every answer you give becomes a permanent test. It also reads your live deployment state and recommends the earned-trust path first when you ask about going live — then follows your call.

The Knowledge importer accepts "www.yoursite.com" the way people actually type it, shows live progress, and the row flips from importing… to ready on its own. Sites behind bot protection (including ours) are now fetched with a real rendered browser instead of failing with an error.

Opening a second conversation while your agent was thinking used to show the thinking indicator there too — and coming back could briefly lose your question when the answer arrived. Every in-flight request is now tied to its own conversation: no leaking across sessions, and your question always stays with its answer. The thinking indicator also picked up ~150 working phrases, so a long wait never repeats itself.

The "Try me" tester beside the agent chat now closes with one click and comes back from a slim tab at the edge — and it remembers your choice.

A rare reply could end with a dangling code marker. Replies are now cleaned at the final boundary on every path — prose only, always.

Your agent greeted you by name on every message, not just the first. It now says hi once and gets on with it.

Sign out used to only clear this browser — the session itself lived on, so after logging back in, Your devices showed the old ghost still “active 8m ago”. Sign out now ends the device’s session server-side too, and it lands in the audit trail.

The reports capability grew from a 31-day overview into a real analyst. Ask “how many messages did we receive today and how many did we reply to?” — or any day, any custom range, even hour by hour — and your agent runs the actual query and answers with the exact figures. Every reports answer ends with an “Open Reports” button that takes you to the full view.

When your agent’s replies carry numbers or facts that matter, they now wear a quiet colour: counts in blue, money in green, dates and times in gold, drops and failures in red, plain key facts on a soft chip. Your eye lands on what matters without reading twice.

Every conversation’s side panel now has an Activity section: the full open/close history with who did it and when — a teammate by name, the customer’s reply reopening it, a macro, or an automatic idle close. The event lines inside the thread now name the person too (“Conversation closed by Aditya”), and the previously silent cases — auto-close, macro close, customer reopens — are recorded as well.

Plain agents used to open the Team screen to an empty table. Everyone now sees the whole team, and your own row is fully self-service: profile photo, name, an “about” line, and your email signature. The signature editor also actually opens now — the button previously did nothing.

A new deployment mode replaces the blind percentage dice: your AI answers customers directly only on topics where it has proven itself. Everywhere else it drafts the answer silently for your team — sending it unchanged builds trust, editing it counts against, and after 25 clean drafts at 85%+ a topic goes live. Stakes like payments and legal always go to a human, decided before the AI ever runs. It even benches itself: three edited answers in a day and the topic goes back to proving. Your agent can propose promotions — under the same permissions as everything else.

Two more capabilities under “What your agent may operate”: it can create and edit pre-chat workflows (ask it for a refund qualification flow — you approve the diff), and it can update the Messenger’s welcome content. It also gained an about line of its own — give your agent a title like “Business Development Executive” and it shows on its profile.

Each “Talk to your agent” session now lives in the URL — reload the page or share the link and you land back in the same conversation, exactly where you left it.

The mic in the agent chat now works the same in every mode: your voice note is transcribed into the composer for you to review and send, whether you’re chatting normally or answering a guided question. A separate button focuses the keyboard, so it never looks like voice is the only way in.

When a request to the trainer hits a network blip or a server restart, Iris now retries quietly a couple of times before saying anything — and if it still fails, you get a plain “server issue, try again shortly” with a Retry button instead of a raw error.

Every long operation in the agent chat shows an immediate, honest loading state — a small terminal-style spinner the instant you click Rehearse or Apply, with the composer locked while a rehearsal runs so half-finished states can’t be interrupted by accident.

The line charts across Reports now show the precise value and date when you hover a point — with a generous hit area, so you don’t have to aim.

The “waiting for you” count on the sidebar orb was clipping in the collapsed rail and truncating to “6 to revie…” when expanded. The collapsed dot now sits cleanly on the orb, and the expanded rail shows the full “6 to review” pill.

Settings now shows your devices — “Chrome · Mac”, when each signed in, which one is this — with per-device sign out and a “Sign out everywhere else” sweep. Signing a device out takes effect immediately: its very next action is refused. Owners and admins can also sign a teammate out of all their devices from the Team screen, and every sign-out is recorded in the audit trail.

The agent behind “Talk to your agent” has grown from trainer to manager. It can see your AI settings and propose changes to where Iris runs and how it rolls out — every change as a diff you approve, recorded with one-click undo. A new “What your agent may operate” panel on the AI screen sets each capability to Off, Propose or Auto-within-limits; billing, team roles and data deletion are never grantable, by construction. And when something isn’t its to change, it walks you there with exact steps instead of saying “I can’t”.

When your AI can’t answer real customer questions, Iris now gathers them, clusters the repeats into one clear question each, and brings them to you: a count on the agent orb, a “Waiting for you” queue on the Trainer, and at most one email digest with a preview of what’s waiting. One tap starts a guided session — it asks, you answer in plain words (typing or by voice), it proposes the change, you approve. Every answer becomes knowledge and its own permanent test, so a question you’ve answered once is never missed again. The emails are on by default and turn off in one click.

Behind the scenes, every AI request now records exactly what it used — the model, the tokens, the true cost — powering the usage line on your AI screen and the transparent per-answer billing Iris is built on. No estimates, no rounding in our favour.

When you change how your agent behaves, Iris now builds the test for that change itself — it finds real questions from your past conversations that the change affects, writes realistic ones where your history has none yet, and saves them as golden test cases automatically. They re-run on every future change, so a rule you set today stays protected forever. You’ll see them in Testing & review, marked as created by Iris — and you can still add your own.

Rehearse runs are now split into small graded batches, so long exams complete reliably and partial results survive a hiccup. If a run does fail, the Trainer says so plainly and offers a one-click Retry — no more blank scores.

When the Trainer proposes a change, the diff card now shows every rule on its own line — what was added in green, removed in red, and the rest quietly for context. Nothing is clipped, so you always see exactly what you’re approving.

The “Talk to your agent” panel could slide in on the wrong side of the screen. It now always opens on the right.

Owners and admins can now pick a team member in the inbox filter and see every conversation they’ve replied in — not just what’s assigned to them. Built for reviewing quality and coaching, one person at a time.

When your team answers while the Messenger is closed, the launcher now shows an unread count and a small preview of the message pops above it — one click opens the conversation right there.

The Needs reply and Unread filters now show how many conversations are waiting, right on the chip — alongside the Open and Closed tab counts.

Resolved chats now stay in the Messenger’s Messages list with a quiet “Resolved” mark — so when your team replies after a conversation closes, the customer actually sees it. And if they write back into a resolved thread, it simply reopens with its full history, right where it left off.

Two new filters sit beside the SLA toggle: Needs reply (conversations where the customer spoke last) and Unread. The Open and Closed tabs now show live counts too — no more scrolling to find who’s waiting. Both filters work with saved views.

The Knowledge screen has a search box — type a few letters and articles filter by title or category as you go.

Branding → Appearance gains a corner-rounding option for the Messenger — Square, Rounded or Round — with the live preview showing exactly how it lands on your site.

The Messenger now greets people by their first name instead of a username, and the Support card wears a proper icon.

Closing a conversation now removes it from the Open list everywhere, immediately — a cached row could previously bring it back after a reload.

Your AI now has a presence — a living orb in the sidebar, always there. Click it from any screen and a chat slides in: ask it to change its tone, explain why it answered something, or teach it from a file or a voice note. It proposes every change as a diff you approve. It’s named Iris by default, or whatever you called your agent.

The agent chat now has a mic — record a voice note right there and it becomes part of what your AI knows, alongside files, PDFs, sheets and screenshots.

Settings → AI is now a clean overview: six cards — Where it runs, Character, Knowledge, Skills, Rollout & handover, Testing & review — each showing its live state and opening its own focused page. No more endless scroll to find one control.

Drop a YouTube link into any Help Center article and it turns into a proper embedded player — right in the editor and on your live help site. Paste a watch, youtu.be or Shorts link; there’s nothing else to set up.

If you ever leave, there’s no maze. One warm screen asks what we could have done better, offers a single fitting alternative — pause your billing instead, or a hand from the team — and “Cancel anyway” is always one click. Whatever you choose, your agent, knowledge and every conversation stay saved, and come back if you return.

When your team answers a customer differently than the AI would have, that’s a lesson. The Trainer now reviews where it was overridden and which questions it couldn’t answer, and proposes the exact snippets, pinned answers and rules that would have prevented them — as diff cards you approve. Your AI compounds: every correction the team makes today makes it better tomorrow.

Open the Trainer and it greets you with last week in plain numbers — how many conversations it handled, how involved it was — the questions it stayed silent on, and one-tap suggestions to close the gaps. The Monday standup your AI runs on itself.

Pin the exact questions your AI must always get right, each with the approved answer to grade against. The set re-runs automatically on every change you make — and flags anything that used to pass but now fails, so a tweak to fix one thing can never silently break another.

Every AI reply in your inbox now carries a “Why?” — open it to see which of your articles grounded the answer, which configuration version produced it, and whether it was a pinned answer sent word for word. No more guessing why the AI said what it said.

Your AI now replies in the language the customer wrote in — Hinglish included — and you can leave per-language notes (“in Hindi, use formal aap”). Turn it off to always answer in your default language.

Give chat and email their own tone, length and extra rules — short and warm in the bubble, formal and complete over email — all from one Character panel.

Train your AI by talking to it. Open the Trainer and say “be more professional” or “never handle refunds yourself” — your AI proposes the change as a clear before/after diff, and nothing applies until you approve it. One screen shows its whole character, the conversation, and a try-me pane to test any question live.

Your AI can now have its own name — Maya, Rajat, whatever fits your brand — and a profile photo, right alongside its tone and rules. Customers see the name on its replies, and both live in the same versioned character file, so a rename is one sentence in the Trainer away.

Drop anything into the Trainer chat — a PDF, a spreadsheet, a doc, a screenshot, even a voice note — and it becomes part of what your AI knows. Sheets are read row by row, recordings are transcribed, and screenshots are read so the facts inside can be saved as knowledge with your approval. Attach, drag & drop, or paste; everything but video.

Every Trainer conversation is saved. Reopen an old session and continue exactly where you left off — your AI remembers that whole thread. Start a new session and it begins fresh. Each proposal in a session keeps its outcome, so old sessions read like a training log.

Any AI setup — including one you haven’t applied yet — can be replayed against your own recent customer questions. An AI examiner grades every answer (grounded? on-policy? right call?) and gives you a score with the failures listed. Simulated customer personas are one click more. Every run uses the same set of questions, so scores are comparable version to version.

Pin an exact answer to a question and it’s sent word for word whenever a customer asks it, however they phrase it. The AI never paraphrases a pinned answer — perfect for refund policies and anything legal signed off on.

Snippets are internal-only knowledge — policies, edge cases, reference facts — that inform AI answers without ever appearing on your help site. You can also import a page from your website or upload a PDF/DOCX, and the AI learns it the same evening.

Write multi-step runbooks for your AI — plain-language steps, simple branches, and your connected tools in sequence. Account changes still always wait for a teammate’s approval.

Ship the AI to a percentage of visitors (the same visitor always gets the same experience) or to chosen segments only. Handover rules are now hard-enforced: a message containing your trigger words goes straight to the team with your holding line — the AI steps aside every time.

See how involved your AI actually is, how each configuration version performed (replies, closes, satisfaction — side by side), and what customers spent after talking to support. Plus auto-attributes: define up to five fields, and the AI fills them in from each conversation as it closes.

A new Character panel on the AI screen lets you shape how your AI sounds and behaves: pick a tone (friendly to matter-of-fact), set how long answers should be, add voice notes, choose words it should prefer — and words it must never use — and write plain-language guidance like “hand any refund dispute to a human.” It applies to chat replies and help-search answers the moment you save.

AI character settings work like a history you can trust: each save becomes a numbered version recording who changed what, and any earlier version can be restored in one click. Nothing is ever overwritten silently.

The new Avatar Studio gives every teammate a look: Live avatars that gently animate (orbits, pulses, waves — they play everywhere, even in the chat bubble), designed styles and emoji tiles in six palettes, or a real photo, auto-cropped. Pick one in two clicks from the Team page.

Team management is now a clean roster — name, role, online status — with everything about a person in a side panel: edit, avatar, role, password reset, transfer, deactivate. No more crowded rows, whether your team is four people or forty.

Press ⌘K to jump to any conversation or screen. Reply to a specific message, WhatsApp-style — the quote travels with your answer, on both sides of the chat. Mark a conversation unread to come back later. ⏳ chips show exactly who’s been waiting and how long. Every reply now carries its own ✓✓. Translate any customer message — and your reply back into their language. Keep team notes on the person, not just the thread. Browse every photo, file and voice note a conversation has shared. Select many conversations and close, assign or tag them together. And when a conversation is handed to a teammate, the AI writes them a catch-up summary automatically. Agents can also mark themselves away — routing skips them until they’re back.

Every conversation now has a reference number — click it to copy, quote it in escalations. Duplicate threads from the same customer can be merged into one (history stays in order; the old thread redirects). Transfers carry a hand-off note and notify the new owner. And you can define sensitive words — legal, chargeback, regulator names — that auto-tag a conversation the moment it arrives, so high-risk cases surface as a filter, not a manual search.

A “New messages” line marks where you left off. See when a teammate is already replying, so two people never answer the same customer. React to messages with emoji. Edit a sent reply within 15 minutes — the customer’s screen updates instantly and the original is kept. Saved replies can include {{first_name}} and fill themselves in. Links show rich previews. Search inside a conversation and jump to the match. Write now, send at 9 AM their morning. And record voice notes right in the composer — customers play them in the chat bubble.

Connecting your systems to the AI now starts from two clear choices: point it at an API you already have (built in a form, no code) or connect an MCP server your engineers expose. Every connected tool shows a live status and its usage.

Read receipts, done the WhatsApp way: ✓ delivered when your reply lands, ✓✓ seen the moment the customer actually views it — updating live, in the thread and the sidebar. Hover any timestamp for the full date and time.

Seven destinations instead of fifteen. Every settings screen now lives in one hub with a quiet grouped menu — the screens themselves are unchanged and old links redirect.

A new inbox view shows every conversation where a teammate @mentioned you, newest first — and mentions now work in ticket notes too, with a notification when someone needs you.

Click any image to open a full-screen viewer — zoom, arrow through a thread’s images, download, Esc to close. Dropping an image into the composer now stages it as a preview: nothing sends until you press Send.

A book icon in the composer searches your published articles and inserts the title with a link to your help center — answer with something you already wrote, in two keystrokes.

Owners and admins can set a new password for any teammate from the Team screen. No email round-trips.

The big one: your AI can now look things up in YOUR systems — order status, bookings, account state — through tools you define in a form (or a connected MCP server, on Scale). Identity values always come from the verified customer, responses are field-allowlisted, changes need a teammate’s approval, and every call is audited. Paid plans; MCP on Scale.

Slice the inbox by who the customer is — country, plan, role, any custom field. Owners define segments once; the whole team gets them as pills above the conversation list, and exports respect them.

CSV exports (summary + filtered conversations), a Monday-morning email digest to owners, per-teammate SLA scorecards, an hour-by-day arrival heatmap — and click any teammate for their personal drill-down: active hours, reply trend, and their numbers.

The canned-reply picker is now a search-first palette: your most-used replies appear before you type, and typing searches every title and body — built for teams with hundreds of macros. Management became a searchable table with usage counts, so dead replies are easy to spot and prune.

Sent the wrong thing? Delete your own message — it vanishes from the customer’s messenger instantly, while your inbox keeps a dimmed record. Strictly the sender: nobody can delete someone else’s words. Plus copy-message on every bubble.

Customer profile photos now flow into the inbox (a dedicated identity field, like Intercom’s avatar). Teammate replies show who actually wrote them — their name and photo, not “You”. And a new Branding toggle can show your customers the real teammate answering them in the messenger, instead of a generic “Support”.

One row per person — a contact with several conversations shows once, with their whole history in the thread behind quiet date seams. Enter now sends (Shift+Enter for a new line), images paste or drag-drop straight into the composer, and the reply box lost its inner border.

Conversations your team answered that go quiet now close themselves after a set time (default 48 hours — tune or turn off in Settings), and closing invites the customer to rate the conversation right in the widget. Ratings feed the CSAT reports.

A new Appearance axis for the dock itself: Classic, Floating, Ghost or Compact — on desktop and the mobile tab bar. The mobile bar also got a proper redesign: four fixed tabs plus a More sheet.

The AI layer is now on for every workspace, not just ours: grounded answers from your own published articles in the widget, AI auto-replies in chat, and Copilot in the composer. Every plan includes AI credits — Free starts with 50 — and when credits run out the AI simply goes quiet while your team and help center keep working.

Growth is now ₹2,000 per seat per month and Scale ₹4,999 — same pooled AI credits (500 and 2,500 monthly). Existing subscribers keep their original price. Credit packs now come in four sizes: 1,000 (₹850), 2,000 (₹1,599), 5,000 (₹3,999) and 10,000 (₹6,999) — packs never expire.

Five slow ambient scenes — aurora, starfield, digital rain, scanline, breathing light — that pause when your tab hides and sit out on phones. And a shelf of Living rooms: the best room-material-scene combinations, applied in one tap.

Make the inbox yours: 200 curated rooms across a Dark and a Light library (searchable, shelved by mood), 8 Materials that reshape every surface (Glass, Ink, Neon, Terminal…), 4 Desk layouts, Space and Reading comfort controls — all composable, per-teammate, synced across devices. Plus a Custom mixer: two colours of your own, run through the same engine.

One click reads your logo and your website — colours, type, tone — and designs room candidates named in your brand’s voice. Tap one to wear it.

Your recent conversations and drafts live on your device: returning to Iris paints instantly (no loading skeletons), half-written replies survive reloads and bad Wi-Fi, and background syncing got dramatically lighter. The server stays the source of truth; everything is wiped on sign-out.

Deactivating a teammate keeps everything they worked on and revokes their access immediately — every session, every device. Reactivate anytime, transfer their open conversations and tickets to a teammate in one move, and see every team change in the new workspace Activity trail.

Returning teammates see your company’s logo and name on the sign-in screen — not ours.

Workspace owners get a direct feedback line in the profile menu — ideas, rough edges, praise. It lands with the team building Iris.

Start free, then per-seat paid plans — paid by UPI Autopay, GST included, cancel anytime, prepaid AI credit packs that never expire. The whole system rides the same payment rails that process real money for thousands of creators. Current prices always live on the pricing page.

Creating a workspace now verifies your email with a 6-digit code — the identity that billing, receipts and ownership hang off.

Tickets, Knowledge, Reports, Settings and friends are now top-level pages with clean URLs — the inbox is just the inbox. Old links redirect.

Author onboarding walks for your team or your visitors (with steps that complete themselves from real product events), and pre-chat workflows that qualify, tag, route and hand off — no code.

Thirteen feature pages, a blog, honest pricing, and a help center built on Iris itself — plus llms.txt so AI assistants can read it all.