Release Notes

A summary of what's new in each production release. Most recent first.

28 September 2026

Three things to look at first. The Team page has moved and is now called People — bookmarks to the old team-settings address need updating. Character limits are now checked on every save, so a deployment holding text written before the limits existed must be shortened before it can be saved at all, including to switch it Active or Inactive. And Cal.eu is retired — Cal.eu itself shuts down on 1 November 2026, so connect Cal.com before then. Everything else in this release is either new or needs nothing from you.

Billing, plans and your allowance

New — Billing & Usage. Your organisation's plan, this cycle's usage and the billing contact now have their own page, under Billing & Usage in the organisation's navigation. A Plans page compares what each plan includes. Only an owner can start or change a subscription. See Billing & usage.

New — New organisations start on a 14-day trial with every Growth feature on a smaller allowance. Choose a plan before it ends to keep your agent answering; otherwise the organisation moves to the Free plan — the agent stops taking conversations, and nothing is deleted.

New — A notice when your agent has stopped. When the organisation uses its whole conversation or message allowance for the cycle, a red notice appears across the top of every dashboard page until the allowance resets or you upgrade, and visitors see a line in the widget saying so. The amber warning at 80% is gone from the top of the page; the meter on the Billing & Usage page still turns amber as you get close.

Changed — Beta accounts now have exactly what Growth has, allowances included — conversations, messages and document pages. See the plan table in Billing & usage.

Your organisation and who is in it

New — Invite your colleagues. The People page — previously Team — lists everyone in your organisation and lets an owner invite people, change roles, remove people and restore them. An invitation lasts fourteen days, can only be accepted by the address it was sent to, and can be resent or revoked. Email addresses are checked as you type. See Who is in your organisation.

Changed — Owner-only actions. Renaming the organisation, saving its settings, deleting it, and managing billing are now owner-only. Deleting a team is open to an owner, or to whoever created that team, and the button now says why when it is unavailable to you rather than simply failing.

New — Switch organisation from the top bar. If you belong to more than one organisation, the switcher at the top left lists them, with each one's plan.

Properties, segments and the rules that route a conversation

Nothing you configured has changed. Every rule still exists, still applies to the same people, and still does the same thing. What changed is what things are called and where you find them.

Changed — Clearer names. Memory types are now Properties and Prospect profiles are now Segments — a property is a fact the agent captures, a segment is a saved population. Chapter 5 is rewritten to match.

Changed — A gate is now a condition, and it lives on the skill it guards. The separate routing screen is gone. A skill's configure page now ends with one list under Who reaches this skill, holding both conditions and nudges, each card saying which it is. A condition decides who may reach a skill; the one-off check it can carry instead of a segment is a test. Your rules moved with the change and were not rewritten. A skill with neither now says what that means — "Anyone can reach Book Meeting" — rather than showing two empty lists. See Configuring skills.

Changed — Conditions say who they admit, not who to turn away, and two conditions on one skill now show an OR between them. They always composed that way — passing either one admits — and a second card read to most people as a second requirement, which fails open: you believe you have tightened access and you have loosened it. Every existing rule was translated automatically and behaves exactly as before.

Fixed — A half-written condition now says what it does. A test with a comparison but no value used to read as an unfinished sentence. It now says No value set and states the consequence: nobody gets in, and the agent still asks the question. A condition in that state can no longer be saved.

Changed — Deleting a segment now tells you what else it deletes. A condition or nudge with no segment has nothing to test, so it has always gone with the segment — the confirmation just never said so. It now counts them.

Content

Changed — Content is added from one menu. Add content opens Scan a website… or Upload documents… in a side panel; the separate Import URLs and Import Documents pages are gone. Content also has its own section sidebar. Uploads are 4 MB per file (PDF, Word, Excel, PowerPoint, HTML, Markdown and RTF). See Ingesting knowledge.

New — See what a scan found, not just what you imported. Discovered pages now appear alongside your library, the page tree is selectable, and pages nobody imported are named and importable where you are already looking — so a scan you part-imported no longer has to be run again from scratch.

Changed — Status says what it means. A page reads In progress, Ready or Failed, and a page the crawler cannot fetch ends as Failed rather than waiting indefinitely. Correcting a page's type saves as soon as you pick it, and a corrected page is marked Yours so re-importing will not change it. Status filters are multi-select and no longer drop what you picked.

Fixed — A bulk delete can no longer remove rows your filters were hiding, and both Content and Prospects now confirm a bulk delete before it runs.

Prospects

Changed — Prospects has a section sidebar, with Segments, Properties and Lead scoring alongside the list. Properties and Segments open in a side panel beside the list rather than on a page of their own.

New — Contact forms and bookings now fill in Properties. What a visitor types into a form is recorded against their prospect as properties, so your segments match them the same way they match somebody who only chatted. Until now a form-filler's answers reached the contact details and stopped there, leaving the Properties tab empty and no segment matching them.

Fixed — Names and companies with accents are no longer dropped. José, Müller, García and Łukasz, and company names like L'Oréal SA or O'Brien & Sons, were silently refused and left blank on the prospect record. Job titles and locations were affected the same way.

New — Five page types for teasers, each with its own glyph on the Engagement tab: Home page, Pricing pages, Product page, Case studies and reviews, and Everywhere else. Your site is classified by what each page is for, and each type gets one teaser. Legal and checkout pages never show a teaser. See Teasers.

Changed — What reaches your CRM. When a visitor fills in a contact form or books a meeting, their details reach your prospect list and your connected CRM whichever way they answer the "remember me" question — they gave you those details on purpose. So does everything else: the properties the agent works out, the lead score, and the session summary. Bookings now record the visitor's first name, last name and email, which they previously didn't.

Changed — The "remember me" question governs one thing: the visitor's browser. It reads "Do you want us to remember you next time?", with Yes and No, and the answer decides only whether a marker is kept on their device so they are recognised when they come back. A visitor who answers No still reaches you in full — their next visit simply starts as a new prospect instead of joining the record they already have. The same is true of Stop remembering me in the Manage data menu, which previously also stopped that person reaching your CRM at all. See Playground testing.

Changed — Deployments. Domains are added one at a time — type one and press Enter. The opening message is always one message. Character limits are now checked on every save, so a deployment holding older text must be shortened before it can be saved, including switching it Active or Inactive. A deployment switched to Inactive stops answering, including in a widget that is already open. The same length checks apply to agents, teams, properties, attributes and segments. See Deployments & widget embed.

Integrations

Changed — Integration tiles say what's wrong. Each tile shows Connected, Refresh failing, Reconnect required, Paused or Paused by plan, and a reconnect is asked for only when the provider says the connection is really gone. The Syncing on / Syncing off switch is removed; an integration you paused with it now reads Paused, and Reconnect switches it back on. See Reading an integration's status.

Removed — Cal.eu. The separate Cal.eu integration is retired and Cal.com is the booking integration. Cal.eu shuts down on 1 November 2026.

Fixed — A booking integration no longer dies from one bad moment. An interrupted token refresh could permanently break a Cal connection; the five separate refresh paths are now one, and a transient failure is retried instead of being treated as a revoked grant.

The look of the dashboard

Changed — A new look. A cleaner light theme with white panels on a light grey frame, a typeface that now loads properly everywhere, a slimmer top bar, and a left rail you can pin open. Tables fill the width of the page, and on Content and Prospects you can choose, reorder and resize columns. Column widths you had dragged before this release have been reset to the new defaults.

Improved — Screens tell you what they are doing. Lists, tiles and panels now show the shape of what is coming while they load rather than a spinner in the middle, a failed load gives you the reason and a Retry instead of raw server text, and a refresh no longer blanks the page you are reading. Animation respects your system's reduce-motion setting.

Improved — Getting around by keyboard and screen reader. Tables are keyboard-operable with sortable columns announced, every control has a visible focus ring and an accessible name, the rail and menus are real links, and there is a skip link to the main content.

Fixed — Nine more screens now warn you before you lose unsaved work, and Save is inert until there is something to save. Each section of a skill's configure page saves its own work, so pressing Save no longer commits two of three sections. A tool's on/off switch takes effect when you flip it, instead of waiting on a button two sections away.

Removed — The Activity Logs page, and the last Back to skills button — use the sidebar, as on every other screen.

The agent itself

Improved — The agent remembers what it has already shown you. It keeps a record of the product it presented, so a follow-up is understood as being about that product, a confirmation is taken on the turn it is given, and it stops re-asking a question you have answered. A "yes" to the agent's own question now settles the thread rather than opening another.

Improved — Answers come from your own content where your content has them, and a question the current skill was not built for is answered in place rather than pushed to a next step.

Fixed — Tapping the body of a teaser no longer puts the agent's own words in the visitor's message box. Mid-conversation chips are hidden while a reply is in flight, so a double tap cannot send twice. The playground session now ends before the screen says it has.

Fixed — The prospect CRM panel shows the real sync state instead of Pending forever. The Beta Programme notice stays dismissed as you move around. Saving a deployment tab no longer empties the form while it reloads. The agent count on a project card counted deleted agents, so a team that created four and deleted one read "4 agents" while the agents page listed three.

Changed — ContentBot is now called Manifest in the list of agent types, and is available only to organisations entitled to it. It is still marked Coming Soon.

Behind the scenes — Around 550 changes landed since the 26 August release, and none of them need anything from you. The largest single strand was security and tenant isolation: member names and addresses are read from the authentication store rather than a mirror that could go stale, one organisation's browser can no longer be served another's entitlement list, platform-admin access is deny-by-default, sessions are verified locally on every sensitive route, and irreversible actions are gated on the Owner role with project access granted explicitly. Alongside that, the prospect record gained a single writer with the tenant check inside it, replacing two pieces of code that wrote the same columns and disagreed about how; the dashboard moved onto one design token system with automated gates that hold it; and the seeds no longer destroy runtime data on deploy.

26 August 2026

Worth doing first: check the AI disclosure line in your widget, and decide whether to switch teasers on. Two things in this release show up on your live site. The widget now always carries a short line under the chat title saying the visitor is talking to AI — it's on already, in English, and you can reword it to your voice or language (see New below). Proactive teasers are the opposite: they're built and ready but switched off for every deployment until you turn them on. Nothing else in this release needs anything from you.

New — Teasers: say something before the visitor opens the chat. A teaser is the small bubble beside the launcher that invites someone into a conversation, with up to three tappable chips underneath that hand them straight to one of your agent's skills. Your site is grouped into three page types — Product page, Product listing and Everywhere else — and each gets one teaser, so there's no ordering to get right and nothing to prioritise: the switch turns that page type's teaser on or off. Wording is capped at 100 characters, previewed exactly as a visitor will read it, and Reset puts the original back. Chips are picked from a fixed pool so they route somewhere reliable, and hovering one names the skill it goes to — a chip whose skill is switched off is greyed out and says so. Some chips fill themselves in from the page the visitor is on, so a compare chip reads "Compare Router X with Router Pro"; if there's nothing to fill in, that chip quietly drops and the rest of the teaser shows as normal. See Chapter 8 — Teasers.

New — You decide whether teasers fire, and when, per deployment. Teaser wording lives on the agent and is shared by every deployment of it; whether a teaser appears at all is a deployment decision, on the deployment's new Engagement page. It's off by default. Switch it on and two gates decide the moment — Time on page, or Scroll depth — whichever happens first, defaults 8 seconds and 40%. The line under the fields reads your two numbers back as one sentence, including the case where a gate of 0 means "no gate" rather than "no delay". That split is deliberate: the same agent on a fast marketing site and on a long-read documentation site wants the same words and different timing. See Chapter 8 — The Engagement page.

New — The widget tells visitors they're talking to AI. A short line now sits under the chat title — by default "AI assistant — here to answer questions". The EU AI Act's Article 50 disclosure duty applies from 2 August 2026, and rather than judge per deployment whether the point is obvious enough to skip, the widget always shows the line. You can reword it in the deployment's Content step, up to 80 characters, to match your voice or your visitors' language; you can't switch it off — clearing the field falls back to the default. Keep any rewrite positive and capability-led ("I can answer questions or book you a demo") rather than a warning: warning-style disclosure measurably suppresses engagement, capability-style doesn't. This is a description of the feature, not legal advice — confirm your own obligations with your adviser. See Chapter 8 — AI disclosure.

New — Widget engagement in Analytics. A new section shows what happens before and around the chat opening: Widget Loads as the honest denominator (loads, not people), a Pages table of where the widget loads with its open rate, and a Teasers table with Shown, Clicked, Engagement and Conversations per teaser. Rows are named by the teaser rather than by its wording, so rewording your copy doesn't fragment your history. With more than one deployment you get a selector, because where a teaser is embedded and how soon it fires is exactly what differs between them. A healthy load count with a low Shown is never a copy problem — the chapter walks the cheap checks in order. See Chapter 10 — Widget engagement.

New — The documents you agree to are named where you sign in. A single line under the sign-in, sign-up and set-a-password screens names the Terms and Conditions, the Data Processing Agreement and the Privacy Policy, each one a link that opens in a new tab. There's no tick box: creating your account or signing in is the act of accepting, and we record the date your organisation's agreement began so we can give you a copy of that record on request. Accounts on the beta also carry a dismissible notice across the top of the dashboard linking the Beta Access Agreement; it comes back on the next page load, and stops appearing on its own when your account moves off the beta. See Chapter 1 — The documents you agree to and If your account is on the beta.

Improved — The agent no longer denies selling something it just recommended. When a follow-up question was scoped to a product reference the agent had invented, the search correctly returned nothing and the reply was indistinguishable from an honest empty result — so the visitor was told the business doesn't sell a product it had recommended forty-five seconds earlier. A scope that names no real product is now dropped and the search runs properly, and a zero caused by a bad scope is no longer remembered as "the catalogue is empty".

Improved — A search that matches nothing says so. When a visitor states a hard requirement and no product meets it, the agent stops presenting products that fail that requirement as though they were answers. Deliberate near-misses still come through, still labelled with what they don't meet.

Improved — Conversations follow through better. Asking to buy after the agent has recommended something now routes to buying, rather than being handed back to a general answer. An answer to the question the agent just asked is taken even when it's hedged — and the agent says the value back in passing, so a visitor can correct it without being interrogated. When the visitor asks something the current skill wasn't built for, the agent answers it in place instead of pushing the next step, and it only holds a next step open while a question of its own is genuinely outstanding. Questions are also scoped to the skill actually running, so one skill can't ask for another's details mid-thread.

Improved — The chat panel keeps up with a long reply. Auto-scroll now follows the reply to its real end — the typewriter finishing, the chips landing, cards and their images loading — instead of stopping when the network stream closes. Scrolling up mid-reply no longer yanks you back down, and coming back to the bottom resumes following. Tapping a teaser chip opens the widget on the teaser's own line rather than replaying a greeting the visitor never read over a question they already asked, and a compare chip now opens a real two-product comparison instead of a generic answer.

Improved — Custom memory types read as you named them. Fields you added yourself now resolve to their labels everywhere they're used — in the agent's prompts, the widget's forms and the skills editor — instead of appearing as their internal snake_case key. Memory types you've disabled are no longer written to, a project's own memory type wins over a same-named one inherited from elsewhere, and duplicate names are prevented rather than left to collide.

Improved — Meeting booking shows the booker's timezone, drops the nested scrollbars, and renders widget errors in language a visitor can act on.

Changed — Importing URLs is limited by your document-page allowance, not by what you select. Select more pages than remain this cycle and Import is disabled with the reason next to it. Selecting and deleting stay available at any allowance level, so you can always clear out pages you don't want — including when the allowance is fully used. See Chapter 4 — Importing from a website.

Fixed — The Widget channel filter is back on the Analytics page, and preset date ranges no longer serve a stale window. A partial read of the internal action catalogue could switch off a live action row on an agent; it now refuses rather than guesses. A chip tap the agent declined is no longer reported in analytics as a chip that worked. Widget domain allow-lists now match a host with a port correctly.

Behind the scenes — The largest engineering release so far, and none of it needs anything from you. Around 130 recorded defects were fixed and retired across chat, billing, integrations, widget and the API layer, including two security fixes (a tenant guard on a billing routine, and locking down integration token columns). Test coverage was backfilled across every previously unmeasured layer and pinned with automatic ratchets, so coverage, file size and defensive-code counts can only move the right way; a database-backed suite now proves tenant isolation against a real database; and a browser-driven end-to-end suite runs against the real deployment before anything is promoted. Every request the agent makes is now attributed to the tenant it belongs to, with dashboards built on top, which is what makes per-customer cost and quality measurable. Alongside that: the chat runtime's context assembly was consolidated onto one path, five oversized modules were split on their real seams, the dependency stack moved to Zod 4 and OpenAI 6 with byte-identical prompts verified, and continuous integration gained build, dead-code, shell-script, migration and prompt-seed gates plus an automated code review on every change.

11 August 2026

Action required: re-import your product pages. This release changes how product content is broken up for search (see Improved below). The change applies when content is ingested, so pages already in your knowledge base keep their old shape until you re-import them. If you use Product Finder, re-import your product pages to pick up the improvement — see Re-importing content when it changes. Nothing else in this release needs anything from you.

New — "I'm not sure" is now a real answer. When Product Finder asks a narrowing question the visitor can't answer — "water or biofilm?", "which sensitivity?" — they get an option saying so, and typing it in their own words does the same thing. The agent then answers its own question from your knowledge base and asks it again in the same reply, instead of leaving the visitor to guess or abandoning the thread. If your content can't settle it, the agent says so plainly and offers a starting point they can change rather than inventing an answer. There's nothing to switch on, but it's worth writing a short guidance page or FAQ entry for each hard question your fields ask — that's the content this draws on. See Chapter 6 — When the visitor cannot answer.

Improved — Product pages are broken up for search more reliably. The chunker now produces one consolidated passage per category — overview, specifications, pricing, use cases, and so on — with a plain-language overview guaranteed on every product page. Previously a page dense with specifications could end up with no overview passage at all, which made it effectively invisible to Product Finder even though the page was ingested and ready. This applies to content ingested from now on; re-import existing product pages to pick it up.

Improved — Qualification is more accurate. The agent is better at telling what a visitor said about themselves from what they mentioned about someone else, so a colleague's name or a company they're comparing against no longer lands in their profile. It also reads push-back more precisely: an objection now means a concern about what's being sold, not confusion at the last message or a correction of the agent.

Improved — Product discovery holds its place better. When the agent asks a narrowing question, that question stays with the skill that asked it rather than being handed off mid-thread.

Fixed — Newly published pages now appear when you map a site. Site mapping was reading from a cache that could be up to seven days stale, so pages added to your site recently could be silently missing from the discovered list — never imported, and never findable by the agent. Every map now does a fresh lookup. If you've added pages to your site in the last week and wondered why the agent didn't know about them, re-map and re-import.

Fixed — Forms render the right fields when two skills share a form. With both Contact Request and Emergency Contact Request enabled, the form could pick up the wrong skill's field list depending on the order the skills happened to be stored in.

Behind the scenes — A substantial rework of how your configuration reaches the agent. Prompts are now assembled fresh each time they're read rather than being rebuilt and stored on every save, so an edit to a skill, agent, or organisation reaches new conversations on its own — there is no longer any "rebuild" step that can be missed or fall out of date. Alongside it: tighter database permissions, removal of a table and eight columns that were no longer read, and a new reconciliation tool that keeps existing agents' skill lists in step with the catalogue. No action required on your end.

4 August 2026

New — Cal is now a proper integration, in both regions. Connect your Cal account once from the Integrations screen and the Book Meeting skill offers your real event types in a dropdown. There's no event URL to paste any more, and no way to point the skill at a link that turns out to be wrong. Cal runs two independent deployments, so the screen has a tile for each: Cal.com for accounts on the global instance, Cal.eu for accounts on the European Union instance — connect whichever one you sign in to. Reschedules and cancellations run through Cal's own confirmation email against your live availability. See Chapter 11 — Integrations.

New — Gate a skill so the agent qualifies before it acts. You can now put a condition on any skill: while it's unmet, the agent asks a qualifying question rather than starting the skill, and what the visitor answers decides whether it opens. Add several and they behave as doors — opening any one lets the visitor through, which is how you write an exception to your own rule. See Chapter 6 — Configuring skills.

Changed — Discourage routing rules are retired; gating replaces them. Existing rules were migrated onto the skills they targeted, so you'll find them in that skill's configuration rather than on the Routing screen. The change is an upgrade, not a move: a Discourage rule only nudged the agent's preference, where a gate holds the skill shut however the visitor arrives — including when they ask for it outright. Chapter 6 now covers Suggest rules only.

Changed — Qualification is no longer a separate skill. The Qualify Prospect skill has been retired. Each skill now declares what it needs to learn from the visitor, so the asking happens inside the conversation the visitor is already having instead of as a detour. Advisory Guidance ships with an empty list and is worth configuring first — see Conversation inputs.

Improved — One tap instead of typing. When the agent asks for something with a fixed set of answers, the visitor gets tappable options rather than a text box, and the answer lands in exactly the shape your segments (then called prospect profiles) match against. The agent is also more disciplined about proposing next steps: one at a time, only at a natural resting point, and it lets go rather than re-offering the same thing every turn.

Improved — Product discovery holds the thread better: the agent keeps track of the product the visitor settled on, shows only the options a stated requirement actually rules out, and can card a shortlist it has just named. Attribute extraction is now grounded in cited source text, so values are traceable back to the page they came from.

Fixed — URL imports were charged twice per page. Pages are now charged once, when they're actually scraped, so URLs that fail or never complete cost nothing.

Behind the scenes — A large restructuring of the chat engine into clearly separated concerns, with substantially expanded automated and conversation-level testing. No action required on your end.

15 June 2026

New — Zoho CRM integration. You can now connect Zoho CRM alongside HubSpot from the Integrations screen. Captured prospects flow into Zoho automatically — new people become Leads with their details, lead score, and source kept current, while chat summaries, booked meetings, and contact requests are added as Notes. Breezee follows Zoho's Lead-to-Contact lifecycle, updating the Contact (not a duplicate) once a Lead converts, and stays connected to your account's regional data centre (US, EU, and others). See Chapter 11 — Integrations for setup and what gets synced.

Improved — Prospect attributes with fixed options can now hold more than one value, so a prospect can match several options for sharper qualification and routing. Conversations feel more natural: the agent paces the details it gathers without repeating itself, answers like a helpful expert (leading with the outcome, not dry facts), and picks its skills and tools more consistently. Qualification is more accurate too — each answer is now extracted against its slot's expected type and allowed options, so a value can't land in the wrong field. The chat widget renders inside an isolated style boundary so it looks the same on any site, keeps the greeting visible when it opens, and now books correctly with EU and self-hosted Cal.com instances. Your dashboard stays cleaner: archived or deleted memory types no longer reappear, memories tied to removed skills are excluded, and a logout error during project loading is fixed.

Changed — Deleting a prospect profile now automatically removes any agent routing rules that referenced it, so you're never left with rules pointing at a profile that no longer exists.

Design & polish — A consistency pass across the dashboard's styling, plus documentation tidy-ups.

Behind the scenes — HubSpot and Zoho now share one batched contact-sync engine, so syncing stays fast even for very large prospect lists. We also completed a major code-quality and type-safety push: stricter linting and TypeScript checks run automatically on every change, clearing thousands of latent warnings and catching whole classes of bug — unhandled background work, unsafe handling of unexpected values, and data-formatting slips — before they reach you, backed by expanded automated and chat-conversation testing. Documentation links are environment-aware, with housekeeping on how skills and actions are stored. No action required on your end.

1 June 2026

New — Public documentation site at docs.breezee.ai, with search, glossary, and troubleshooting. The Docs link in the dashboard top bar takes you straight there, themed to match wherever you're working.

Improved — Faster chat widget on customer sites, with steadier sessions: we fixed edge cases where the widget could get stuck in a "session ended" state or fail to resume after a refresh. The widget can now expand to a larger size or collapse back to compact, and it opens full-screen on mobile for easier conversations. Routing rules that compare text values (e.g. job title equals "Engineer") now ignore capitalisation.

Changed — Prospect classification now flows entirely through Profiles. The legacy single-value Status field has been retired; switch any filters or reports that used it to a Profile. HubSpot sync no longer writes the Status field; Profile matches and Lead Score continue to sync as before.

Design & polish — A refreshed look across the dashboard, documentation site, and chat widget. A striking near-black dark mode, a consistent Breezee logo and wordmark, and a unified brand-green accent now run throughout. The documentation site matches the dashboard's look and links straight back to your dashboard, and top-bar controls are clearer and more legible.

Behind the scenes — Security hardening: we rotated platform credentials, tightened how internal services authenticate with each other, and added automated secret-scanning to our build pipeline. No action required on your end.