Changelog

The transcription step of operations agents can now split a call's transcript by speaker and mark each turn as agent or customer, so the system you send it to, a CRM for example, knows who asked and who answered.

Operations runs now have a calendar period, a total for the period, the cost and what was charged per run, and an export to Excel; a new Monitor > Runs page lists every agent's runs together, the Logs show the cost per run, and the agent page shows the plan minutes you have left.

Email templates: keep your own email layouts under Settings > Email, start from the built-in Voiceland AI design, edit with a live preview, validate, send yourself a test, ask the assistant to change it, import and export, and pick one in any workflow's Send email step.

The dashboard now includes operations agents next to the voice and chat agents, with a filter by agent type and a custom period.

The send-email and webhook steps show a preview of exactly what will go out, with real data, and can be sent on their own with the other side's answer shown live; the webhook step now has method, query parameters, headers, authentication and JSON, form or text bodies, and the email step Cc, Bcc, Reply-To and HTML.

The phone system integration now keeps a log of every request the PBX sends, with the answer it received, so a call that never arrived or an authentication that failed can be read on the spot; it is kept for 30 days and never stores a password or a token.

The PBX integration card shows whether the phone system is connected: waiting until it authenticates, connected from the moment it does, with the time of the last authorization and the last activity, and the vendor guide link now opens the right page.

A call with no recording, such as a missed call, no longer parks an operations workflow for a day: the download step is skipped with its reason and the run finishes, instead of failing and retrying every hour.

Every trigger an outside system fires, a finished PBX call, a webhook or a call from another agent, can be listened for: press Listen, make it happen, and read exactly what arrived before the board acts on it.

In the workflow designer every change can be undone and redone (Ctrl+Z, Ctrl+Y, Ctrl+Shift+Z, Cmd+Shift+Z), a selected connection or step is removed with the Delete key, the selected step's name is shown at the top of its panel, every trigger has a Run through this trigger button, and the revision history is one click away.

A library of ready-made workflows by integration (Yeastar, HubSpot, generic), all tested end to end, with a try-it button, and an AI designer that draws the workflow from your description with checked connections.

Every save of an operations agent's workflow is kept as a revision: see what changed, compare two revisions side by side, and restore any of them without losing anything.

A workflow can be exported to a JSON file and imported from one: the file is checked first, you see the current and the incoming board side by side, and you decide whether to replace.

The Webhook received step now has a listen button: press it, send the event from the other system, and see the request as it arrived, in formatted JSON, without it being run.

Selecting a step in the designer now shows what it received and what it produced in the last run, in full and formatted, not just the size in bytes.

In the operations agent designer you can now select the line that connects two steps and delete just that line, without deleting the step.

On a phone call the agent no longer reads out the little reference numbers from your knowledge base; they still appear in chat and in the transcript.

Voice now pronounces acronyms your content writes with dots, such as EOF, as words instead of spelling them out letter by letter.

The Try a search tab now says plainly that only published documents are searchable, and says it again when a search finds nothing.

A document that came from a web page can now be checked for changes every so many hours, days, weeks or months, from one hour up to a year, and the document shows when it was last checked and when it is checked next.

The knowledge base is now included from the Growth package and up, so your agent can answer from your own documents without moving to Business.

Knowledge base answers arrive sooner: the lookup starts while the caller is still finishing their sentence, so the agent answers in one step instead of two.

The content gap report, what your visitors asked that your content does not answer yet, now appears on the knowledge base's Report tab in the console.

In chat, an answer now shows only the sources it actually used, and one entry per document instead of one per excerpt.

An agent no longer ends the conversation when you answer yes to its offer to tell you something more.

Operations agents are billed as automation minutes: 30 seconds minimum per run, a quarter of your plan's per-minute call rate, drawn from your included minutes first, with five report questions included in every run.

The PBX integration is now set up once per project instead of per agent, under the project's new Integrations page, and the HubSpot integration lives under the Contacts app; an operations agent opens on an overview with its status, history and settings, and the workflow designer can align the steps or open in its own window.

Operations agents now retry: every step has a retry policy (up to 24 attempts, for example once an hour until your endpoint answers 2xx), a failed step can be retried by hand from the run, every step shows the JSON it received and produced and the exact answer of the other side when it failed, a Dispatcher step starts several agents or posts to several addresses at once, and a Summary step writes a summary of the transcript for the email and the webhook.

Your monitoring now covers operations agents: the metrics endpoint exposes runs, outcomes, step durations by module, cost and schedule fires, and a fifth ready-made Grafana dashboard for them is listed under Dashboards.

Operations agents now keep working through restarts and updates: every step is saved as it completes, a run picked up by another server continues from where it stopped, and a new Schedule trigger starts a workflow every hour, every morning, on weekdays, or on your own cron expression.

Operations agents: a new agent kind that does work instead of talking. Draw a flow of actions on a board that runs when a PBX call ends, a webhook arrives or you press Run: download the recording, transcribe it into one or more languages, answer your prepared questions in a report with an Excel file, email it, fire a webhook, or call another agent. Steps run in series or in parallel, every run shows what each step received and produced, and runs are billed at 0.05 euro per minute with a one-minute minimum.

HubSpot integration: connect a HubSpot account with a private app token and keep it in step with your contacts, pulling on a schedule with the conflict policy you choose, pushing only what is missing or everything, and sending each new contact to HubSpot the moment it is added; the integrations catalog now shows every supported system with its logo.

Custom fields on contacts: define your own fields (text, number, date, choice, yes/no) and publish any of them to the PBX so a home number, an address or a customer tier shows on the phone's screen when that person calls.

Contacts app and PBX integration: keep your contact book in the console and connect a Yeastar P-Series Cloud PBX to it, so callers are recognised by name, unknown callers can be added automatically, and every finished call is journalled on the contact with its recording link.

An agent no longer starts answering while the caller is still speaking: a safety limit meant for a stuck conversation was also firing on people who simply had more to say, cutting them off mid-sentence when they listed several things in one breath, and it now gives way for as long as new speech keeps arriving.

Greek-speaking agents now say sizes the way a person behind the counter says them, instead of stumbling over the Latin abbreviation, so a drink listed in your catalogue as 500 ML is spoken naturally; the written order, the transcript and the call record keep the short form.

Voice agents reply noticeably sooner: the model that decides when a caller has finished speaking now runs at the same time as the short pause the agent waits out, instead of before it, and that pause shortens automatically when the model is certain the caller is done. Agents on the newest language models also gain a Thinking before answering setting on the Behavior tab, where Minimal trades a little deliberation for the fastest possible reply.

A new per-agent option, \"Always ask the closing question before hanging up\", for flows that must end on a confirmation such as orders and bookings: the agent may only hang up right after it asked its closing question and the caller answered, or after the caller said goodbye, so a bare \"ok\" after the order summary no longer ends the call with the order unconfirmed.

Each agent now has a \"Do not hang up on a yes\" setting on its Behavior tab: when the agent has just asked a question and the caller answers briefly, the agent acts on the answer instead of ending the call, and you can add your own closing questions (after which a yes does end the call) and your callers' own goodbye words, so the behaviour fits your agent's wording without waiting for a platform change.

The audio enhancement an agent may have for phone calls no longer applies to voice sessions from the web widget: the browser's own processing is used there, so one agent serves both the phone and the widget without a duplicate setup.

Share your agent with a link: from the preview you create a public test link, hosted by us with the agent's web widget, send it to colleagues or clients, and withdraw it whenever you like, with an expiry and a session limit you set.

Agents can now say a short 'one moment' when a reply takes a little longer, so the caller never wonders whether the line dropped: a switch under Behaviour and timing, with ready phrases in the agent's language.

The agent no longer ends a call when the caller answers yes to a question it just asked: a short reply such as \"ok, thanks\" to an offer now counts as acceptance and the conversation continues, while a clear no, a goodbye, or a yes to \"anything else?\" still ends the call as before.

Connect live data sources from your own system (a menu, a price list, availability) to your agent: give the address, pick a provider such as iotSoft or describe your own format, and the agent has the current picture from the first word of every call, with no waiting, while the access key you give is never shown again.

Data sources can now refresh as often as every five minutes, and a refresh that finds nothing new leaves your data untouched: you can see when a source was last checked and when it actually changed.

Web widget chat conversations now draw on the same plan minutes as calls, at a lower chat rate set on your plan, and only for the time from the first message to the last one: idle time is never charged, a chat that ends without a reply costs nothing, an idle conversation closes by itself after five minutes, and pressing the widget's close button on a live conversation now asks whether to end it.

Chat conversations in the web widget now count toward your plan's concurrent-call limit, the same limit your phone calls share, so a plan that allows two conversations at once allows two across calls and chats together; a visitor who arrives while every slot is taken sees a short busy message instead of waiting in silence.

Try tools that call your own systems without a backend of your own: new demo agent templates with ready-made examples (orders, weather, bookings) work the moment they are created, so you can hear the agent call a system and answer before you connect yours.

An agent tool can now run inside the visitor's browser, using the session they already have on your site: the agent answers about that customer's own order or appointment without you ever handing the platform a key to your system.

The reason for a transfer to a representative can no longer be heard by the caller: when the model leaks the transfer's internal details as speech, the system recognises and suppresses them whole, and the caller hears only the normal \"connecting you to a representative\".

Call recording and direct transfer to a representative now start at the Growth plan: Starter is the trial tier and no longer records calls or hands them to a person, so its plan card states no recording retention window rather than a promise about audio it does not keep, while every paid plan keeps the window it had (30 days on Growth, 60 on Pro, 90 on Business, a year on Enterprise). Starter still includes email notifications, the web widget, voice and chat on one agent, versioning with rollback, and the customer satisfaction survey, and these are now named on the plan so you can see what the trial covers.

Fixed the spot that made the voice stumble on the same phrase in three different calls: when a quoted screen label such as \"Charge now?\" is followed by a full stop, the doubled punctuation that confused the speech engine is now simplified before it is spoken.

Every agent's instructions are now automatically checked for the text patterns proven to break speech delivery - saving is blocked with a clear explanation, the AI assistant detects and fixes them itself, and the platform's built-in templates were cleaned of the same patterns at the source.

When the voice engine glitches and leaves seconds of dead air in the middle of an answer, the system now removes it before the caller hears it, and very long sentences are spoken in smaller natural segments so such a glitch costs half a sentence instead of the whole answer.

Whatever a caller manages to say in the final moment before a call ends is now kept in the transcript, instead of being lost because speech recognition delivered it a fraction of a second after the call record had already been sealed.

The safeguard that stops an agent hanging up while the caller is still talking no longer requires a satisfaction survey: you can set your own closing question in your own words and the agent will ask it before ending the call, instead of asking every caller to rate you from one to ten.

When the agent wrongly tries to end a call while the caller has just asked a question, it now answers the question instead of asking them to rate the service; the satisfaction question is only asked once the caller has actually said goodbye.

When you write an exact phrase in quotation marks in your agent's instructions for it to say, the agent can now say it: until now the safeguard that stops an agent revealing its own instructions could mistake that scripted line for a leak and answer \"I don't have that information\" instead.

Three fixes for calls in languages that do not use the Latin alphabet: voice billing counted the text's encoding rather than its actual characters and was roughly double on Greek, the agent no longer deletes Latin words a caller says such as your company name or \"internet\" - and when every word was Latin it used to fall silent altogether - and the agent no longer hangs up when the caller speaks again during its goodbye.

You can now duplicate an agent and move an agent between projects from the console. Duplicating creates a draft copy under a new name in the same project, with every setting carried over, so you can try changes without touching the original. Moving places the agent in the project you choose as a draft: phone numbers must be detached first so no number is left ringing an agent that is gone, and knowledge collections stay with the project they belong to - the confirmation lists which ones were detached so you can attach the right ones in the new project.

If you have paid for a year and want to move to monthly billing, the new plan now starts the day you choose it instead of waiting for your year to run out: the part of the year you have not used becomes credit, that credit pays for the new plan at its monthly price, and nothing is charged on the day of the change. Before you confirm, the screen tells you how much you have already paid, how long it covers on the plan you picked, and the exact date of your first monthly charge, and we email you a reminder a week before that charge is taken.

Call recordings are now kept for a number of days that depends on your plan (15 on Starter, 30 on Growth, 60 on Pro, 90 on Business, a year on Enterprise), shown on every plan and in the plan details, and once that window has passed the audio is deleted while the call record and its transcript stay: a deleted recording says so instead of failing to play, and changing to a plan with a shorter window warns you first.

Three small things at the end of a call are now right: the agent no longer reads out stray quote marks or brackets from its own internal instructions, it says goodbye once instead of twice when it has to close the call itself, and two agent lines spoken well apart (such as the hold message and the message that a colleague could not be reached) appear as separate lines in the transcript instead of one run-on sentence.

Every plan now grants exactly the features it lists: the calendar, live call events sent to your systems, the export of call records, the chain of custody, the activity log, downloadable recordings and shareable recording links, and the limits on simultaneous calls and on the number of agents are all held to your package, in the console and on every call, so a plan change takes effect from the next call rather than the next time an agent is saved, while listening to a recording inside the console stays available on every plan.

Changing your plan now keeps your renewal date and charges only the difference: an upgrade is priced on what you have already paid for the rest of the current cycle, moving to a yearly plan extends your coverage from where it stands instead of restarting it, a downgrade or a switch from yearly to monthly waits for your renewal and is confirmed to you by email on the day it takes effect, every plan change you make is confirmed by email, and each package description shows its current prices rather than a figure typed in once.

When a caller says they need nothing more, the agent now asks the short closing survey as intended, instead of answering that it has no information on that: a safety check that stops the agent from reciting its own instructions was also catching the survey question wherever the agent's instructions quote it word for word, and the check now knows the lines the agent is meant to say, such as the greeting, the goodbye, the survey, the hold messages and the chat messages for when nobody is available, and no longer treats an answer quoted faithfully from your knowledge base as something to hold back.

Calendar: holiday lists on the Availability page. See the Greek national holidays (including the movable ones that follow the Orthodox Easter), clone them into your own list, add or remove days, and apply one or more lists per member so those dates offer no slots and accept no bookings.

Agent testing: a Testing tab on every agent runs scripted UAT conversations against the live agent brain - real calendar tools, per-turn checks on tool calls and replies, chat and voice channels, full transcripts, and the same scenarios runnable from your own code via the API.

Calendar: connect Microsoft Outlook alongside Google, and give any event type a Video call location that adds a Google Meet or Microsoft Teams link to every booking.

Calendar: get an email on every new booking (to your account address or any list of recipients), and send all calendar emails through your own mail server with your own From address - inherited from the AI application settings or configured separately.

Calendar: mark a booking as a no-show and add host-only internal notes; the no-show rate now appears in Insights.

Calendar: private booking links for hidden event types, a booking window and per-day/week/month limits, and out-of-office dates that block every schedule.

Scheduling: let people book time with you from a shareable link, the API, or an agent on a call, with availability, reminders, Google Calendar sync, and no double-booking.

Putting records on hold now actually keeps them: a hold stops deletion and stops the clock that would have removed them on schedule, it can cover a whole project including conversations that have not happened yet, and lifting it puts any erasure request that was waiting back in the queue and tells you which ones. A colleague who takes over a chat now reads what the visitor already said before they arrived, instead of joining mid-sentence. Your call records show who took a conversation over and how long the visitor waited. Suggestions from unmatched phrasings now show the match level they were judged against and how far each one fell short, so the setting can be tuned from your own conversations rather than guessed at. And two things that were quietly narrower than they looked have been made plain: handing a conversation to a person on a guardrail is a real handover in chat and a refusal on a phone call, which the assistant's settings now say per channel; and generating an operations report on demand is limited per hour, with your allowance on every response.

Calls your assistant places out now have a stated daily allowance and a stated pace, and both are visible in your console before you reach either, so a script that loses its footing or a key that falls into the wrong hands cannot run up a phone bill overnight, and a planned campaign can be paced instead of discovering the limit by hitting it; a call refused by either limit is never placed and never billed, and you are warned once as you approach the day's figure rather than only when it stops. The identifier we keep on a privacy request is now a coded value that cannot be turned back into a phone number by anyone who reads it, and because that would otherwise have taken away your ability to check a request, there is now a way for you or an auditor to confirm that a request really is about a given person without ever holding the key that makes the code; requests filed before this change keep working exactly as they did, are still found, and are still protected. And putting records on hold now covers the saved profiles of recognised visitors too, so a hold preserves what it says it preserves: your own forget-me button and our privacy desk answer the same request the same way, and neither quietly removes what the other is keeping.

A visitor who is already signed in on your own website can now be recognised by the assistant: your site signs a short-lived token, the assistant greets them by name, and, only if you switch it on separately, carries a short, bounded list of what it may recall into their next conversation, which you can read in full on one screen and erase on the spot, and which deletes itself after the number of days you choose. Your team gained reply drafts too: when a colleague takes over a chat, one button drafts an answer out of your knowledge base with the passages behind it, an unsupported draft is called out as one, nothing is ever sent on its own, and a reply that came from a draft is marked on the record so your quality numbers stay honest about what a machine wrote and what a person did.

The monthly or quarterly quality report now has a screen of its own: read any past period in the console, see the target marked met or missed, open the numbers per agent, download the same spreadsheet the scheduled run emails, and generate a period on demand without sending it to anybody. Live chat also gained opening hours: give your project a time zone and the windows you are staffed, and a visitor who asks for a person outside them is told straight away instead of waiting for a desk nobody is at, while a chat a colleague already picked up is never cut off by closing time.

You can now tell an agent what your callers actually ask for, and every request is recognised as it is spoken and routed to the right answer, procedure, tool or person, with a list of the things people asked for that your definitions do not yet cover; every finished conversation is scored for how the caller felt at its start and at its end, which way it moved and what it was about, with a review queue of the ones that ended badly and were handed on, and a discovery view of recurring themes your topic list is missing; and every caller turn and every reply now passes safety screens that catch attempts to talk an agent out of its instructions, refuse a reply that would repeat those instructions or read out a credential, and optionally check an answer drawn from your documents against the passages it actually retrieved before it is spoken.

Your test questions can now re-run themselves on a schedule and again whenever the models behind your calls change, alerting you when the score falls against the previous run rather than against a fixed floor, so an assistant that quietly got worse without anybody touching it stops being something you find out about from a customer. And a wrong answer that a reviewer confirmed and pinned on one of your documents now flags that document for whoever owns your knowledge base: a queue ordered by how often each one caused a wrong answer, closed either by fixing and republishing the document or by dismissing it with a reason, plus a view of how much of your wrong answers can be pinned on the knowledge base at all.

An erasure request now reaches everything written about a caller, including the ratings, review notes and quality records that newer features had begun keeping their words in; a call needed for a legal case can be put on hold so an erasure stops cleanly instead of half finishing; and transcripts can be stored with names, numbers, emails and tax, social security or bank identifiers replaced by placeholders, tuned for Greek and English, with the limits of that detection written down openly.

Website chats can now reach your people: a visitor asks for a human, your team claims the chat and replies in the same window as a clearly human voice, and the agent takes over again when they finish. Every answer can also be rated with a thumbs up or down, and a wrong answer can be reported and reviewed together with the documents it relied on, so confirmed mistakes become test cases that stay fixed.

Your agents are now graded on what they say, and tested before they change. An automatic judge samples real conversations and measures how often answers are backed by your documents; the dashboard shows grounded answers, negative feedback and chats resolved without a person, with alarms you set per project. Before a deploy, a golden set of test questions (seeded from answers you confirmed wrong) runs against the draft, and a failing run can warn on the deploy or block it. A monthly or quarterly quality report lands in your inbox as a spreadsheet.

When a caller asks for their data to be erased or handed over, the request is now tracked from start to finish: recordings, transcripts and call records are erased everywhere or exported in one bundle, with a signed acknowledgement the moment the request is filed and an alert that keeps every request inside the legal one month deadline.

Put a search box in your own colours on your site so visitors search inside your published documents, and you decide which pages may show it and what it may read.

A knowledge document can now be sent for review before it goes live, and whoever checks it either approves it by publishing or hands it back as a draft.

Every knowledge document can now name the audience it is written for and how authoritative its source is, and your searches can be narrowed to exactly those.

Give us your website address and its pages become a knowledge document: we find the sitemap ourselves, follow the links when there is none, and when your site changes you can send us back for a fresh copy.

Connect your own file storage and bring knowledge documents in from it: running the import again refreshes only what changed, with no duplicates.

When the agent answers from your documents, the chat widget now shows what the answer was based on underneath it, with a link where there is one.

One knowledge document can now carry several languages, and a page you point at can be fetched again on a schedule.

Knowledge: your agents answer from your own documents, and every call record shows which passage was used.

Web widget: visitors talk or chat with your agent straight from your website, and you choose which tab the bubble opens on.

The post-call summary email now carries your own branding: logo, colours and subject line.

Developers section in the console: a live API reference and self-service API keys.

Start from 50 ready-made industry templates, and let the AI review your agent and diagnose a specific call with suggested fixes.

Multiple sub-projects under one account, each with its own branding, domain and users.

Warm transfer to a person, end to end: hold music, a spoken briefing before connecting, a readiness check, an urgent mode, and the whole transfer recorded.

Call-transfer controls: an allow-list of destinations and a per-call transfer limit, with every transfer's outcome recorded on the call.

After a long pause in the conversation the agent could lose its next reply; it no longer does.

Faster first answer, and the agent stops the moment you interrupt, even on echoey lines.

A Greek end-of-turn model: the agent waits until you have actually finished your sentence.

Full visibility per call: transcript, step-by-step timeline, and recordings with controlled access.