AGEphone, virtualized in the cloud.
The one talking is AI.

AI answers, introduces itself as part of your company, and finds out what the call is about.

Questions it can answer, it answers there and then. Who it speaks as, for what purpose and how far it may go are designed on screen as a business profile, and assigned per number or per line.

• Business profiles — a design document for the conversation: its purpose, the manner of speaking, what may and may not be said, and the voice to use.
• Different handling per number — the profile is chosen automatically from the number dialed, so one service can front several brands and several departments. The choice is made per call, so mixed traffic is handled correctly.
• An immediate, pre-prepared greeting — the greeting plays the instant the call connects, with none of the silence that leaves a caller saying "hello?". As well as writing the greeting, you can play a recording or generate speech from text (voice, 0.5× to 2× speed, and manner). Inbound and outbound are configured separately.
• Automatic FAQ answers — register the expected questions and answers per profile. AI searches them during the call and answers from what is registered. It does not invent an answer to a question that is not there.
• A natural conversation, properly closed — when the caller starts speaking, AI stops and listens (barge-in). Once the business is done it says a closing greeting and ends the call. It never simply goes silent, and never forgets to hang up and leave someone waiting.
• Several lines at once — several extensions are held at once and several calls handled in parallel, with a per-line cap on concurrent calls.

"Could I speak to Mr X, please?" — AI puts them through, like a switchboard operator.

A person's name, a job title ("could I speak to the president?") or the name of a desk are all accepted. AI asks what the call is about, rings the extension, tells only that person what it concerns, and then connects the caller.

1. Ask who they wantBy name, job title or desk
2. Establish the reasonIf a decline rule applies, it is politely turned away
3. Ring the extensionRinging tone plays while the extension rings (5–120 seconds)
4. Whisper the reasonThe reason and the number, spoken to that person only (whisper)
5. ConnectThe caller and the colleague are joined directly, and AI leaves
6. If nobody answersThe caller returns to AI with an apology, and an alternative desk can be offered

• Three transfer lists — names, job titles, and fixed destinations such as general reception, managed separately (up to 200 each).
• Linked to the extension directory — the company's extension directory can be included as it is, and updates to it flow through to transfers.
• Getting names right — phonetic readings can be registered for kanji names, and AI both speaks and listens using them (東海林 is read Shoji, for instance). Duplicate family names are flagged at registration.
• Decline rules — define whether to transfer and what to say, by type of enquiry such as sales calls (up to 50 rules). Keyword blocking can be used alongside, so even if AI is talked round, the transfer does not happen.
• Choice of transfer method — attended transfer (connect after the colleague answers) or blind transfer (connect immediately), set per profile. What happens when someone asks for a person not on the list is also configurable.

It does not only answer. AI makes the confirmation and notification calls too.

AI places calls using a profile written for outbound use. It waits until the call is answered before speaking, so nobody picks up to silence. The moment they answer, AI introduces itself.

• One call at a time — give it a destination and a profile and place the call from the console or through the API, including from your own business systems.
• Scheduled calls — three ways to schedule: daily at a given time (with a time zone), once at a given moment, or weekly on a given weekday and window. For regular notices and reminders.
• Campaigns (calling a list) — import a list from CSV (up to 500,000), set how many run at once (1–100) and work through it. Start, pause and stop from the console, with progress and results totaled in one view.
• Retries and automatic reporting — set how many retries and how far apart, and shift the window ("if the morning fails, try the afternoon"). Results are recorded on two levels: what AI reports (connected, no answer, busy, voicemail, objective achieved) and the call-level outcome.
• Do-not-call list (DNC) — registered numbers are never called. Matching absorbs hyphens and formatting differences, and is always checked immediately before dialing. Bulk registration (up to 100,000) supports obligations such as not re-soliciting.
• Calling windows, maximum duration and no double-dialing — a request outside the window does not fail; it is deferred to the next time the window opens. A second call with the same request key is never placed.

Traditional predictive dialers were bound to produce abandoned calls: they dial more than they can handle, and someone answers to nobody. Because Virtual AGEphone's AI can always take a call, it dials only as many as the concurrent capacity it holds — and abandons none.

One call leaves one complete record.

Every call AI handles is stored with the recording, the transcript, the summary and the outcome joined together, and can be reviewed and searched later.

• Two-track recording — the caller and AI are recorded on separate tracks, so where a person's speech ends and AI's begins is always distinguishable in the audio itself.
• Recording is optional — recording is switched on or off per business profile (on by default). With it off, what AI said and what the call was about still appear in the history as text, and no audio is kept at all.
• Transcription — a transcript by speaker with timestamps. On screen the conversation appears as speech bubbles by speaker, so the call reads as a piece of text.
• AI summaries — title, short summary, detailed summary, classification, priority, tags, action items, key topics and the temperature of the conversation, all generated. Summaries can be corrected by hand, re-generated under different conditions, or turned into a single image (photographic, illustrated or a four-panel strip) to share.
• Custom vocabulary — register company names, product names and personal names to improve transcription accuracy. As well as a company-wide list, each user has their own (up to 200 terms).
• Call history — direction, date and time, duration, the profile used, the outcome and the summary, in one list. Outbound job results and call records are linked to each other and followed from one screen.

The failures that AI on the phone is prone to are anticipated and held down in advance.

Each of these controls was worked back from a failure seen in real operation, and all of them are included as standard.

• Preventing an opening mishap — to prevent the classic failure of AI reading out its own configuration the moment it answers, the greeting is produced once by a dedicated mechanism. The first thing a caller hears is always the greeting you designed.
• Watching for silence — if the caller says nothing for a while, AI checks once ("can you hear me?") and, if the silence continues, ends the call. The interval is adjustable.
• Not hanging up too early — a guard against AI mistaking a noise at the start of a call for the end of the conversation and hanging up.
• Not answering itself — while AI is speaking, false detection of the caller's audio is suppressed, which prevents the loop where AI reacts to its own voice through a speakerphone. Towards the end of an utterance, only a loud interruption gets through.
• Barge-in with restraint, and conversational pace — AI stops when interrupted, but exceptions during the greeting and the closing words keep the conversation from becoming odd. How the end of a turn is judged (by meaning or by volume) and how sensitively is set per profile.
• It does not answer a call it cannot place — where AI cannot tell which business profile should handle a call, it does not guess. Starting a call with the wrong company name or the wrong premise is structurally impossible.

Configuration, monitoring and reviewing records all happen in a browser.

There is no work to commission and nothing to change on the handsets. The business side runs the day-to-day and IT handles the line and encryption settings, with permissions dividing the two safely.

• Editing business profiles — greetings, FAQs, transfer rules, voice and scope, edited on screen. Editing moves safely through draft → publish → apply, and each publication is kept as a version, so earlier ones can be looked back at.
• Unapplied changes made visible — if something is saved but not applied, a warning stays at the top of the screen. "I changed it but nothing happened" is prevented by the system.
• Line management — register and configure extension accounts (encryption, registration timers, NAT traversal, profile assignment, concurrent call limits). AGEphone configuration files can be imported, so nothing is lost in the move.
• Inbound routing and live monitoring — maintain the table that says which profile answers which number. See current calls and each line's registration state at a glance: whether it is properly attached to the phone network, and when it last succeeded. A line can be de-registered temporarily without deleting it.
• Outbound work, the extension directory and vocabulary — schedules, campaigns, the do-not-call list and the outbound history in one part of the console. The extension directory used for transfers and the vocabulary that sharpens transcription are maintained here too.
• Reading the SIP log — view and export protocol-level logs as a sequence diagram. Secrets are masked automatically, so the output can be handed to a carrier or used to isolate a fault as it is.

Calls and their records are sensitive company information, and this can be run to your IT department's standards.

Encryption in transit, separation of permissions and an audit log are all in place, which makes it straightforward to take through an internal security review.

• Call encryption — signaling over TLS and audio over SRTP, with client certificates (mTLS) available to prove where a connection comes from.
• Roles — eight roles separate who can do what in detail. The decision is made both in the interface and on the server, so it is not a restriction that exists only on screen.
• Safeguards on permissions — you cannot change your own role (which prevents locking yourself out), and an administrator cannot grant a role above their own (which blocks privilege escalation). A role change takes effect from the next action.
• An audit log of actions — who changed what and when. Configuration changes, applications and role changes can all be traced afterwards.
• Single sign-on (SSO) — SSO over OpenID Connect, with Google Workspace supported out of the box. Domains can be restricted, and accounts can optionally be created on first sign-in.
• Protecting secrets — line passwords, certificates and other secrets are write-only: once registered they are never displayed again, which closes off leakage through the interface.

A phone system must not stop. Assuming things will break, automatic recovery sits at the center of the design.

A brief network outage or a PBX restart does not require anyone to attend; the lines recover on their own.

• Automatic re-registration — after a brief outage or a PBX restart, lines re-register and recover by themselves. Retries back off as they go, so the phone network is not loaded.
• Lines are independent of each other — registration state, retries and concurrent capacity are managed per line, so a fault on one line or one call does not spread to the others.
• Continuous health monitoring — registration state, last success time and calls in progress are snapshotted every second across all lines, and the monitor checks continuously that things are alive. When they are not, the process restarts automatically.
• Making the state visible — the live screen lists the registration state of every line, so a partial fault — the system up but one line unregistered — is not missed.