Voice, knowledge and follow-up,
in one operating layer.
Six areas that were designed together rather than integrated afterwards. Nothing here needs a second product to be useful, and nothing has to be reconciled between them.
One agent per department, and they hand over to each other.
An agent has a job, a voice, a set of rules it cannot break, and the knowledge it is allowed to answer from. Four templates come built in; a new agent starts from one and is yours to change from the first minute.
- Rules are a list you can read, not a paragraph buried in a prompt.
- Draft, test, publish, roll back — live calls only ever use the published version.
- Switching between your own agents is silent; handing over to a person is announced.
- Language and accent can change mid-call when the caller asks for it.
Customer service
Answers general questions and looks up a caller’s details before helping.
Support
Handles problems, raises tickets, and hands over to a person when it cannot.
Sales
Answers questions about products and prices, and captures interest.
Reception
Greets the caller, finds out what they need, and routes them onward.
It answers from your business, or it does not answer.
Upload anything — documents, spreadsheets, a website — and it is structured into two kinds of knowledge, because a policy and a price list are not looked up the same way. What it finds is put in front of the agent before the turn starts.
- You review and confirm what was extracted before any of it is saved.
- Prose is searched by meaning; tables are looked up exactly, by identifier.
- The agent is grounded in the knowledge bases you bind it to, and refuses outside them.
- Contradictions between two documents are surfaced rather than silently averaged.
- Documents, PDFs, a websitePassages, searched by meaning
Policies, hours, how a thing works
- Spreadsheets, price listsReal rows, looked up exactly
A part number, a tariff, a stock level
One conversation history, whichever way they reach you.
A caller who phoned on Tuesday and messaged on Thursday is one person with one history, not two tickets nobody connected. The agent knows what was said last time, whichever way it was said.
- Your existing phone number — nothing has to be re-pointed to try it.
- Voice and chat on your own site, with no app to install.
- A WhatsApp call and a WhatsApp message are two channels rather than one, because they are two different things.
- One inbox behind all of them, not one product per channel.
- Phone call
- WhatsApp callNot yet
- WhatsApp message
- Web chat
- Voice on your site
The work is filed before the call ends.
A helpdesk built into the product rather than bolted beside it: queues, states, priorities, macros and service-level clocks, with the caller, the transcript and the summary already attached.
- Fields are filled from the conversation, not typed again afterwards.
- Service-level targets count business hours, including your holidays.
- Macros apply a set of changes and a reply in one action.
- Every change is on the ticket’s own history, with who and when.
Double charge on the March invoice — refund requested
- Queue
- Billing
- Priority
- High
- State
- Open
- Source
- Phone call
The numbers a contact centre is actually run on.
Not a page of charts about page views. Containment, resolution, speed of answer and cost per conversation, computed from the conversations themselves rather than from a counter that can drift.
- Every metric is defined once, so two screens cannot disagree about it.
- Where a satisfaction score is inferred rather than asked, it says so.
- Filter by agent, queue, channel, language or time of day.
- Export anything you can see, as the rows behind it.
- Containment rateResolved by the agent with no human involved
- First-call resolutionNo further contact on the same intent in seven days
- Average speed of answerArrival to the agent’s first word
- Service levelAnswered inside your own threshold
- Transfer rateHanded to a person, and why
- Cost per conversationTelephony, model and messaging, divided by volume
Isolation you can point a reviewer at.
Each business is its own tenant, separated at the database row by policies the application cannot talk its way around. Access is re-checked on the server for every request, whatever the interface showed.
- Row-level security in PostgreSQL, alongside the application’s own checks.
- Provider credentials encrypted with AES-256-GCM, never in a repository.
- Every privileged action written to an append-only audit log.
- Retention you configure, deletion that actually deletes, export on request.
- Transport
- TLS 1.3 only, HSTS
- Credentials at rest
- AES-256-GCM
- Tenant isolation
- Row-level security in the database
- Every privileged action
- Append-only audit log
- Recordings and transcripts
- Retention you set, deletion that runs
Three ways to answer a phone.
The differences are architectural rather than incremental — which is why they show up in the first thirty seconds of a call.
| Capability | Kasper | Speech-to-text plus a chatbot | Touch-tone menu |
|---|---|---|---|
| What the caller hears | A voice that speaks their dialect | A synthesised voice reading generated text | A recording and a list of options |
| How the turn is built | Audio in, audio out, one model | Speech to text, text to model, text to speech | A fixed tree of key presses |
| Understands a sentence it was not expecting | Yes | Yes | No |
| Answers from your own prices and policies | Yes | If somebody wired a search to it | No |
| Says it does not know rather than guessing | Yes | No | Yes |
| Leaves a transcript, a summary and a ticket | Yes | No | No |
| Hands over to a person mid-conversation | Yes | No | Yes |
| What it costs to add a new question | Add it to the knowledge | Retrain or rewrite the flow | Re-record the menu |
What the caller hears
- Kasper
- A voice that speaks their dialect
- Speech-to-text plus a chatbot
- A synthesised voice reading generated text
- Touch-tone menu
- A recording and a list of options
How the turn is built
- Kasper
- Audio in, audio out, one model
- Speech-to-text plus a chatbot
- Speech to text, text to model, text to speech
- Touch-tone menu
- A fixed tree of key presses
Understands a sentence it was not expecting
- Kasper
- Yes
- Speech-to-text plus a chatbot
- Yes
- Touch-tone menu
- No
Answers from your own prices and policies
- Kasper
- Yes
- Speech-to-text plus a chatbot
- If somebody wired a search to it
- Touch-tone menu
- No
Says it does not know rather than guessing
- Kasper
- Yes
- Speech-to-text plus a chatbot
- No
- Touch-tone menu
- Yes
Leaves a transcript, a summary and a ticket
- Kasper
- Yes
- Speech-to-text plus a chatbot
- No
- Touch-tone menu
- No
Hands over to a person mid-conversation
- Kasper
- Yes
- Speech-to-text plus a chatbot
- No
- Touch-tone menu
- Yes
What it costs to add a new question
- Kasper
- Add it to the knowledge
- Speech-to-text plus a chatbot
- Retrain or rewrite the flow
- Touch-tone menu
- Re-record the menu
See it answer one of your own calls.
Bring a question your customers actually ask. Fifteen minutes is enough to know whether this works for you.