Relay
In buildLaunch AI voice and web agents on your own number.
Relay is our platform for building, testing and running voice and chat agents — a visual builder, your choice of model, telephony or an embeddable widget, and the evaluation and analytics to know whether it is actually working. We build it because running our own product keeps us honest about what happens after launch.
- RelayGood evening, you're through to the automated line. This call is recorded — is that alright?
- CallerYeah that's fine. I'm chasing a delivery, reference GL double eight two one three.
- RelayThanks. That consignment left the depot this afternoon and is booked for tomorrow morning.
- CallerCan I move it to the afternoon instead?
- RelayI can. There's a slot after midday tomorrow — shall I take it?
- CallerPlease.
- RelayDone, and the confirmation is on its way to the address on the booking.
The problem
- 01
Phone support does not scale the way a ticket queue does. Every call needs a person for its whole duration, so capacity is headcount, and the demand arrives in spikes nobody staffs for.
- 02
The calls that get dropped are the routine ones — a status enquiry, a booking change, a question the systems could already answer. They go to voicemail out of hours and get handled by hand the next morning.
- 03
Outbound qualification has the opposite problem: the conversations are short, the volume is high, and the output that matters is a structured record rather than a call disposition somebody typed afterwards.
What it does
Visual agent builder
Design the conversation, the tools and the escalation paths. Versioned and testable, so a change is reviewable rather than a rebuild.
Bring your own model
Claude, GPT, Gemini or open weights you host yourself. Model choice is a configuration, so a better one is a switch rather than a migration.
Telephony and web
Inbound and outbound calling on your own numbers over SIP, or an embeddable widget running the same agent definition.
Grounded knowledge base
Retrieval over your documents with citations, so an answer can be checked back to the source it came from.
Tools, webhooks and integrations
Call your systems mid-conversation — CRM, calendar, helpdesk — and write the record before the call ends.
Transcripts, evals and analytics
Every call transcribed and scored against your own evaluation set, with A/B testing between agent versions.
Consent, redaction and audit
Recording consent in the flow, PII redaction before anything reaches a provider, configurable retention and a full audit trail.
How it works
A call is a budget spent across six hops, and only one of them is the model. Each note below says what happens at that hop and what governs its cost.
- 01
Caller
An inbound call on your own numbers, or an outbound one. Nothing about the published number changes and the daytime routing is untouched.
- 02
SIP / WebRTC
The carrier hop. You do not control it and you have to count it — a budget measured from the model call outwards is the measurement that flatters.
- 03
Speech to text
Streaming, so the agent has partial text while the caller is still speaking. Endpointing decides when they have stopped rather than paused, and tuning it aggressively buys latency and costs interruptions.
- 04
Agent loop
Retrieve, reason, call a tool, decide. The turn either completes, asks one clarifying question, or hands over — there is no fourth branch, which is what keeps the failure mode legible.
- 05
Text to speech
Measured to the first byte of audio rather than to the finished utterance, because that is the moment the caller stops waiting.
- 06
Back to the caller
And a transcript, the tool calls made, and a score against your evaluation set — written before the call ends rather than reconstructed after it.
What we are building toward
These are targets, not measurements. Relay is in build and has not taken a live call — when it has, these become numbers with a date on them.
- Target
Sub-second turn-taking on a live call
- Target
Every answer grounded in a cited source, or no answer
- Target
A published cost per call before you commit to a volume
- Target
Barge-in handled as behaviour, not as an error state
- Target
Every call transcribed, scored and replayable against your eval set
What people use it for
- Inbound support triage
- Answer, identify the caller, resolve the routine cases and route the rest with context attached.
- Appointment booking
- Check real availability, book it, confirm it and write it to the calendar during the call.
- Outbound qualification
- Structured conversations that produce a qualified record rather than a call disposition.
- IVR replacement
- Natural language in place of a menu tree, with the same routing guarantees behind it.
- After-hours coverage
- Handle what can be handled overnight, and queue the rest for the morning with a summary.
Integrations
Categories rather than logos, deliberately. Relay is in build, so it has no integrations to claim yet — the tool layer is typed tools and webhooks, and reaching your systems is part of the engagement rather than a checkbox.
- CRM and sales tools
- Look up an account mid-call, write the outcome before the call ends, and hand a qualified record to the system your team already works in.
- Calendars and scheduling
- Check real availability rather than a cached copy, book it during the conversation, and confirm it while the caller is still on the line.
- Helpdesk and ticketing
- Open the ticket with the transcript, the caller's identity and the tool calls already made attached, so nobody starts from the beginning.
- Telephony and carriers
- SIP trunking on your existing numbers. No porting, no change to daytime routing, and the agent answers only what you route to it.
Security and compliance
Recording consent in the flow
The agent says what it is in the first sentence and takes consent before recording, because a caller who does not know they are talking to software cannot consent to being recorded by it.
PII redaction before the provider
Card numbers, identifiers and anything you nominate are redacted from the transcript before it reaches a model provider, not after.
Configurable retention
You choose how long audio and transcripts live, separately. Deletion is a real deletion, including from the evaluation set if you ask for it.
A full audit trail
Every tool call, every retrieval and every model version, per call, replayable — which is also what makes an evaluation set worth anything.
Regional pinning
Inference and storage pinned to a region you name, with the provider agreements to match. Where a provider cannot offer it, we say so rather than routing around it quietly.
Where this sits
There are good platforms in this category, and they compete on latency and on price per minute. Relay's difference is not that it is cheaper. It is that the team building the platform also builds the integration — so a company that needs the agent wired into its own systems gets one engineering team rather than a platform plus an agency, and the person tuning the prompt is the person who wrote the tool it calls.
Join the waitlist
Join the waitlist
Relay is in build. Leave your email and we'll tell you before it opens — that is the only thing we'll use it for.
Questions
Can Relay use our existing phone numbers?
Yes. It runs on your carrier over SIP, so there is no porting and no change to how calls are routed during staffed hours. You choose which calls reach the agent — out of hours, an overflow queue, one line — and the rest are untouched.
Which model does it use?
Whichever you choose. Claude, GPT, Gemini or open weights you host yourself, set as configuration rather than built in — so moving to a better one is a switch and a re-run of your evaluation set, not a migration.
What happens when the agent cannot handle a call?
It hands over, and the transcript, the caller's identity and the tool calls it already made go with them. Handing over too late is the expensive failure, so the escalation conditions are part of the agent definition and are scored in evaluation like anything else.
How do we know it is working?
Every call is transcribed and scored against an evaluation set built from your own traffic — task completion, handover quality, turn-taking latency and interruption recovery, scored separately so a fluent answer from the wrong lookup is still a failure. Releases are gated on the suite.
Where does our data go?
Inference and storage are pinned to a region you name. PII is redacted before anything reaches a model provider, retention for audio and transcripts is set separately by you, and nothing is used to train a model — ours or a provider's.
Are you certified — SOC 2, ISO 27001?
No, and we will not imply otherwise. We are a young company and hold no certifications yet. What we can show you is the posture: our subprocessors, our provider agreements, how access and secrets are handled, and what happens to your data. That is on the security page, and we are happy to answer a questionnaire.
Want the team that runs its own product?
Whether you need Relay or something built from scratch, the useful first conversation is about the constraint rather than the technology. Tell us what has to be true, and we will tell you what it takes.
contact@algologix.coWe reply within 24 hours.