AI agents

Not a chatbot in front
of a phone system.

The AI is an endpoint the router dials exactly as it dials a person — on the phone, and on web chat, SMS, WhatsApp, email and social. That one architectural decision is why a handoff keeps its context, why a supervisor can listen to the AI mid-call, and why the same reports cover both.

Every channel can be handled by the AI or by a person. Which one is a branch in your flow, not a property of the channel.

What happens on the call

Voice first, because it is the hardest case — a caller will not wait for a model to think, and will talk over it. Text works the same way with more slack.

  1. 01

    It answers in about a second

    1.17 seconds from the call being answered to the AI speaking, measured on the production system. Fast enough that nobody says "hello?" into silence.

  2. 02

    It listens while it talks

    Interrupt it and it stops, the way a person does. Speech-to-speech rather than transcribe-think-speak, which is what removes the pause that makes bots obvious.

  3. 03

    It answers from your content

    Upload what your business knows and it retrieves from that, rather than guessing or escalating a question you have already answered in writing.

  4. 04

    It does things, not just says things

    Books the appointment, takes the details, looks up an order — through a permissioned, audited connection to your systems.

  5. 05

    It hands over cleanly

    To the right queue, with the reason and everything gathered on the agent's screen before they say hello.

Example call

AI
Thanks for calling — how can I help?

Caller
Do you have the 18-millimetre birch ply in stock?

AI
We do — 14 sheets at the Oakland yard. Collection or delivery?

Caller
Actually — what's your cut-off for next-day?

AI
Four o'clock. Shall I reserve the sheets?

Caller
Can I speak to someone about a trade account?

AI
Of course — putting you through to sales now.

An illustrative conversation. The interruption at line four is the part worth noticing: it changes subject mid-sentence and the AI follows.

Guardrails that aren't optional

An AI answering your main line is a liability question before it is a technology question. These are structural, not settings someone can forget.

It cannot be talked out of its instructions

Platform rules sit in a locked layer that the instructions you write cannot override, and a caller cannot argue past.

It says it is an AI

Disclosure is part of the locked layer, not a line of your prompt that could be edited away.

Every action is audited

Tools are schema-validated and classified read or write. Writes require confirmation, carry an idempotency key and are rate-limited. All of it is logged.

It fails to a human

If the model is unavailable for any reason the call routes to a person. There is no path where a caller reaches dead air.

You control when it is used

The flow decides. Branch around the AI for a repeat caller, out of hours, or for a queue you want people on — it is a widget, not a global switch.

It is graded like an agent

AI-handled conversations are transcribed, scored against your QA form and reported on beside your team's.

The same agent, everywhere else

Digital is not a lesser version of this. It is the same agent, the same guardrails and the same knowledge — plus the AI work that happens around a conversation rather than in it.

On every text channel

Web chat, SMS, WhatsApp, Messenger, Instagram and email — the same agent and the same knowledge, escalating to the same human queues.

Agent assist

A suggested reply grounded in your knowledge base, next-best actions and live sentiment while the conversation is still open.

Automatic wrap-up

A summary and a suggested disposition written for the agent after the call, from the transcript.

Quality scoring

Every transcribed call and ended chat graded against your form, with reasoning — not a sample.

Outbound

The AI places reminder and follow-up calls, behind a do-not-call check and an audit gate.

Learning loop

Recurring escalations are clustered into draft knowledge entries you accept in one click, so "the AI couldn't help" becomes the content that fixes it.

What it is good at, and what it isn't

Useful to know before you point your main line at it. We would rather you found the edges here than on a live call.

Reliably good
  • Answering questions your business has documented — hours, prices, services, policies.
  • Booking, changing and cancelling appointments against your availability.
  • Taking a structured message: name, number, address, job description, urgency.
  • Qualifying an enquiry and routing it to the right queue.
  • Looking up an order or account through a permissioned connection and reading it back.
  • Recognising when it is out of its depth and handing over cleanly.
Not what to use it for
  • Anything clinical, legal or financial that needs a qualified person. Route those straight past it.
  • An emotionally difficult conversation — a complaint, a bereavement, a cancellation someone is angry about. Branch these to a human on the first signal.
  • Negotiation, or a decision that has to be someone's judgement call.
  • Anything you have not written down. It answers from your content; it does not guess what your business would say.

All of that is a flow decision, not a limitation you have to work around. The flow builder has a dedicated widget for choosing between AI and a person, so you decide exactly which calls it ever touches.

Setting one up

An afternoon, not a project. There is no model to train and no script tree to draw.

  1. 01

    Tell it about the business

    Who you are, what you do, how you want it to sound, and what it must never do. Plain sentences — this is not a scripting language.

  2. 02

    Give it your content

    Paste in or upload what you already have: price lists, service descriptions, policies, the answers you find yourself repeating. It retrieves from these rather than inventing.

  3. 03

    Decide when it escalates

    Set how eagerly it hands over, and which queue it hands to. Start eager; you can tighten it once you have listened to a few calls.

  4. 04

    Connect anything it should act on

    Your calendar, your order system, your CRM — through a permissioned connection where every call is schema-checked, classified read or write, and logged.

  5. 05

    Put it in a flow and call it

    Drop the AI widget into a flow, point a number at it, and ring your own line. Listen before you commit to anything.

Judge it yourself

Point a number or a chat widget at it.

It takes a few minutes and costs nothing. Hearing it answer your own line — or watching it handle a chat — tells you more than any page we could write.

Google or Microsoft · 30-day free trial · $15 credit · no card required