FAQ

The product

What is CoPhrase? An AI agent platform. You build an agent, train it on your business, and connect it to the messaging channels your customers already use. It handles inbound sales and support conversations — answering questions, qualifying leads, booking calls, resolving support requests — and pulls in a human when it should.

Is it a sales bot or a support bot? One agent, both jobs. A support question gets a grounded answer or an escalation. A sales conversation gets qualified against the questions you configured and driven toward the action you chose. It is intent-aware, not two separate products bolted together. See persona and actions.

Which channels does it work on? Six: web chat, Discord, email, WhatsApp, Telegram, and Instagram DMs. Messenger, WhatsApp Business API, iMessage, SMS, and connected Gmail/Outlook inboxes are being built. See channels.

Can I run more than one agent? Yes. An organization can have several agent configurations, and each channel connection is attached to one of them — for example a support agent on your help widget and a sales agent on Instagram.

Getting started

How fast can I go live? The setup itself is short: connect a channel, point the agent at your website so it has something to answer from, set the persona, test it. See quickstart.

Web chat is the fastest — a script tag, no account to link and no approval to wait for. Email is close behind: one forwarding rule. Discord and Telegram take minutes. WhatsApp takes minutes to connect but you should read the ban-risk warning first. The channels still being built (Messenger, WhatsApp Business API, SMS) carry calendar-time provider approvals that nobody can rush.

What does the agent need before it's useful? Knowledge. Phase 2 accepts a website URL or hand-written FAQs. File upload is future work, and live Firecrawl still needs provider acceptance. See knowledge.

How the agent behaves

What happens when the agent doesn't know the answer? It does not guess. When the knowledge base has nothing confident to say, the agent gives a templated can't-answer response or escalates to a human, rather than free-generating a plausible-sounding fact. Where it lands is configurable — see handoff.

Will it make up a price or a policy? It is not supposed to, and the guardrails are built for exactly this. Answers about prices, policies, and hours have to come from your structured business profile or a cited piece of your knowledge base. If it isn't in there, the agent won't invent it — it will escalate. The most common cause of a wrong price is that the right price was never added, or that the knowledge is stale. See troubleshooting.

Can a human take over? Yes — from the inbox. A thread can be taken over, replied to by a human, and handed back. Approval mode — where the agent drafts and a human sends — is being built. See handoff.

What language does it reply in? The agent mirrors the language the customer wrote in.

Can it handle images and audio messages? Discord, WhatsApp, Instagram, and email carry inbound images, files, and audio, and the agent reads them. The web chat widget is text-only. Billing is described in the rate card.

Planned Phase 4 credits and billing

What does a reply cost? 1 credit — 10 cents — on Pro, or 0.25 on Core. You are billed per agent reply, not per customer message: five rapid-fire messages answered together count as one reply. On BYOK the AI is free, because you pay your provider directly. See billing.

What else costs credits? Tool actions (0.25, and only when they succeed), web search (0.5), knowledge training (a small per-page charge), and a monthly charge per connected WhatsApp number. Messages your agent doesn't answer, replies your team sends by hand, and anything you try in the playground are free.

What happens if I run out of credits mid-month? Your agent does not go dark on a live customer the instant you hit zero. You get a loud low-balance and then empty-balance alert, and a grace window rather than a hard stop. Top up from your dashboard at $0.10 per credit — purchased credits never expire, so a buffer costs you nothing.

Do my credits roll over? Plan credits refresh at the start of each billing cycle; they do not accumulate. Credits you purchase never expire.

Can I use my own AI provider key? BYOK configuration is a Phase 3 target. Its designed failure posture is fail-closed: a dead key pauses the agent rather than silently switching payer/provider.

Agencies

Can an agency resell CoPhrase? Agency resale, white-label domains, client workspaces, and the Agency plan are PROJECT Phase 4 targets. See agencies.

Can one client see another client's data? No. Each client is a separate tenant with its own channels, knowledge, contacts, and conversations, and isolation is enforced in the database, not just in the interface. Even the agency itself does not get automatic read access to a client's rows — someone from your team has to be explicitly placed into that workspace, and the client can see that they're there. See security.

Data and integrations

Do you train models on my conversations? No. Conversation content is sent to model providers to generate a reply and is not used to train CoPhrase's own models.

Is there an API? There will be a versioned public API with organization-owned keys, plus signed outgoing webhooks. The reference isn't published yet — see API and webhooks for the shape it will take.

How do I delete my data? Delete an individual contact from the dashboard, or request full account deletion by email. See Data Deletion.