Questions that come up during setup

No. Every workspace gets share links and QR codes on the console’s Share page that open your assistant on the shared sandbox with your routing command pre-typed. Set a routing tag in Settings and you can be talking to your own assistant in a couple of minutes. See the sandbox link.

Most teams are answering real messages in an afternoon: connect a channel, upload your documents, turn it on. There are no scripts or decision trees to build. The part worth spending real time on is the material itself — see what makes a good knowledge base.

Check its status on the Knowledge base screen. A source has to finish processing before it can answer, and the page polls so you can watch it. If it failed with “no readable text found”, the file is a scan or an image-only PDF with no text layer to extract — re-export it as a text PDF or paste the content in directly.

Because nobody would be. Handing over to a person counts as available only when your workspace has at least one confirmed escalation recipient, and an address is confirmed by clicking the link it is emailed. Add one under Settings → Escalation alerts and the promise becomes true — which is the only condition under which Tephlo will make it.

No, and not on any plan or in any workspace. There is no ordering system, no cart, no checkout and no booking system anywhere in the platform, so there is nothing for a setting to switch on. A reply that promised one of those is replaced before your customer sees it. The full list is on what it cannot do.

It can quote the stock recorded in your catalog at your last update, and it is required to describe it that way. There is no live inventory lookup, and a reply claiming to check current or real-time stock is treated as a fabricated capability and replaced. Keep the catalog sheet current and treat stock in a chat as indicative.

Tephlo answers from what you have uploaded, not from your website and not from general knowledge. If a fact is not in your knowledge base or catalog, there is no grounded answer and the conversation escalates rather than the assistant guessing. Upload the page.

It says so and hands off. The conversation lands in your escalation queue with its full history and the reason, and an agent takes over in one tap over the same channel. The reasons are listed under human handoff.

Yes. Exact phrases on the Never say list are enforced in code, not merely requested: a reply containing one is withheld and the conversation goes to a person instead. You can also list situations where you would rather a human took over regardless.

No, and this trips people up. The conversation-goals screen describes what your team does; the assistant guides customers toward it and explains how you take it from there. Writing “checkout” as a goal does not create a checkout. The same screen shows the live list of what your assistant can genuinely do right now, so the two can never be confused.

No, and this is enforced in code rather than requested of the model. A customer saying yes records an action as under way; only a step that genuinely completed can be reported as complete, and a failure is stated plainly and handed to a person. See working memory and action honesty.

WhatsApp Business, Telegram, and an embeddable web chat widget for your own site — all answering from one knowledge base and landing in one queue. See channels.

Yes, with nothing to configure. It detects the language each customer writes in and replies in it, including handoff notices and confirmations. Your knowledge base can stay in one language; identifiers like order references are repeated verbatim rather than translated.

Understanding media is still rolling out and is off by default. Until it is enabled for your deployment, a customer who sends one gets a polite note in their own language offering a person — never silence, and never a guess about what the file said. Location pins and contact cards are handled today.

WhatsApp only permits free-form messages inside a 24-hour service window after the customer’s last message. That is Meta’s rule, not ours; the window is re-checked at send time, and a message outside it is not sent rather than failing quietly somewhere else. Telegram has no such window.

Yes. Every query is scoped to your workspace on the server, and the browser never supplies a workspace identifier. Isolation is tested against a decoy workspace holding the same products at different prices, so a leak would show up as a wrong price rather than as something nobody notices. See tenant isolation.

Yes, from the console’s Privacy Center — no database access or support ticket required. Both actions need a reason, deletion needs the identifier typed twice, and both are recorded in an audit that holds hashes and counts rather than raw identifiers. See privacy tooling.

Only if you say so. Permission defaults to off, and a workspace with no setting recorded is treated as having refused. Even then, conversations are named one at a time and approved by a person — there is no bulk export and no automatic sampling. See consent and retention.

No. Card numbers, CVVs, PINs, passwords, one-time codes and identity-document numbers are refused, and the refusal is applied both when a collection plan is built and when your own checklists are rendered. Adding one to a checklist does not make it collectable.