Saltar al contenido principal

AI tools

AI tools do a specific job for you while Wiz works: today Web search (Wiz searches the web when a question needs public information) and Image Studio (a photo a customer sends becomes a styled, framed portrait). Wiz decides for itself when a tool helps. Each tool has its own switch, its own providers, its own limits and its own history.

Open Settings → AI → AI tools. You see the tools your plan includes and that you may manage (the AI management permission). Under each tool you see where it works today — for example Wiz, while it replies, or for Image Studio Incoming photos, Wiz, while it replies, Web widget and API — and, after Coming soon, the places it will reach later (flows, the voice assistant…).

Two switches decide whether a tool runs: your plan must include it, and the tool's own Turn on switch must be on. Both are off-by-default questions you answer once.

Settings​

Click Configure on a tool. The Settings tab has:

  • Turn on — the workspace switch.
  • Whose keys — Platform key (runs on our keys and is billed from your wallet) or My own key (your own provider account).
  • Provider order — the providers and models the tool may use, in order. The first one answers; if it fails or takes too long, the next one is tried. Add any model the tool supports, move it up or down, or remove it. You are not tied to one model: Image Studio offers, among others, Magnific, Google Gemini, fal.ai and OpenAI's image models; web search offers Serper, Tavily, Brave, SerpApi, SearchApi and Google Programmable Search. On platform keys, a provider the platform does not offer right now is marked Not on platform keys — skipped.
  • Your keys (with My own key) — for each provider in the order, pick a key you saved under Settings → Credentials. Only keys of the right kind are offered, and only their names are shown — never the key itself.
  • Channels — where the tool may answer. None selected means every channel the tool supports.
  • Limits — per person per day, per workspace per minute or per day, how many at the same time. 0 means no limit.
  • Behaviour — for example Check the result, Require wallet balance, Stay quiet when a person has taken over, or, for web search, Correct the question's spelling.
  • Automatic deletion — after how many days the customer's photo and the result are deleted. Empty or 0 keeps them forever — in a conversation, in the web widget and through the API alike. Nothing is deleted automatically unless you set a number here or in a profile; a person's request to erase their data still removes them.

Save sends only what you changed.

Profiles: events and campaigns​

A profile is one use case of a tool — an event, a campaign, a stand at a fair — and it is data you edit, not something we build for you. A workspace can run many profiles at once: twenty different events in one month is twenty profiles.

On the Profiles tab: New profile, edit (pencil) or delete (bin). A profile has:

  • Name, a short code (used in links, cannot be changed later), Status (active or paused), Priority (when several profiles fit, the higher number wins), Starts and Ends, and its Channels.
  • Trigger — Keywords: a message that starts with one of them links the conversation to this profile, for example from a QR code that opens WhatsApp with the word already typed. Only after a keyword makes the profile answer only people who sent a keyword first. Start on every photo (on unless you turn it off) starts a portrait from every photo by itself, without Wiz — the right choice at a stand. Turn it off and the photo goes to Wiz, which starts a portrait when the customer asks for one (see Wiz decides).
  • Look — the theme, the style variants and accents. A variant can be marked for groups (group, or min_people/max_people) or crowds (crowd): a group photo gets a group style that keeps every person, and above Max people (up to Crowd max) a crowd style on a best-effort likeness, with the crowd note text added to the caption. Larger crowds are asked for a smaller group.
  • Frame — your own PNG frame from the media library, or one we draw from your words and colours:
    • Band — a colour band at the bottom with title, subtitle, brand and name lines.
    • Brand strip — the photo stays the hero: the person's name on a tag in your main colour and the title over a soft shadow at the foot of the photo, and below it a strip in your strip colour with your logo, a thin divider, the subtitle (event), a details line (date · place) and a small credit line. Corner accent adds a curved wedge in your main and second colours to the strip's bottom-right corner; no text ever touches it.
    • Minimal lock-up — the photo fills the whole picture; only one edge (Lock-up position: bottom or top — keep it away from the face) gets a soft shadow with your logo and the title as one lock-up (Lock-up shape: logo above the title, or logo | divider | title), and a tiny credit logo in that edge's corner. The name line is optional. The same lock-up on every portrait. Safe area (width, height, anchor) keeps everything inside one shape: with a 1080 × 1920 story output and a 1080 × 1350 safe area at the top, cutting the bottom off gives a complete 4:5 post, and the providers are asked for a tall 9:16 picture where they can make one.
    • Logo — the media-library ids of your logo files: one for a light strip and, if you have it, the reversed (white) version for a dark strip; the frame picks the one that matches the strip colour. Logo height and Logo clear space follow your brand guide (empty = automatic). If a logo cannot be read, the portrait is still delivered, without it.
    • Long names are made smaller, then wrapped onto two lines — never cut off.
  • Output — size and file format.
  • Consent — Ask for consent before a photo is sent to an AI service, and whether to ask once per text version or every time. A photo whose owner says no is never used.
  • Web widget and API — whether visitors on your site and your own app may use this profile (both off until you turn them on; see below).
  • Automatic deletion — overrides the tool's setting for this profile.
  • Overrides — a different provider order or limits for this profile.
  • Texts — every sentence the tool sends (welcome, the consent question and its buttons, the refusals, the caption…), in as many languages as you like. An empty field uses the text shown in grey.

Deleting a profile keeps its past runs in the history.

Wiz decides​

When Image Studio is switched on, Wiz can start a portrait itself while it talks to a customer: someone writes "can you turn my photo into a portrait?", or sends a photo to a profile that does not start on every photo. Wiz picks the profile — only active profiles that need no keyword, plus one the conversation was linked to — and, if the customer asks for one, a style from that profile. It never chooses the photo: it is always the customer's latest photo in that conversation, from the last 24 hours. Without one, Wiz asks for it.

From there everything is as with a photo: the same checks and limits, the same consent question, the same "on its way" message, and the portrait arrives in the conversation on its own, and the run is listed under Recent runs like any other. Wiz starts at most one portrait per reply.

Wiz sees a photo only if your assistant can read images, or if the customer also writes something. If neither is true, keep Start on every photo on.

A flow's Agent step can be given Image Studio too: tick Image Studio among its tools.

On the stand screen​

When a live feed board shows AI portraits, the visitor is also asked whether their portrait may appear on the stand screen — only then, and only on the conversation channels (not in the widget or the API). The question, the button labels, the replies and the remove from screen keywords are the profile's texts screen_consent_question, screen_consent_yes, screen_consent_no, screen_consent_reply_hint, screen_consent_yes_ack, screen_consent_no_ack, screen_revoke_keywords (comma-separated) and screen_revoked, editable per language like the others.

In your web widget​

A profile can appear inside the chat widget on your website, with its own small screen. Edit the profile and, under Web widget:

  • Show in the web widget — on, and the profile is offered to visitors. It is off until you turn it on, profile by profile.
  • Widget icon — the icon of its button in the widget.
  • Only on these sites — hosts such as shop.example.com, or *.example.com for the domain and every subdomain. Empty: every site your widget runs on.
  • Also post the result into the visitor's chat — off by default. When on, the finished result is also added to the visitor's widget conversation: your team sees it in the inbox, the visitor sees it in the chat, and it is kept and deleted like the rest of that conversation. Only the result is posted, never the photo. It does not apply to the API, which has no conversation.

Under Texts, Widget: button label and Widget: short description name it in the widget. The consent text the visitor accepts follows the switch above, so it never promises more or less than happens:

  • Posting off: Consent question (widget and API, result not posted to a chat); empty = a default that promises nothing about a conversation.
  • Posting on: Consent question (widget, result posted to the chat); empty = your own Consent question (the conversation one), and if that is empty too, a default that says the result is kept in the chat.

What a visitor sees: a new button in the row above the message box (with several profiles, a list to choose from) → your consent text with your two buttons → Take a photo (the camera on a phone) or Choose a photo, and, if the profile has styles, a choice of style or Surprise me → Create → your progress text while it works → the result with Download, or your own refusal text (not a person in the photo, today's limit reached…). Nothing is uploaded before the visitor agrees, and nothing is posted into the chat unless you turned that on.

The same limits, wallet, kill switch and automatic deletion apply as in a conversation. On top of them, a visitor needs a live chat session, and a session, a network address and the whole widget can each start only a few runs in a short time.

From your own app (API)​

Your own app or backend can run a profile too — for example a photo booth app at your stand. Edit the profile and turn on Allow through the API under API. Then create a key under Settings → Developer → API keys and tick Run AI tools (tools:run) — or add that scope to a key you already have; its token stays the same. Your developer finds the calls, with an example, in the AI tools API reference. A retried call with the same Idempotency-Key never starts or charges a second run.

API runs have no conversation, so the widget's "post into the chat" switch does not apply to them.

A run from your app carries your own reference for the person (for example your user id), and the daily limit per person counts by that reference. Your app must show the profile's consent text and send back its version — that is how it confirms the person agreed. If you subscribe a webhook to tool_run.completed, your server is told when the result is ready.

Recent runs​

The Recent runs tab lists the latest runs: when, the status, which provider answered, the channel (or Web widget / API) and the cost. It is read-only. Web search records only which engine answered and how many results came back — never what the customer asked.

Test​

Tools that can be tried, such as web search, have a Test tab: type a query and Run. It runs once with the saved settings; it is a real call and is billed like one.

When a question needs public information your knowledge base does not have — an event's dates, a venue's address, a public fact — Wiz searches the web and answers from the results. Wiz decides when: it writes the search itself (so "and on Sunday?" becomes a real query), searches at most twice per reply, and does not search for greetings or small talk. Each search is billed as before and listed under Recent runs.

The answer still has to rest on something. On email, Wiz replies only when the answer comes from your knowledge base or from a search in that reply that found something; otherwise the email waits for a person. In Sources only mode, an answer with neither sends your fallback message. Where automatic handover for unsure answers is on, such an answer goes to a person instead.

A flow's Agent step can be given web search too: tick Web search among its tools.

Web search used to be a tab of Settings → AI assistant; that tab now brings you here, and your settings carried over unchanged — the same engine, the same key, the same switches. If you had typed your engine key directly into the old screen, it now appears under Settings → Credentials as Web search key — <engine>.

Your MCP servers​

If you run your own system — a catalogue, a booking engine, an order database — and it offers an MCP server, you can connect it here. The assistant then uses that system's tools while it talks to customers, deciding for itself when a tool helps: looking up a fair's dates, checking stock, sending an enquiry.

Add as many servers as you like under Settings → AI tools → Your MCP servers → Add server:

  • Server address — the server's Streamable HTTP address. Only public https:// addresses are accepted; an address inside a private network is refused when you save and again on every call.
  • Authentication — a bearer token, a key in a header you name, or none. The token is stored encrypted, sent only to this server and never shown again.
  • Offer to the assistant — on: the assistant can use the server's switched-on tools in every conversation. Off: only flow agent steps that pick them use them.

When you save, the server's tool list is fetched. No tool reaches the assistant until you switch it on. Open Show tools on the server's card and switch on the ones you want. For each tool you can also:

  • mark that it changes something outside the conversation — sends an enquiry, signs someone up. The assistant then asks the customer to confirm the details first and calls the tool at most once per reply. Tools the server marks as read-only start unmarked; everything else starts marked.
  • write your own description, which replaces the server's for the assistant.

If the server changes a tool. The assistant always uses the description you approved. When you press Refresh tools and a switched-on tool now says or takes something different, it is switched off and marked Changed — review. Read it, and switch it on again if it is still right. A server cannot change what an approved tool tells the assistant behind your back.

What the assistant does with the answers. It treats them as information, never as instructions. If a server is slow or down, the call stops after the time limit you set; after five failures in a row the server is skipped for five minutes, so one broken server does not slow every conversation. The reason shows on the server's card.

Many tools. There is no limit on servers or tools. When the assistant holds more than 30 tools in a conversation, the tools from your MCP servers are no longer all shown to it at once: it is told which servers it can use and how many tools each has, and searches them by what the customer needs. The tools it finds are added for the rest of that reply. ConnectWiz's own tools are always shown. This keeps a long tool list from slowing replies and from blurring the assistant's choice; a clear tool description helps it find the right one.

Personal data a customer gives the assistant for such a tool — a name, an email address for an enquiry — is sent to your server. You are responsible for that system and for telling your customers about it.

MCP servers are set up on the web; the phone app does not list them.

On the phone​

Open Menu → AI training → AI tools. Each tool has its switch, Whose keys, the provider order (add, remove, move up and down), the key for each provider, channels, limits, behaviour, automatic deletion, a test where the tool has one, its profiles and its recent runs. Profiles are listed on the phone, with Web widget and API when they are offered there; Create or edit profiles on the web opens this page in the browser, because a profile's texts, frame and styles are easier to write at a desk. Under each provider, Add a key stores a new key for it and comes back with it chosen (Credentials).