AI Insights & Actions

Where: Communication → Message automations → an automation that answers from a catalog

Module: Answer Catalog — the reading and the suggestions are part of the Plus plan and up; acting automatically is part of Prime and up. See Prices and plans.

What it is

An AI Answers automation answers the questions it finds in your catalogs. This page is about the two things the same automation can do besides answering:

Each action has a mode, set on the automation:

• Off — nothing happens, and the AI is not even asked about it.
• Suggest — the action appears in a bar above the Inbox’s message box, and one click does it.
• Automatic — FlyB does it, and logs it in the conversation.

Every action starts Off. Suggest is what the editor offers first; Automatic is switched on action by action, and only on the plans that include it.

What the AI never does

• It never decides what your records already know. Whether the contact has an appointment coming up, an open deal, an open task, which stage they are in, who is assigned, whether the WhatsApp window is open — FlyB reads those from your data and never asks the AI.
• It never writes a message — with one exception you choose. A booking link, a payment link and a follow-up written by you go out exactly as you wrote them. The only text an AI writes is a follow-up set to Written by the AI, and that text is checked before it leaves (see The follow-up’s text).
• It never explains itself in words of its own. Every suggestion says why it was made with the question the AI was asked and how sure it was — “The customer seems to want to book (91%)” — never a sentence an AI wrote.
• It never repeats itself. Whether an action may happen again is decided by FlyB, by fixed rules per action (see Repetition), never by the AI.

Reading the conversation

Open an automation that answers from a catalog and go to AI reading in the editor’s left column. Read the conversation turns the whole layer on — the actions below act on what it reads, so none of them can be switched on while it is off.

Under the switch, the editor shows an estimate: how many readings the automation would have made in the last 30 days, from its own history — “about 430 pauses → about 430 Flights a month” — or No history yet for a new automation.

A few things to know:

• Only a customer’s message starts a reading. A reply from your team pushes a reading that is already waiting further back — the conversation has not paused yet — but never starts one.
• An automatic message keeps the answer’s pace. While the booking link or the payment link is set to Automatic, the reading no longer waits for the pause above: it waits as long as the automation’s AI answer does — the automation’s delay, or a few seconds when it has none — counted from the customer’s (or your team’s) last message, and it comes right after that turn’s answer. The customer gets the booking options while still in the conversation, not a minute later. The pause above applies again once neither is Automatic.
• It covers what the automation covers. The automation’s channel, numbers and conditions decide which conversations it reads. Its keyword filter does not: keywords decide which questions get an answer, not which conversations are worth understanding.
• A voice message is read as its words, when your account transcribes them — the reading waits a little for them.
• Spam, automatic messages and requests for the AI’s instructions are not read (and not charged) — the same messages AI Answers ignores. An abusive or off-topic message is read: that is exactly when dissatisfaction matters.
• Two automations reading the same conversations each read it, and each is charged. The editor warns you when another running automation already reads the same conversations.

The signals

Buying intent is not a forecast: Hot means the conversation reads like someone settling details or deciding, not that 75 out of 100 such customers buy. That is why the Inbox shows the band first and the percentage only on hover.

The reading also says who is expected to write next — the customer, your team, nobody (the matter is settled) or unclear — and flags a customer who declined (“not interested”, “I chose someone else”, “don’t contact me”) or asked for a person. Those drive the actions’ guards, below, and show as chips in the Inbox.

Actions

AI actions → Overview, in the editor, lists every action with its mode and a one-line status (3 delays · written by the AI, 2 tasks, a deal has no name). An action can be switched to Suggest or Automatic once its own section is complete — a task needs at least one task, a deal at least one deal.

Suggest and Automatic

Behind every action is a question the AI answers with a probability — does the customer want to book?, which of these offers has the customer chosen? Two thresholds decide what happens:

• at or above FlyB’s suggestion threshold, the action is suggested (FlyB sets it; a bar that suggests too much is fixed by dismissing suggestions, not by a setting);
• at or above the action’s Minimum confidence to act automatically — a percentage from 50% to 99% you set on each action set to Automatic — it is done.

An Automatic action the AI was not sure enough about is suggested instead, with an amber dot. The follow-up is the exception: set to Automatic, it is scheduled after every customer message and its minimum is applied when it is due — below it, it is not sent (see When it goes out).

When an automatic action is held back

Even above its minimum, an automatic action becomes a suggestion — with an amber dot and the reason in its tooltip — when:

And some things are not suggested at all:

• A customer who declined (“not interested”, “don’t contact me”) gets no booking link, no deal, no payment link and no follow-up suggestion. An automatic follow-up stays scheduled: when its moment comes, the AI reads the conversation again, sees the customer declined, and sends nothing. Tasks, stages and resolving still apply: losing a customer is when “send the cancellation form” makes sense.
• Resolving is only suggested when the matter is settled and nothing else is suggested in the same reading — a booking link and a resolve in the same breath contradict each other.
• A suggestion you dismissed is not suggested again in that conversation for 24 hours, nor done automatically.

Repetition

How often an action may happen is fixed, per action, by FlyB:

What your team does by hand is not limited by these, with one exception: a second open deal of the same kind is refused — Open the existing deal — because the first one is still open.

The follow-up

Brings back a customer who stopped replying while something was left open — a quote, a decision, a document, a date to pick.

The follow-up uses the automation’s own quiet hours; there is no second set.

When it goes out

• Automatic: every customer message schedules it, and the decision waits for its moment. When the customer writes, the AI’s reading schedules the follow-up — or re-schedules the one already waiting — and decides nothing else: not whether it fits, not whether the customer declined, not who owes the reply. All of that is decided when it is due (below). If the reading could not happen (no Flights left, the day’s limit reached), it is still scheduled, at the default delay. A follow-up a member scheduled is never changed by a reading.
• Suggest: only suggested when the conversation is waiting on the customer — your team answered or offered something, and the customer went silent.
• The delay follows what the customer said. “I’ll talk to my husband and get back to you tomorrow” picks the option closest to a day; “next week”, the one closest to a week; nothing said, the default. With Let the AI pick off, it is always the default.
• Any new message postpones it. The delay counts from the conversation’s latest message, from either side — a reply from your team, a message from the customer, another automation’s message — so a follow-up of one day goes out one day after the conversation actually went quiet. A snooze holds it until the snooze ends.
• While the customer was the last to write, it waits for your team’s reply: the Inbox says Follow-up on hold: waiting for your reply to the customer. A customer still waiting for your answer never gets a nudge.
• At its moment, the AI decides. It reads the conversation as it is then, and the follow-up goes only if a nudge fits — on Automatic, at least the follow-up’s Minimum confidence to act automatically — the customer did not decline, and nothing was settled. So a conversation your team closed with “Great, we’re done!” without resolving it gets no follow-up. This check costs nothing.
• Then, right before sending, FlyB checks the conversation one last time: if somebody wrote in the meantime, the follow-up waits again.

Attempts. With 2 or 3 attempts, each one is scheduled after the previous one went out, with its own delay. They are for a customer who never answers: if the customer writes after any attempt, the next ones are cancelled — and, on Automatic, that message schedules a new follow-up.

The follow-up’s text

• My message — your text, its variables filled in for the contact.
• Written by the AI — a language model writes one short message in the customer’s language (at most 400 characters, with one question), from the conversation and your instructions. Before it is sent, every number, link, e-mail address and variable in it must appear in your own messages of the conversation or in your instructions — never only in the customer’s words — and it must not repeat the previous attempt. A text that fails goes to the next model; when none passes, your My message goes instead if you allowed it, otherwise nothing is sent and the history says so. The customer’s first name may be used as a greeting.

The 24-hour window

On a number connected through the Cloud API, a free-form message can only go out within 24 hours of the customer’s last message — and a follow-up of a day or more almost always lands after that. So on those numbers you choose, for when the window has closed: an approved template (sent from the conversation’s own number, when it has that template approved) or send nothing. The editor asks you to choose as soon as a delay of about a day or more can reach a Cloud API number. WhatsApp Web numbers and FlyB Chat have no window.

On FlyB Chat, where the visitor may have left the page, the follow-up is also sent to their e-mail address, the same way your team’s unread replies are (see When the visitor has already left).

When a follow-up is cancelled

Follow-up or “Customer quiet for X hours”?

Both can run on the same numbers; they do different jobs.

Since any message postpones the AI follow-up, a Customer quiet for X hours rule that fires first pushes it back by its full delay. The editor says so when both cover the same numbers.

Tasks

A list of up to 10 tasks the AI may suggest or create. Each one:

Several tasks can apply to the same conversation at once (send the proposal and check the credit). A task becomes a note on the contact, written by the automation — no member is its author — saying which automation created it, how sure the AI was and what the customer said, with one reminder task inside it. At its moment it does what every reminder does: a note in the conversation, a reminder in the Inbox of the member it is for (or the whole team) and the optional e-mail. Only the owner or an admin can edit the note’s words; whoever may update the contact can complete or change the task.

Accepting a task suggestion opens a small window with the due time and the assignee, ready to change.

Booking

Up to 5 options, each a calendar or a booking distribution link, with the message sent with it — which must hold {{booking_link}} (and may hold {{booking_times}}) — and, with more than one, an optional when that helps the AI pick (“Dental cleaning appointments”).

• Automatic, link — your message with the calendar’s booking page (or the link’s short address). A calendar’s page opens with the form already filled with the contact’s details, through a short signed link that carries none of them (how).
• Automatic, link and the next free times (calendars only) — the next day with free times, up to 3 of them, sent the way Suggesting a time sends them, so the bubble shows the same card with Confirm and where each time stands. Nothing is reserved.
• On FlyB Chat the visitor gets the same card: what you wrote before the placeholders, then each time with its own Confirm — the booking page opened on that time, the form already filled — then what you wrote after them, and one See all times button under the message (Choose a time when only the link was sent). Text written between the two placeholders (“You can also pick another one here:”) is left out, because the button says it. A time already taken or gone by shows as unavailable; one the visitor booked, as booked. On WhatsApp — both a Cloud API number and a WhatsApp Web number — the message goes as written, with the times and the link in its words: a WhatsApp message cannot carry several links as buttons.
• Suggested — accepting opens the team’s agenda with the option’s calendar selected; the suggestion is done once a booking is made or a time suggested from it.

Never while the contact already has an appointment coming up on one of the options’ calendars, nor when a calendar was deleted or switched off.

Deals

Up to 10 deals, each with a name, a value (in your organization’s currency, like every deal) and an optional when (“The customer chose the annual plan”). A reading picks at most one — two offers in one conversation are a person’s call — and only when the conversation is at least at evaluating something specific.

The deal is created on the contact with a note quoting what the customer said, owned by the conversation’s assignee when there is one, and shows in the conversation as Deal created · by AI. Never twice while a deal of the same kind is open for the contact — FlyB’s database refuses it — or, with Once per contact, ever, never again even after it is closed.

Resolve

Resolves the conversation when the matter is settled and nobody is waiting for anything.

Automatic resolve needs, all at once: the AI sure enough (92% by default), the matter settled, nothing else suggested in the same reading, a calm customer, no member message in the last 30 minutes (you choose how long), nobody assigned (switch Only unassigned conversations off to allow it anyway) and no snooze still ahead. A message that arrives while it is resolving keeps the conversation open. Resolving cancels the conversation’s suggestions and its follow-up, and the divider in the thread reads Resolved by the AI followed by the automation’s name.

Escalate to the team

Hands the conversation to a person when the customer asks for one — or, if you leave it on, when they are angry.

It assigns the conversation, notifies, and keeps AI Answers quiet on that conversation for the pause you chose — the automation’s history says the conversation was escalated to the team. In the conversation it shows as an internal note saying why, and who received it. Once per conversation.

CRM stage

Up to 10 stages — a funnel, a stage and an optional when — the AI may move the contact to, at most one per reading, never into the stage they are already in (and, with Only if not in the funnel yet, never within a funnel they are in). The stage’s own automations run as they would for any move.

Payment link

Up to 5 prices of your Sales module, each with the message sent with it — which must hold {{payment_link}}, filled in with that price’s checkout — and an optional when. Asked only when the customer wants to pay now (“how do I pay?”, “send me the link”). It follows the same guards as the booking link.

The link goes as a button wherever the channel can show one, labelled with the preset’s Button label (up to 20 characters; left blank, Pay now in the customer’s language):

A colon that introduced the link becomes a full stop when you wrote more after it. A message that is only {{payment_link}} goes as written everywhere.

Tags

Categories — a tag name, its colour and when it applies (support, billing, complaint…) — added to the contact automatically when the conversation is about them. Tags have no Suggest mode: a suggestion to add a tag would only be noise. A tag the contact already has is left as it is.

Signal rules

Rules that act on a reading with no question of their own: when buying intent (or dissatisfaction, or urgency) reaches a percentage, move the contact to a CRM stage, mark them as a qualified lead or add Lead Score points. Up to 10 per automation, each at most once per conversation. They act automatically, so they need the plan’s automatic actions.

In the Inbox

The suggestions bar

Above the message box (under the line that says who else is in the conversation), a bar lists the AI’s suggestions for the conversation:

✨ Suggested [📅 Suggest booking · Cleaning ▸] [💼 Deal · Annual plan ▸] [⏰ Follow-up in 1 day ▸] [✓ Resolve]

• Click a chip to do it: one click for a deal, resolving and a follow-up (with the delay shown); a small window for a task; the team’s agenda for a booking. ▸ opens Adjust before doing — the delay, the due time, the assignee, the value.
• Hover a chip for why: the question the AI was asked, how sure it was, and when the conversation was read.
• An amber dot marks an automatic action that was held back, with the reason.
• ✕ on a chip dismisses it for 24 hours in this conversation; the bar’s ✕ dismisses them all.
• At most four are shown; the rest are under ⋯. On a phone the bar scrolls sideways.
• The bar only shows what you may do (see Who can use it), and is not there at all when there is nothing to suggest.

Suggestions belong to the conversation, not to the member: everybody who opens it sees the same ones. Accepting is first come, first served — a second member gets Somebody already did this. A suggestion a later reading no longer supports goes away by itself, and any left are cleared after 24 hours or when the conversation is resolved. Accepting a suggestion is free and works whatever the plan: the reading that produced it was already paid.

The follow-up chip

A scheduled follow-up shows at the start of the bar, even with no suggestions:

⏰ Follow-up Wed 10:00 (1 day without reply) · Send now · Change · Cancel

The time is when it will really go, after any postponement, in your own time zone. Change offers the automation’s delays or a date and time of your choice; Send now runs the follow-up’s checks at once (including the AI’s last look); Cancel cancels it.

You can also schedule a follow-up by hand from the message box’s ⋯ menu — Remind them in 1 day — without waiting for a suggestion; it uses the automation’s follow-up settings, or the defaults when no automation covers the conversation, and it is kept when an automation’s follow-up is switched off.

The AI reading panel

In the contact panel, right under Conversation: Buying intent and Dissatisfaction, each as its band with the percentage on hover, an arrow against the previous reading and how long ago it was read; chips for Waiting on the customer, Declined and Asked for a person; urgency and the main objection when they are on; and a small line of the conversation’s last ten readings.

• A new conversation starts with no reading: the panel shows the previous conversation’s last one, greyed, until the first reading.
• A reading that is out of date — the customer wrote since, and the next reading was not made — says so in grey, with why when there is a reason (Not read: no Flights left).
• Analyse now reads the conversation on the spot, for 1 Flight: what it finds is suggested, never done automatically.

The header of the conversation shows 🔥 Hot or 😟 At risk when the band is reached, and so does the person’s card in the queue.

Views, order and notifications

• Hot leads and At risk, among the views: open conversations whose latest reading is Hot, or At risk.
• Sort by buying intent puts the open conversations with the highest reading first.
• An At risk notification goes to the member a conversation is assigned to when its reading becomes At risk, and Escalated to you to whoever an escalation assigns or names — see Notifications.

In the conversation

Every action leaves exactly one trace in the thread:

Suggestions that are created, dismissed or expire leave nothing in the thread.

In the automation editor

Besides the sections above, the editor of an answering automation has:

• Test — paste a conversation (one line per message, starting with contact: or team:) or pick a recent conversation of the automation: the AI reads it with the settings on screen, before you save, and the result shows the signals and what would happen — would suggest…, would do… automatically, would hold back… because…. Nothing is done or sent; each test counts towards the daily limit and is not charged.
• History — every reading of the automation and every action it suggested or did, with its outcome and, when nothing happened, why.
• Review — the automation in one sentence, the reading and the actions included.

See Message automation for the editor’s layout.

Prices and plans

Example. A conversation that paused 6 times and received 4 catalog answers costs 4 (the answers, see AI Answers) + 6 (the readings) = 10 Flights — plus 1 if a follow-up written by the AI went out.

A reading is charged only once it is done; one that cannot be paid is not made — the panel says Not read: no Flights left — and readings start again with the next Flights. A follow-up with your own message still goes when Flights run out; one written by the AI falls back to your message when you allowed it, or is not sent.

If the plan changes, readings and automatic actions stop at once — suggestions already in the bar can still be accepted, and scheduled follow-ups are skipped when their moment comes. The automations keep their settings.

Limits

Who can use it

• To set an action on an automation — or change it — you need Message Automation edit and also the right the action itself needs: CRM create for deals, stages and signal rules, Notes & Tasks create for tasks, Calendars view and Inbox create for booking, Inbox edit for resolving and escalating, Inbox create for follow-ups, Inbox create and Plans view for payment links, CRM edit for tags. Nobody can make the AI do what they could not do themselves. Switching an action off needs only Message Automation edit.
• To accept a suggestion you need the same right as the action; the bar hides the others.
• The reading and the suggestions are seen by whoever can open the conversation in the Inbox — the per-number access and own contacts only rules included.
• Automatic actions act for the organization, logged as the automation — never as a member.

Privacy

To read a conversation, FlyB sends TypeSafe — the AI provider listed in Sub-processors — your organization’s name and sector, up to the last 30 messages of the conversation (with how long ago each was written and who wrote it: the customer, your team, an automation or the AI), the customer’s messages since the last reading, and the questions the automation’s actions need: the titles and when descriptions of your tasks, the names of your deals, calendars, stages, prices and tags. E-mail addresses, numbers of eight digits or more and the contact’s own name are removed from the messages first. The contact’s record is never sent. TypeSafe returns probabilities; FlyB keeps them on the reading for 180 days.

A follow-up written by the AI sends a language model, through OpenRouter on FlyB’s own account (also listed in Sub-processors), your organization’s name and sector, your instructions and up to the last 12 messages of the conversation, with the same removals. FlyB asks OpenRouter to use only model providers that neither store nor train on what they receive.

The readings are scores about a person, produced automatically from what they wrote, and the automatic actions act on them. You decide which actions are on and whether they act alone; see Privacy Policy, section 3.1 for what that means for you and for your customers’ right to ask for a person’s review.