SquidSupport
Product reference / AI

AI Advanced settings

In the app: AI → Advanced → Advanced settings

Needs the 'View AI Config' permission.

AI Advanced settings, titled "AI Agent Configuration" on the page, is where every detail of the AI agent is set by hand. The top half, "Store settings", holds the choices for one store: whether the agent is on, whether it sends replies itself, how it behaves in live chat, when it hands over, and its limits. The bottom half, "Advanced", holds the config set the store runs and the contents of that set: prompts, specialists, tools and variables.

What belongs to a store and what is shared

Two kinds of settings sit on this page, and they reach different customers.

  • Store settings belong to the store selected in the store picker at the top. They change that store only. They are a form: nothing is saved until you click "Save AI settings" in the bar that appears at the bottom, which names what is unsaved. "Discard" puts the form back.
  • Config set contents, under "Advanced" → "Config Set Resources", belong to a config set and are shared by every store that uses it, marked "Shared". Most of these save as soon as you change them.

One part does not follow this split. The discount caps and the order cancellation rules under "AI action limits" are shown with the store settings and are saved with "Save AI settings", but they are stored on the config set chosen in "Config Set Resources". They apply to every store that uses that set.

A "Setup issues" row under the title lists what stands in the way of the agent for this store: warnings about the store's AI settings such as unmapped departments, tools that are off because an integration is not connected, a missing ParcelPanel connection, and Refund Hub being switched off. Most come with a link to fix them.

Seeing the page needs the 'View AI Config' permission. Changing store settings, prompts, specialists and variables needs 'Edit AI Config'. Switching tools on or off, saving the discount caps and order cancellation rules, and making or changing custom tools need 'Edit AI Tools'; seeing custom tools needs 'View AI Tools'. Creating and activating releases and switching version control needs 'Release AI Config'.

Some changes can be refused with "Review failed policy checks before going live". This happens when the AI Builder made the configuration, one of its policy checks failed for exactly this version and this store, and no administrator has reviewed it in the Builder. It then blocks switching the AI on, turning on auto-send or "AI-only live chat", choosing or changing the config set of a store, copying AI settings from another store, activating a prompt and switching on a specialist. A configuration that was not made in the Builder has no such checks.

Store settings

Needs the 'Edit AI Config' permission.

The first block is titled "AI agent for" and the store's name.

"AI Agent Status" is the on and off switch of the store's AI agent. Off by default. When it is on, the AI processes the store's customer messages. When it is off, messages wait for your team. The switch cannot be turned on until three things exist, which a checklist under it names with links: a config set assigned to the store, an active main prompt in that set, and an agent name. Turning the AI on, or changing its config set while it is on, is also refused while a specialist the main prompt hands off to is switched off or has no active prompt.

"Auto-send AI Responses" decides what happens to a reply the AI writes.

  • Off: replies are saved as drafts in the ticket and a teammate sends them. This is the default.
  • On: replies go to the customer directly, without review.

Turning it on asks "Turn on auto-send?" and lists what the AI will then do without review: discounts, automatic refunds on verified returns, order cancellation and every other tool that changes something. Auto-send can only be turned on after the store's configuration passed a test under "Test & go live"; until then the switch is locked and says so. The test must be of the configuration as it is now and not older than 30 days: after a change to the config set, test again. Turning auto-send off needs no confirmation and also turns "AI-only live chat" off.

Turning auto-send off has a side effect on tools. Every tool that makes changes is switched off in the store's own tool settings, and also on the config sets the store uses when no other store uses that set. A set shared with another store keeps its switches, but the AI does not get those tools for a store that saves drafts. The switches are not turned back on when auto-send is turned on again: switch the tools on again on the "Tools" tab.

The AI agent name is the name customers see and replies are signed with, up to 100 characters. It is required to switch the AI on.

"Max AI Turns" limits how many replies the AI gives in one conversation. Draft replies count too. Once the AI has reached the limit, the next customer message triggers a handover before the AI answers it. In AI-only live chat, the choice under "When the AI would escalate" applies instead. From 1 to 50; the default is 5. With the switch off there is no limit.

"AI Disclosure" controls whether the AI may say it is an AI:

  • "Disclose if asked": it answers truthfully when a customer asks directly. This is the default.
  • "Always disclose": it introduces itself as an AI assistant in greetings.
  • "Never disclose": it presents itself as a human agent. Some regions require disclosure on request; you are responsible for meeting those rules.

"Reply in the customer's language when it differs from the store language" is on by default. When on, the AI uses the customer language already recorded for the ticket, or tries to detect it from the message. A supported language takes priority over the store language. An unsupported language falls back to the store language. With no usable reading, the store language is the fallback, although the AI's general instruction to match the customer language remains. When off, the AI is instructed to use the store language. An explicit language preference supplied for a reply takes priority over this switch.

"Enable vision (process images in messages)" is on by default. When on, the AI attempts to process customer images, normally up to five per run. Only successfully processed images reach the AI; turning this on does not guarantee that every image can be read. When off, this image-processing step is skipped.

"Response Delay (seconds)" makes the AI wait a random time between "Min delay" and "Max delay" before an email reply, each from 0 to 3,600 seconds. 0 means instant, which is the default. A minimum that is higher than the maximum cannot be saved. It applies to email and every other channel except the chat widget and WhatsApp, which use the pause under "Live chat" instead and never this one.

"Test an answer" opens a test conversation with the agent this store runs. Nothing is sent to customers; see the Playground for how tests work.

"Copy from" at the top of the block opens "Copy AI settings". It replaces all of this store's own AI settings with those of another store. That includes whether the AI is on, whether it replies without review, the agent name, the config set and the sets per channel, and the live chat, return, supplier, escalation and department settings. The two stores run the same config set afterwards. Store URLs are not copied. If the source store is live and auto-sending, this store starts doing both at once, without the normal activation-readiness and configuration-test checks. Copying still checks failed AI Builder policy reviews for the destination. It is refused if a copied config set is restricted to the source store. If this store was auto-sending and the source store is not, the tools that make changes are switched off as described above. It cannot be undone except by changing the settings back.

Live chat

Needs the 'Edit AI Config' permission.

These settings apply to the chat widget. The reply pause, the typing and writing-style switches as described below, the handoff message, the waiting acknowledgment and the email alert are also used for WhatsApp. "AI-only live chat" is for the chat widget only.

"Humanize live chat" makes chat replies feel written by a person. Its four switches start off. Typing speed starts at 10 characters per second, and the maximum messages per reply starts at 3. These choices belong to the selected store and take effect after "Save AI settings":

  • "Pause before replying": waits a random time between "Min delay" and "Max delay", each 0 to 300 seconds, before a chat or WhatsApp reply. Switching it on proposes 3 to 12 seconds when both saved delay fields are zero; otherwise it keeps the existing range. The minimum cannot exceed the maximum. With it off, chat and WhatsApp replies are not delayed at all, whatever "Response Delay (seconds)" says.
  • "Simulate typing": in the chat widget, shows the typing indicator before each message, for a time based on the length of the message and "Typing speed (characters per second)", from 1 to 100. The indicator normally shows for about 1.2 to 10 seconds per message, with a small random variation. If the planned typing and pauses exceed 45 seconds in total, they are shortened together, so an indicator can last less than 1.2 seconds. With typing off, no indicator is added by this setting. On WhatsApp the typing indicator is shown when the AI starts on a reply.
  • "Split replies into multiple messages": in the chat widget, the AI writes its answer as a few short messages and they are sent one after the other, up to "Max messages per reply", from 2 to 5. With this switch off, this delivery step keeps the reply in one message. It does not split WhatsApp replies into chat bubbles.
  • "Casual writing style": contractions, short sentences and a conversational tone, in the chat widget and on WhatsApp. Spelling stays correct. It does not change email replies.

"Escalation experience (live chat)" sets what a visitor sees when the AI hands a chat to your team:

  • "Handoff message": sent the moment the AI hands the chat over. It is a fixed message, not an AI reply, so it is also sent for stores in draft mode. It is skipped when the handover comes from an AI tool in a store that sends replies itself, because the AI's reply supplies the transition. It is also skipped if the chat has closed, a person has already replied since the handover, or a handoff message has already been sent for that handover. A duplicate-prevention check can also suppress another handoff message within five minutes. Leave empty to use the default. Up to 2,000 characters.
  • "Waiting acknowledgment": sent at most once per handover when the visitor writes again while waiting for a person, and not after a person has replied. Up to 2,000 characters. Left empty, the handoff message is used, or the default when that is empty too.
  • "Email alert after (minutes)": when the chat is still unanswered this many minutes after a handover, an email goes to your active admins and managers and to the teammate the chat was assigned to, at most 10 people. 0 to 1,440; the default is 5 minutes, and 0 switches the email off. Closed chats and chats no longer handed over to your team get no alert. The recipient limit can leave the assigned teammate out when ten other recipients have already been selected.

"AI-only live chat" lets the AI handle every chat in the chat widget for the store. It needs the AI on and auto-send on; it does not apply to WhatsApp. It starts off. With it on, "When the AI would escalate" chooses what happens instead. "Boundary message" and "Ticket confirmation message" each allow up to 2,000 characters; leaving them empty uses the built-in wording:

  • "Escalate to the team live": hands the chat to your team as usual. This is the default.
  • "Don't escalate": the AI sends the "Boundary message", at most once every 6 hours per chat, and keeps handling the chat. The turn limit then no longer hands chats over.
  • "Escalate via follow-up email ticket": the AI sends the "Ticket confirmation message", an email ticket linked to the chat is created for your team, and the chat closes. It needs a verified sending domain. Without one no ticket is made: the visitor gets a message asking them to email the store, the AI keeps the chat, your team gets a notification in Squid, and your admins are alerted, at most once a day, that the sending domain is not verified. A visitor who left no email address is handed over live.

Two normal overrides still send chats to your team. If the escalation settings cannot be resolved, Squid also falls back to a live handover. A chat that was waiting because the AI could not answer during an outage is handed over live. And a conversation about an order with an open payment dispute goes to your team before the AI runs, when "Active dispute handling" under Settings → Integrations is set to send such tickets to people.

Escalation rules

Needs the 'Edit AI Config' permission.

Escalation rules decide when the AI hands a conversation to a person.

"Keywords" are hard triggers. Before the AI answers, Squid searches everything the AI is given about the ticket for each keyword. A match triggers a handover before the AI writes a normal reply. The email acknowledgment or chat handoff message can still be sent. In AI-only live chat, the choice under "When the AI would escalate" applies instead.

What is searched is wider than the newest customer message: it includes earlier messages and replies from your side, together with the customer and order details the AI gets with it. It is not an unlimited search of the ticket: the normal context keeps the latest 50 messages and is shortened if it exceeds its size limit. A keyword therefore does not have to be in the latest message to trigger.

Keywords are matched as whole words and without regard to capitals: "member" does not match "remember". A keyword of several words must appear exactly as written. Enter one per line or separate them with commas; the suggested keywords under the box add themselves with a click.

"Escalation Instructions" start empty and are added to the AI's instructions. In AI-only live chat with "Don't escalate", they are replaced by the instruction to keep handling the chat. They describe, in your words, when the AI should decide to hand over after reading the conversation.

"Block uncategorized transfers to human team" is off by default. When on, the AI may only hand over with a specific handoff reason. A missing reason, a reason that is not an active one of your list, or a catch-all such as "other", is refused and the AI keeps handling the ticket. It applies to handovers the AI decides on itself, not to keywords, the turn limit or the categorizer. It needs handoff reasons to be set up under AI → Handoff reasons: with none, every such handover is blocked. If Squid cannot read the handoff reasons or the store's switch, this restriction is skipped so that a lookup failure does not prevent a handover.

"Acknowledge escalations" is on by default. When on, the customer of an email conversation gets a short note when the AI hands it to your team without having written a reply. When the AI did write a reply with the handover, that reply is used instead, also with this switch off. "Acknowledgment message" holds the text, up to 2,000 characters; leave it empty for the default wording, which is written in the customer's language for the common languages. The customer's name and the store's name can be filled in automatically. In draft mode the note waits for review like any AI reply. Your own acknowledgment wording is kept as written, rather than translated automatically. An acknowledgment is skipped if a person has already replied or the conversation has closed before it is delivered. Live chat and WhatsApp use the chat handoff message instead.

Reassess Escalations

Needs the 'Edit AI Config' permission.

This section is shown only when Squid has switched reassessment on for your organization.

"Enable reassessment for this store" lets the AI take a ticket back after it handed it over. Squid checks this when the customer writes again on a handed-over ticket, and only after a teammate has sent the customer a reply since the handover. An internal note does not count. A customer who asks for a person in so many words stays with your team.

When on, "Reasons eligible to return to AI" lists each handoff reason with its own switch. For a new store they start switched off, as does reassessment itself. Older stores with reassessment already enabled can start with eligible reasons switched on. With explicit reason rules saved, the latest handover must match an enabled rule. A handover without a reason can match an enabled uncategorized rule; older settings without explicit rules can also allow it. A reason marked "Always human" cannot be switched on.

For a reason that is switched on, two texts say how, each up to 12,000 characters:

  • "When should AI take over?": the conditions that must be met before the AI resumes, such as the confirmation a person must have given. Squid checks them against the conversation and the current order details. When the check does not confirm them, or cannot run, the ticket stays with your team. Left empty, there is no analyser check; the other routing conditions still apply. If the ticket, latest message, handover or saved rules change during an analyser check, its result is discarded.
  • "Special instructions for AI when it takes over": how the AI should continue. They are given to the AI when it resumes.

These settings belong to the selected store and take effect after "Save AI settings". "Open full screen" opens a larger editor for a reason; changes there still need "Save AI settings". When the AI takes a ticket back, the ticket gets the note "AI reassessed this customer reply and resumed handling."

Returns

Needs the 'Edit AI Config' permission.

Return windows are not set here: they are on Settings → Return Policy and apply to the return portal and the AI alike.

"When a customer can't use the return portal" chooses what the AI does when a customer says they cannot finish a return in the portal:

  • "Hand off to a human agent". This is the default.
  • "Let the AI process the return over email": the AI handles that return itself in the conversation, with its return instructions and offers, and stops sending the customer back to the portal.

Either choice is only used when Squid can confirm a reason to leave the portal route, such as a real problem with the portal. Without one the AI sends the portal link again.

"When a return is verified" chooses the action once the carrier has scanned the return parcel and it is addressed to your return warehouse. This confirms the parcel is on its way, not that it arrived.

  • "Do nothing": the AI is told the tracking is valid and continues with its own instructions. This is the default. The setting itself creates nothing and moves no money, with one exception: when the return is proven only by a photo of the shipping receipt, a pending refund request is submitted for review in Refund Hub.
  • "Hand off to a human agent": the ticket goes to your team, and the AI does not promise a refund.
  • "Submit for review in Refund Hub": creates a pending refund request for what the customer paid for the returned items, less the return shipping fee when "Subtract shipping fee when customer returns items to warehouse" is on. A person approves it before money moves.
  • "Refund automatically": refunds the customer in Shopify as soon as the tracking verifies, with nobody reviewing it, also when refunds otherwise need approval. Choosing it asks "Refund verified returns automatically?". Some refunds still go to review in Refund Hub instead: the carrier reports a delivery city that does not match your return warehouse; the customer sent two orders back in one parcel; the return is proven only by a photo of the shipping receipt; the refund reaches the amount under "Send large refunds for approval" in Settings → Refund Hub; the order is waiting for a return-policy review; or the order already had one automatic refund.

For the last two, Squid requests the remaining paid value of the returned items and checks it against the "Return automation refund" category in Refund Hub. That category starts with a 100% ceiling and can be changed there. Lowering the ceiling does not automatically resize this refund: if the requested amount after any shipping fee exceeds the ceiling, preparation is refused and the ticket goes to your team. Both need the AI's refund tool to be switched on for the store's config set; when it is off, the ticket goes to your team instead. Already refunded or cancelled quantities and add-ons are excluded. Items already settled by an accepted keep-item offer are excluded too; if those are all the returned items, the ticket goes to a person. Missing or stale order details stop preparation until corrected. Existing refund requests are not duplicated. A refund that cannot be prepared goes to your team.

While the store saves drafts instead of sending replies itself, none of these actions is carried out: the AI only reports that the tracking is valid.

High-value orders are handled per return policy, under Settings → Return Policy.

AI action limits

Needs the 'Edit AI Tools' permission.

Both cards in this section are saved with "Save AI settings", but not on the store. They are stored on the config set selected in "Config Set Resources", which starts on the set the store runs. They apply to every store that uses that set. If you picked another set in "Config Set Resources", the limits are read from and saved to that set. When version control is on for the set, a change to these limits goes into the draft and reaches customers with the next release.

Discounts

"Discounts" holds the caps for discount codes the AI creates:

  • "Max Discount Percentage": 1 to 100, shown as 20. It limits codes the AI sizes as a percentage.
  • "Max Discount Value": 1 to 1,000, shown as 25. It is an amount in US dollars and is converted to the currency of the order. When no exchange rate is available for that currency, this cap is not applied. It limits codes worth a fixed amount, including a voucher from an accepted offer of the "Advanced offer system".
  • "Max Discounts Per Customer": 1 to 20, 3 by default. It counts the codes the AI already created for the same customer email address in this store.

When a code would go over a cap, or the customer already has the maximum number, no code is created and the ticket is handed to your team. The AI does not lower the amount to make it fit.

Switching "Discounts" on asks "Let the AI offer discounts?". The switch does two things only: it shows the caps, and it lets "Save AI settings" store them. It does not decide whether the AI can create discount codes. That is decided by the discount tool on the "Tools" tab of the config set: the AI creates codes when that tool is on and none when it is off.

Caps that were never saved are not applied. Until the caps are saved once with the switch on, the set has no percentage cap and no value cap, and the limit per customer is 3.

Order Cancellation

"Order Cancellation" sets what the AI does with a request to cancel an order. Two tools on the "Tools" tab use it: "Order Placed Time", which only checks the window, and "Orchestrate Order Cancellation", which runs the whole flow. The card shows its fields when at least one of them is on. With none on, the card says cancellation is off and none of the settings below are used. "Orchestrate Order Cancellation" makes changes, so the AI does not have it while the store prepares drafts instead of sending replies itself; "Order Placed Time" stays available.

"Cancellation Window" is how long after an order was placed a request still counts as in time. Choose minutes, hours or days; the minimum is 1 and the default is 5 minutes. It uses whole minutes from the order to the customer request. If the request time is unavailable, it uses the time the tool runs. There is no maximum in this field.

"Escalate To" is the team that gets a request made inside the window: "Human Team", which is the default, "Fulfillment Team", "Dispute Team" or "No escalation". The ticket is handed to that team with a note that it must be processed urgently. With "Orchestrate Order Cancellation" on, the supplier is also asked to stop the order when nothing of it has shipped and the store has a supplier connection. With "No escalation", the normal complete-order request inside the window is not handed to a team: "Orchestrate Order Cancellation" sends an order that has not shipped to the supplier as a cancellation request instead, and "Order Placed Time" on its own only tells the AI that the order is inside the window.

A partial cancellation inside the window still goes to a person if the requested items are missing or cannot be matched. A config set can also carry a separate in-window action configured elsewhere that cancels and refunds a whole unshipped order; this card preserves that choice. A partial request or an uncertain main order goes to a person instead.

"Outside the cancellation window" holds the rules for a request that comes too late. They are shown and used only when the tool "Orchestrate Order Cancellation" is on. Two sets exist, switched with "Normal" and "Dispute"; both are saved together. "Dispute" applies when the AI judges that the customer threatens a chargeback or payment dispute.

  • "When the order is unfulfilled outside of the cancellation window": "Refuse cancellation", "Assign to human team" or "Create supplier ticket". For the last two, "Only if the order has been unfulfilled for at least … days" sets a minimum age, counted in whole days since the order was placed; a younger order is refused. 0 means any age. An order that has partly shipped follows this rule for the part that has not shipped.
  • "When the order is fulfilled (not shipped)": the same three choices, for an order marked fulfilled whose parcel the carrier does not have yet. Anything other than "Refuse cancellation" first checks with the carrier data. Only items confirmed not to be in transit are stopped or handed over. When that cannot be confirmed, the request is refused.

Defaults when nothing was saved: under "Normal", an unfulfilled order gets "Create supplier ticket" after 14 days, and a fulfilled order gets "Refuse cancellation". Under "Dispute", an unfulfilled order follows the "Normal" choice at any age, and a fulfilled order gets "Create supplier ticket" when the unfulfilled choice is a supplier ticket, otherwise "Refuse cancellation".

"Create supplier ticket" needs a supplier connection: Basecamp, or a supplier that receives its tickets in the Supplier Portal. Without one, no ticket is made at any order age. For an unfulfilled order the AI is then told to hand the request to a person; Squid does not force that handover, the AI does it by its instructions. For a fulfilled order the items that can still be stopped are handed to your team. A handover from these rules goes to the team chosen under "Escalate To", or to "Human Team" when that is "No escalation".

None of the choices offered on this card directly cancels the order in Shopify or refunds it. A supplier ticket asks the supplier to stop the order, and a handover leaves the cancellation to your team. An order that is in transit is never cancelled, whatever is chosen here. The AI is told the order has shipped, can no longer be cancelled and can be returned once the customer has it. A delivered order is treated as a return request, not as a cancellation.

Department mapping

Needs the 'Edit AI Config' permission.

AI tools hand conversations to a role, not to a department. This section says which of your departments each role is: "Human Team", "Fulfillment Team" and "Dispute Team". Choose a department per role, or "Not mapped". The mapping starts empty, belongs to the selected store and takes effect after "Save AI settings". Departments are created under Settings → Departments; with none, the section asks you to create them first.

A handover to a role that is "Not mapped" still happens. Squid then looks for the department in this order:

  • for "Human Team" and "Fulfillment Team": a department named exactly "Human" or "Fulfillment"; capitals do not matter;
  • for "Dispute Team": a department with "dispute" in its name. When there are several, "Disputes Team" is preferred, then "Dispute Team", "Disputes" and "Dispute", and otherwise the first one in alphabetical order;
  • the store's fallback department, when one was set for the store. This page has no field for it.

When none of these exists, the ticket's department is not changed, so a ticket that had none stays without one. It is taken off the AI agent, reopened if it was snoozed, and waits in the inbox until someone picks it up. This fallback does not automatically assign a teammate without a target department. A teammate already assigned to the ticket is kept.

When a department is found, the ticket moves to it and Squid tries to assign an available teammate from that department. If nobody is available, the ticket waits in that department's queue.

Store variables

"Store variables" are values filled in from the store's own settings that every prompt can use.

"Store URLs" holds seven addresses: "Store Website URL", "Order Tracking URL", "Return Policy URL", "Return Form URL", "Privacy Policy URL", "Terms of Service URL" and "Contact Page URL". Leave "Return Form URL" empty to use the address of the store's return portal; an address you fill in is used exactly as written. They start empty and are saved for the selected store with "Save AI settings". Saved values are used when the next prompt is prepared. Other than the return portal fallback, an empty address supplies no value.

The table below lists the variables the system provides, with "Variable name", "What it says" and "Source". These cannot be edited here; change the setting they come from.

Prompt configuration

"Prompt Templates" lists the ready-made prompts Squid provides, grouped by category, each with its name, description and kind. With none available the section says "No templates available yet". Seeing the list needs the 'View Templates' permission; "Import" needs 'Edit Templates'.

"Import" makes a copy of the template as a new draft prompt. The template itself stays as it is. The copy is stored on the selected store, not in a config set: it shows in the "Prompts" tab right after the import, but it is not part of the config set, it is gone from that tab when the page is loaded again, and it cannot be activated there. No agent answers from it. To use a template's text for a store that runs a config set, create a prompt on the "Prompts" tab and paste the text in.

"Import" is refused when the store already has a prompt with the same short name as the template.

Advanced

"Advanced" is folded by default; a link that points at something inside it opens it. It holds the config set the store runs, a picture of the AI's route, a tool check and the contents of the set.

"Deployed Config Set" chooses which config set answers for this store. It applies as soon as you choose, without "Save AI settings". You are asked to confirm first in two cases: the first time a set is chosen for the store or for a channel, and when you change the set while the AI is on for the store. It needs the 'Edit AI Config' permission.

  • "Default" is the set for every channel that has no set of its own.
  • "Chat widget & WhatsApp" and "Email" can each use a different set. The chat choice also applies to WhatsApp. "Use default set" removes the channel's own set.
  • A store with no set chosen under "Default" runs the config set that is marked "(Default)" for your organization.

While the AI is on for the store, a change is refused when neither the chosen set nor the store has an active main prompt, or when its main prompt hands off to a specialist that is switched off or has no active prompt. A set made for one other store cannot be chosen. After a choice, a message names the tools of that set that are off for this store because an integration is not connected. A note under "Default" says when the set is also used by other stores: editing it affects all of them.

"Describe instead: open the AI Builder" in the card opens the AI Builder. "Prefer describing instead of configuring?" appears in its place when there are no config sets yet.

How your AI agent works

"How your AI agent works" is a fixed picture of the route a message takes: "Customer Message", "Categorizer", "Orchestrator" (the main agent), "Specialists", "Tools" and "Response". It starts folded to one line, which also gives the number of specialists that are switched on. It is an illustration and changes nothing.

Only four things on it come from your setup: the number of specialists that are switched on, the names of up to three of them (on a wide screen, after clicking "Specialists"), the number of tools that are switched on, and "Draft" or "Auto-send" on "Response", taken from the store's saved setting. The specialists and the tool count are those of the config set selected in "Config Set Resources", which can be a different set from the one the store runs. Every other text on the picture is the same for everyone, including the example actions under "Tools", which are not a list of your tools.

Clicking a box unfolds a few lines and a link: "Configure Prompts", "Manage Specialists" and "Configure Tools" open that tab of "Config Set Resources". The picture does not show which tools the AI can really use for this store; "Validate config tools" does.

Validate config tools

"Validate config tools" is a read-only check of whether the AI of the selected store can use the tools its config set switches on. It changes nothing. A line says how many checks pass; "Show checks" lists them. The list opens by itself while a check does not pass, also while it is still loading.

A tool reaches the AI only when three things are true: it is switched on for the config set on the "Tools" tab, the active main prompt allows it under "Allowed Tools", and nothing holds it back for this store. The two common things that hold a tool back are an integration the store has not connected, and draft mode: while "Auto-send AI Responses" is off, the AI gets no tool that makes changes and cannot resolve, close or snooze a conversation or schedule a follow-up itself.

The rows do not all look at the same config set. "Enabled function tools", "Return-address tool access" and "ParcelPanel tool access" count the switches of the set selected in "Config Set Resources". "Active orchestrator prompt", "Shopify tool access" and "Side-effecting tool mode" are worked out for the set the store runs by default.

  • "Config set": the set selected in "Config Set Resources" and where it stands for this store: the set the store runs, the set it runs on a named channel only, or "not deployed to this store, so the tool rows below are a preview".
  • "Deployed set": an extra row for a store that has no set chosen and therefore runs your organization's default set. It says that a change to that set reaches every other store without a set of its own.
  • "Active orchestrator prompt": whether a main prompt is active in the set the store runs. A legacy active main prompt on the store can serve as a fallback; without either, the agent cannot answer. Activating a prompt in a set the store does not run does not count.
  • "AI agent name": whether the store has the name replies are signed with.
  • "Enabled function tools": how many tools are switched on for the selected set. This is a count of switches, not of what the AI gets.
  • "Shopify tool access": how many of the tools that need Shopify the main agent gets in any conversation, and which of the conditions above stops the others. Tools the AI only gets in some conversations are named separately and not counted.
  • "Return-address tool access" and "ParcelPanel tool access": whether the return address (a fixed one, or one on a supplier) or the ParcelPanel connection that switched-on tools need is set up for the store.
  • "Side-effecting tool mode": whether the store sends replies itself or saves drafts, and which tools draft mode holds back. Held-back tools are a warning, not an error: it is how draft mode works.

The three rows that are worked out for the store turn grey, with a question mark, when their answer could not be loaded. Read that as not checked, not as a pass.

What the check does not cover: it describes the main agent on the store's default config set. A specialist applies its own "Allowed Tools" on top and can have fewer. A different set chosen for "Chat widget & WhatsApp" or "Email" is not checked. Tools can also differ per conversation: while the AI sends a customer to the return portal it cannot give a return address, ask for photo proof or take back a voucher, and in "AI-only live chat" with "Don't escalate" it cannot hand a chat to your team.

Config Set Resources

This card holds the contents of one config set. The picker at the top chooses the set; it starts on the set the store runs. Everything here is shared by every store that uses the set.

Above the tabs, the "Main AI Agent" card holds settings of the main agent. They save as soon as you change them, without "Save AI settings", and apply to every store on the set.

  • "Response verbosity": how detailed the agent's replies are, in four levels. The default leaves the style to the prompt. It is never chosen automatically. It applies from the next message, also when version control is on.
  • "When a customer argues with the return policy": applies while the AI is sending return requests to the portal. Choose between leaving it to your prompt, keeping the AI on it, handing over only when the customer threatens a dispute, and handing every such dispute to a person. When version control is on for the set, a change here goes into the draft and reaches customers with the next release.

Model intelligence and intelligence routing

"Model intelligence" is how capable a model the main agent uses: low, medium, high or extra high. A higher level costs more per reply. Specialists have their own level on the "Specialists" tab; a specialist without one uses the main agent's model. The level and the routing mode apply from the next message, also when version control is on.

Who sets the level depends on the organization. By default it is set automatically: the card shows the level with the note "Chosen for you", there is no slider and no routing switch, and a main agent without a stored level normally runs on medium. If Squid enables situation-based routing for the organization, a separate classifier can raise the level from the ticket situation; it never lowers the set's level or the conversation's current level. In its comparison mode the chosen level is recorded without changing the reply. This platform setting is off by default. The organization cannot change the level, and a config set that is imported arrives without its levels. Squid staff can give an organization the manual controls described below.

With the manual controls, "Intelligence routing" has three modes. "Manual" is the default. Squid can disable this routing at platform level; then the configured level is used in all three modes.

  • "Manual": the main agent always uses the level set with the "Model intelligence" slider. With the slider on its default, the platform's standard model is used.
  • "Shadow": customers still get answers at the level you set. In the background, Squid works out which level "Automatic" would have chosen and, when that gives a different model, answers the same ticket again at that level and stores both answers side by side. The second answer is never sent, and actions in it that would change something are simulated. One comparison is stored per ticket, up to 500 tickets; after that nothing more is recorded until you leave "Shadow" and choose it again. This page does not show the comparisons.
  • "Automatic": Squid chooses the level for each customer message, and the slider is switched off. Your own level is no longer used for the main agent.

How "Automatic" chooses, using what the categorizer found in the message:

  • a simple message starts at low, a standard one at medium, a complex one at high. When the categorizer gave no verdict, the level is medium, never low;
  • it is raised to at least medium for a possible refund exception, a needed identity check, a financial offer, or an unclear first message;
  • it is raised to at least high when the message has an image the AI will look at;
  • it is raised to extra high when the customer threatens a payment dispute or legal action.

The level only goes up inside a conversation: once a ticket has run at a higher level, later messages stay at least there, until the inactivity window has passed since the AI last chose a level for that conversation. That window defaults to 24 hours and can be changed by Squid. Follow-ups the AI sends by itself reuse the conversation's level. "Automatic" changes the model only; verbosity stays as set. If the choice fails for any reason, the reply is made at the level you set.

When version control is on for the set, a banner says changes go to the draft and reach customers with the next release.

The tabs are "Prompts", "Categorizers", "Specialists", "Tools", "Variables", "Return shipping fees", "Tone of voice" and "Version control".

AI Prompts

Prompts are the instructions the AI follows. The "Prompts" tab lists the prompts of the config set, grouped by type, one row per version, with a badge: "Active in this set", "Draft" or "Archived". A line says when a version is not live and which version customers get instead.

"Create Prompt" and "Edit" open the prompt form:

  • "Name", 1 to 100 characters, and "Slug", 1 to 100 lower-case letters, digits, underscores or hyphens. The slug is the prompt's fixed short name; it cannot be changed after creation.
  • "Type": "Orchestrator" for the main prompt, "Specialist", or "Return Deflection" for what the AI says when it sends a customer to the return portal. The type is fixed once the prompt exists.
  • "Description": a note for your team, up to 2,000 characters.
  • "Content": the prompt text. Variables are written between double curly brackets and tools are referred to with an @ sign. "Insert an outline" adds a starting structure to a new, still empty prompt, "See this config set's variables" lists what you can use, and "Open full-screen editor" opens a larger editor in a new tab. The main prompt can hold 250,000 characters, a prompt of another type 110,000. A notice lists behaviours the platform manages itself, which instructions cannot change.
  • "Allowed Tools", for an "Orchestrator" prompt: the tools the main agent may use. "All enabled tools (default)" allows every tool that is on for the set. Ticking a tool here cannot switch it on; tools that are off on the "Tools" tab stay unavailable.
  • "Allowed Specialists", for an "Orchestrator" prompt: the specialists the main prompt may hand off to, or "All specialists (default)".
  • "Change note", when editing: optional, up to 500 characters, shown in the history.

Saving an existing prompt saves a new version as a draft: customers keep the live version until you activate the new one. On a row, "Activate" makes a version live from the next message, or puts it in the draft when version control is on. For a version older than the live one the button reads "Roll back to" and asks "Roll back to an earlier version?". "View history" shows every version with its change note and "Past versions" lists archived ones.

"Delete" removes one version, not the whole prompt, and cannot be undone. The active version cannot be deleted. Deleting a version of a specialist prompt also deletes the specialist that runs on that version.

Categorizer

The categorizer classifies each incoming customer message before the main agent runs. It reads the customer's newest messages together, gives them a category and decides whether the ticket needs a person. When it decides that, the normal path hands the ticket to your team before the main agent answers. "Disable human team assignment" and the store's AI-only live chat handling can change that outcome. If classification is throttled, the reply is deferred for retry; it is not treated as a successful classification.

"Disable human team assignment": when on, the categorizer can no longer send a ticket to your team. Its verdict is still recorded, and the main agent answers instead. Handovers from tools, keywords and the turn limit are not affected. Off by default. The switch belongs to the config set, so it applies to every store on the set, and it saves as soon as you change it. When version control is on for the set, a change goes into the draft and reaches customers with the next release.

"Create Categorizer" or "Edit Categorizer" opens its prompt. "Use default categorizer prompt" saves the platform's standard categories and rules as the set's categorizer prompt; it is a copy of the standard text as it is at that moment. Turning it off lets you write your own, starting from the default. "Save and Activate" saves a new version and makes it the active one for the config set at once; with version control on, it reaches customers with the next release. The card under the switch shows the prompt in use, marked "Default prompt" or "Custom prompt", with "History" for its versions and "Activate" on a version that is not active. A set without a categorizer prompt uses the built-in one, shown as "Using the runtime fallback".

Specialist Agents

A specialist is an expert sub-agent the main agent can call for one topic, such as returns or shipping. The "Specialists" tab lists them with a switch each. A specialist that is switched on but whose prompt is not active is flagged, and the AI cannot be enabled while the main prompt hands off to it.

"Create Specialist" and "Edit" open the specialist form:

  • "Name", up to 100 characters.
  • "When to route here": a short description that tells the main agent when to use this specialist. From 10 to 5,000 characters.
  • "Specialist Instructions (system prompt)": what the specialist follows once it is called. The form requires 10 to 110,000 characters. Saving a change creates and activates a new version, which goes into the draft when version control is on; the history is on the "Prompts" tab.
  • "Trigger Keywords", separated by commas. They are not searched for in the customer's message and do not send a message to the specialist by themselves. They are compared with the category the categorizer gave the message, and only matter when the config set has more than six specialists that are switched on: the main agent's instructions then describe six of them, first the specialists without keywords, then those with a keyword equal to the category or to one of its words, then the rest. The main agent can still call a specialist that is not described.
  • "Description", a note for your team, up to 5,000 characters. The AI does not see it, with one exception: for a specialist that has no "When to route here" text, the description is what the main agent is told about it.
  • "Enable this specialist": on for a new specialist; off means the main agent does not use it.
  • "Allowed Tools": the tools this specialist may use, out of the tools that are on for the config set. Leaving it empty does not grant every tool: the specialist then has no tools at all. It normally answers from its instructions only. On a supplier-portal workflow, the platform can supply the tools for recording supplier actions, relaying replies and forwarding proof even when they were not selected. Tools that make changes are also held back while the store saves drafts.
  • "Model intelligence" and "Response verbosity" for this specialist: low, medium, high or extra high, or the default. Default intelligence inherits the main agent's model; default verbosity adds no style instruction. Manual intelligence controls appear only for organizations allowed to choose levels. Automatic routing can raise the specialist's level to the level selected for the conversation.
  • "Only for filed returns", off by default: hides the specialist while the AI is sending the customer's return request to the return portal. It becomes available once the conversation is handled as a return, for example when a return is on file for the order.
  • "Hand off the conversation", off by default: when on, the specialist takes over and writes the reply to the customer itself. When off, the main agent consults it and writes the reply.

Specialist settings belong to the config set. Saving them affects every store on that set on subsequent runs; with version control on, release-controlled changes wait for a release.

Advanced offer system

The form of an "Orchestrator" prompt and the form of each specialist can carry an "Advanced offer system". It lets you fix in advance which compensation the AI may offer, in which order, and what is carried out when the customer accepts. The AI does not choose amounts: it asks Squid for the next offer and passes on what it gets. The card is shown when the offer tool exists for the config set.

The switch "Advanced offer system", off by default, turns the system on for that one agent, the main agent or one specialist. Off, or on without any group, that agent gets no offers from it. The switch and the groups are saved with the prompt or the specialist, and are shared by every store on the config set. With version control on they go into the draft.

Offer groups

An offer group is one ladder of offers for one situation. "+ Add offer group" makes a new one. "Reuse existing group" attaches a group another agent in the same config set already uses; a reused group stays shared, so an edit applies everywhere it is used.

  • "Reason code": the short name the AI uses to ask for this group. Lower-case letters, digits and underscores, up to 75 characters.
  • "When should AI invoke this group?": the situation the group is for, in your words. Up to 5,000 characters. The initial group list gives the AI this sentence and the reason code. Offers already revealed can also appear in its order context.
  • "Require the return to have been registered within its return period": off by default; when ticked, the group gives no offer for an item whose return was registered after the return period ended.
  • "How AI sees this offer ladder" shows the exact text the AI receives: per group the reason code and the situation, never the offers. It is placed where the prompt text contains the placeholder named in that panel. A prompt without the placeholder does not tell the AI about the groups.

Offers, options and what acceptance does

"Offer ladder" lists the offers of a group in order; "+ Add offer" adds a step. An offer can hold several options ("+ Add option") that are presented together; the customer takes one.

  • The offer text is what the AI presents to the customer, in the customer's language. It is required and limited to 5,000 characters: the form cannot be saved while an option has no text.
  • "Agent note (optional)" is guidance for the AI, up to 5,000 characters. The AI is instructed to keep it out of the customer reply.
  • "Identification tags" are added to the ticket as tags, in lower case, at the moment the offer is first revealed. They are shared by all options of the offer.
  • "What should happen when customer accepts this option" holds the actions, at most one of each kind: one refund ("Refund (percentage)" or "Refund (fixed)"), one voucher (a fixed amount, or a percentage of the paid or of the listed value of the items), one "Escalation to human team" with a handoff reason, and one "Basecamp ticket (supplier investigation)" with a reason. An option without any action is text only: accepting it carries out nothing and settles nothing.
  • "Subtract shipping fee" takes the return shipping fee for the order's delivery country off the refund. It starts unticked and needs a refund action and a fee on the "Return shipping fees" tab. Money actions need an amount above zero.
  • "Return option" marks an option that means sending the item back. It starts unticked. While an agent has an eligible group with such an option and can use an offer tool, it gives the return address for an item only after a return option was shown to the customer for that item, in an earlier reply of the same conversation. An item registered as defective or wrong needs no offer first, unless its photo proof was rejected. A group without any return option does not hold the address back. A group requiring an unkept return window is also excluded from this hold. If the group settings cannot be read, this hold is not applied.
  • "Return address with this offer", unticked by default, lets the AI give the address in the same reply that shows the return option. Unticked, the address can only be given in a later reply.
  • "Last offer after a return pick" takes one offer of the group out of the ladder: a rejection never reaches it. It is given once per item, together with the return address, when the customer chooses to send the item back. One per group, unticked by default, and it cannot itself contain a return option. It requires working offer-acceptance tools and eligible unsettled items. It is not given on an order above the store's high-value limit, or when the AI judges that the customer threatens a payment dispute.

How the ladder runs

  • The AI must first have looked up the order, and each item needs a return request on file that is eligible: one the AI registered with its reason, or one the customer filed in the return portal. An item whose return still waits for an eligibility review gets no offer.
  • Offers are tracked per order and per item, also across tickets about the same order, so a new ticket does not start the ladder again.
  • Normally the first request reveals offer 1. Offers already shown for those items under another group can cause equivalent or smaller percentage offers to be skipped. The next offer is revealed only when the customer clearly rejects the current one or asks for more. Otherwise the current offer is repeated. The ladder advances at most once per reply, although it can skip steps already covered by offers from another group.
  • A fixed refund or voucher amount is used as that number in the currency of the order. It is not converted: 10 means 10 dollars on an order in dollars, although the form shows a euro sign.
  • A percentage refund is a share of what the customer paid for the items the offer covers. A percentage voucher is a discount code for a fixed amount worth that share of the paid or listed value, not a code for that percentage off a next order. One exception: in a group whose reason code is exactly fit_issue, the voucher "based on the % from the paid value" is also worked out on the listed value. Vouchers count against the discount caps under "AI action limits".
  • A refund from an accepted offer follows the store's refund rules in Refund Hub, the same as any refund the AI submits. If it cannot be submitted, the ticket normally goes to your team. If a teammate took over during the run, the refund is left to that teammate without another handover.
  • Accepting an option settles all items the offer was presented for, as soon as one of its actions other than a supplier investigation went through. An accepted offer that only opens a supplier investigation leaves the items open for other offers. If a later action fails after compensation went through, the items stay settled; the earlier payment is not undone. A supplier investigation that fails after compensation does not undo acceptance.
  • No offer is given for an item that already has an accepted offer or a submitted refund, for an item your team already resolved or replaced, for add-ons such as a shipping warranty, or on an order above the store's high-value limit, where the high-value rule of the return policy applies instead.
  • While the store prepares drafts instead of sending replies itself, the AI can present offers but cannot carry out an accepted one.

"After all offers were exhausted" is set once per agent and applies when the customer rejected the last offer. "Assign to human team", the default, hands the ticket over, with the handoff reason chosen under "Reason code when all offers are exhausted" for the group, or a general "offers exhausted" reason. "Tell customer nothing else can be done" makes the AI repeat the last offer and say nothing more is possible; the ticket stays with the AI.

Built-in Tools

Tools are the actions the AI can take. The "Tools" tab lists every built-in tool with a switch. A switch saves at once for every store on the config set; with version control on, it reaches customers with the next release. Initial switches depend on the tool: the offer tools start on, while money tools such as refunds and discount codes start off.

Badges on a tool say what it needs and does: "Requires Shopify", "Requires ParcelPanel", "Requires Return Address", "Requires confirmation", "Read-only" for tools that only look things up, "Makes changes" for tools that change an order, the store or the conversation, and "Unavailable in draft mode" for tools the AI does not get while the store prepares drafts. Switching on one of nine tools asks you to confirm first: refunds, registering a return, discount codes, taking back a voucher, the two order cancellation tools, the two checks of a return's tracking or label, and the return portal lookup. Every other tool switches on without a confirmation, also when it makes changes.

A tool marked "Unavailable in draft mode" cannot be switched on here while the selected store saves drafts: the page says "Cannot enable while in draft mode", and auto-send has to be turned on first. In the same way, a tool that needs Shopify, ParcelPanel or a return address is refused while the selected store has none, with a link to the page where you set it up. The integration checks look at the selected store only, and only when the config set shown is the one that store runs by default. The draft-mode switch lock uses the selected store even when another set is shown. "Makes changes" is broader than this draft lock: tools that tag, add notes, hand over or open supplier tickets can still be available. For another set, or when another store on the same set saves drafts or lacks the integration, the integration switch can be turned on for the set, while runtime checks can still withhold the tool for a store.

"Show technical details" shows the tool's technical name and description. "Duplicate" makes a custom tool from it.

A tool that is on here can still be unavailable for a store, for example when the store lacks the integration. "Validate config tools" under "Advanced" shows that per store.

Custom Tools

Needs the 'View AI Tools' permission.

Custom tools are copies of a built-in tool with a "Custom Name" and "Custom Description" of your own, made with "Duplicate" on a built-in tool. The section appears once the first copy exists. They belong to the whole organization, not to one config set. Each has its own switch and can be edited or deleted; a new copy starts switched off. The custom name uses letters, digits and underscores, up to 64 characters, and must not duplicate another custom or built-in tool name. Its changes save immediately for the organization, outside config-set releases.

A copy does exactly what the tool it was copied from does. The AI gets it under the custom name, but with the description of the built-in tool: the "Custom Description" is a note for your team and is not sent to the AI. A copy of a tool that makes changes is held back in draft mode like the original.

Custom Variables

Variables are named pieces of text a prompt can include, such as a return policy or FAQ answers. They are shared by every store that uses the config set. On the Assistant page they appear as "Knowledge".

"Create Variable" asks for a "Variable name", "What it says" and an optional "Description". The name uses lower-case letters, digits and underscores, starts with a letter or underscore, and has 1 to 100 characters. Built-in names and names already used in the set are refused. "What it says" needs 1 to 10,000 characters; "Description" allows up to 500. The name cannot be changed afterwards. Creating a variable here makes its first value live immediately, including with version control on. Editing a value saves a draft: customers keep the live value until you click "Activate" on that draft. "History" shows every version and lets you activate an earlier value. Activation applies on subsequent prompt preparation, including with version control on. A description-only edit applies immediately. "Delete" removes the variable and its whole history after a confirmation. If prompts still use the variable, the confirmation lists those prompts and the stores answering on the set, so you can take it out of the prompts first.

Return shipping fees

Return shipping fees are amounts taken off a refund, per delivery country, for the stores that use this config set. "General" is the fee for every country that is not listed; rows below it set a fee per country. Each fee has an amount of 0 or more, with no maximum set here, and a currency that starts at EUR. There can be at most 250 country rows, with one row per country. A country set to 0 pays no fee, and an empty "General" row means no fee for unlisted countries. Orders without a delivery country use "General" too.

The fees are used in two places only:

  • offers that have "Subtract shipping fee" ticked;
  • refunds after a verified return to the warehouse, and only when "Subtract shipping fee when customer returns items to warehouse" is on. That switch is off by default and can be turned on once a fee is set. It applies to stores whose "When a return is verified" setting is "Submit for review in Refund Hub" or "Refund automatically". The receipt-photo exception under "Do nothing" also uses this refund path and its fee switch. Removing all positive fees switches the warehouse deduction off.

A fee in another currency than the order's is converted at the day's rate. When no rate is available, the fee cannot be read or the fee is as large as the refund or larger, nothing is refunded and the ticket goes to your team.

Nothing is saved until you click "Save changes". Saved fees apply at once, also when version control is on.

Tone of voice

The "Tone of voice" tab sets how the AI sounds. With no tone set, the AI writes in the style of its prompts. A tone adds a short block after the prompt; it never replaces the prompt text.

  • "Personality": "Friendly", "Professional", "Sophisticated" or "Custom", where you describe the voice yourself in up to 1,500 characters. "Custom" requires a description. "Set a tone" starts from "Friendly", "Sparingly" and empty wording fields.
  • "Reply length": the same setting as "Response verbosity" on the "Main AI Agent" card. It saves as soon as you move it and applies at once, also when version control is on.
  • "Emoji": "Never", "Sparingly" or "Allowed". "Never" instructs the AI to use none. The other choices instruct it to use at most one in email. In chat and WhatsApp, "Sparingly" asks for at most one per reply and "Allowed" permits more where appropriate. These are prompt instructions.
  • "Wording": "Greeting", "Names and terms" and "Avoid", each in your own words, up to 600 characters. The greeting opens every email reply and the AI's first reply in a chat; the AI is instructed to put the customer's first name where you write {first_name}, omit an unknown name and translate the greeting into the customer's language.
  • "Preview" shows a fixed example reply for chat or email in the chosen personality. It is not written by the AI: it follows your personality, greeting and emoji choice, not your other wording texts, and "Custom" is previewed as "Friendly". "Test in Playground" opens the Playground with the tone as it is on screen, saved or not, so you can replay a real conversation with it.

A tone changes wording and style only; it does not change facts, prices or policies. "Save" puts the tone live for every store on the config set on subsequent runs; with version control on, "Save to draft" waits for a release. "Remove tone" returns the AI to the prompt's own wording after a confirmation, or removes it from the draft until the next release when version control is on. The tone never adds a sign-off: email replies keep the store's email signature, or the config set's custom sign-off when that is on.

Version control

Version control makes changes to a config set wait for a release.

"Enable Version Control system (for advanced users)" is off by default. When off, the AI runs the set as it is edited. When you turn it on, a first release named "Default" is made from the current state and the AI runs that release. It cannot be turned on while a configuration test is running on the set. From then on, changes to prompts, specialists, tool switches, offers, the tone of voice, return instructions, the sign-off and the other settings of the set go into a draft and reach customers only with a new release. Intelligence and verbosity levels, the automatic intelligence mode, variables and return shipping fees still apply at once.

"Releases" lists every release with when it was made, its QA score and its QA result. "Live" marks the one the AI runs. "Unreleased changes" says what the draft holds that is not released.

  • "New release" freezes the draft as it is now. Give it a "Release name" of 1 to 200 characters and optionally an "Identification tag" of up to 50 characters, which is added as a tag to every conversation the release answers. The AI runs the new release from the next message. A prompt version that was saved but not activated is not part of the release; a warning names it.
  • "Passing threshold", from 50% to 100%, is the share of AI QA reviews rated good that the release needs. It is shown when Update Compliance QA is switched on for your organization; without it a release gets no QA score. The initial "Default" release gets no evaluation, and a review quota of zero also creates none. The threshold starts from the organization's QA setting, limited to the slider's range. "Update Compliance QA" shows how far the review is. A release fails as soon as it cannot reach the threshold even if every remaining review is good; passing waits until all required reviews are in. When the live release fails, the latest earlier release that passed or has no verdict takes over. Releases with a failed or superseded verdict do not qualify. When no earlier release qualifies, the failed one stays live even though its row says "Auto-disabled".
  • "Revert to" on a row makes the AI run that release at once. It is not offered for the live release or for a release that failed its QA.
  • Opening a release shows its frozen prompts and specialists. "Edit from here" makes the AI run that release, rebuilds the draft from it and discards unreleased changes. The current intelligence, verbosity, routing mode, variables and shipping fees are retained. This cannot be undone.

Turning version control off asks "Turn off version control?": the draft then goes live, including any unreleased changes. The releases are kept. A config set with version control on cannot be deleted.