AI Configuration

AI Configuration in the sidebar of the admin console at app.flexivity.ai decides how the AI behaves with your customers. It has four pages and one card:

Conversation Handling Rules: what the AI may do on each ticket
Answering: how sure the AI must be before it answers, and whether its replies are checked
Conversations: how long a conversation may run, and how it ends when the customer goes quiet
Summary: the instructions behind the summary an agent reads
Unattended changes: what has happened to each kind of change, the circuit breaker’s thresholds, and any halt waiting for you to resume it (see Execute)
Replies to customers: a switch that stops every reply to customers at once, shown beside whether Flexivity has paused unattended changes

Brands and account-wide settings

Each page applies to one brand at a time. Use the Settings for selector at the top of the page to choose one.

The selector also has an account-wide entry, shown as your organization’s name followed by (all brands). A value set there applies to every brand that has not set its own. A value set on a brand applies to that brand only.

Every setting shows where its value comes from:

Built-in default: nobody has chosen this value. It is the value Flexivity starts with.
Inherited from …: set on the account-wide entry, and this brand follows it.
Set here: set on this brand.

Override sets a value for the brand you are viewing. Reset removes it, and the brand goes back to the account-wide value or the built-in default. An overridden setting also shows the default it replaced.

Brands come from your help desk. To turn the AI on or off for one brand, open its help desk on the Help Desk page.

Your instances are configured differently. If your help desk instances hold different values for a setting, the page cannot show one value, so it shows the built-in default. Set the value on this page to apply one value everywhere.

Conversation Handling Rules

A rule says what happens to the tickets that match it. A rule matches on a ticket’s attributes, such as its topic or sentiment. See Help Desk for how attributes are set up.

A ticket that matches no rule gets Assist: the AI drafts a reply and an agent decides whether to send it.

Each rule chooses one of seven levels. They are listed from least to most autonomous, and the badge colour in the rule list follows the same order.

Level	Badge	What happens
Ignore	Black	Nothing runs on these tickets, and not even a category is written. Use it for automated mail and anything the AI should stay out of.
Label	Light blue	The AI works out the ticket’s attributes and writes them to your help desk, so your own triggers and automations can use them. Nothing else runs.
Gated	Blue	A person handles the ticket. No answer is drafted and nothing is retrieved. The ticket still gets its category and summary.
Assist	Dark blue	The AI drafts a reply and an agent decides whether to send it. This is the default.
Reply	Light green	The AI answers the customer directly, with no agent in between.
Clarify	Green	When a request is too vague to answer, the AI asks the customer one clarifying question and then answers their reply. A clear request is answered directly.
Resolve	Dark green	The AI keeps the conversation going until the request is resolved or a limit on the Conversations page is reached.

When more than one rule matches a ticket, the more specific rule wins. A rule refines another when it matches on everything the other one does and more. Among rules that are equally specific, the most restrictive level wins.

A value the system does not recognise is shown as Unrecognised (treated as Gated). Tickets matching that rule go to a person until the rule is corrected.

The rules page also lists the attributes that are switched off for the brand.

Answering

These settings decide how sure the AI must be before it answers, and what it offers when it is not sure.

They control how well an article must match a question, not what following it would do. Nothing here flags a suggested step as destructive or irreversible. Cautions written into your articles are carried through into replies. Risks that are not written down are not inferred, so put them in the article.

The AI finds the articles closest to the ticket, then applies these settings in order:

Is there one clear answer, or several options? (Single-answer margin)
Which articles ground the reply? (Match strength required, Maximum articles used)
Is every claim in the reply supported by those articles? (Answer verification)

When the knowledge base has little to offer, the agent sees related past tickets instead (Maximum related tickets).

Single-answer margin

Default 0.2, range 0 to 1.

How far ahead of the second-best article the best one must be before it is offered as the answer. Below this margin, the agent or customer is offered several options instead.

Lower values mean more tickets get a single answer. Higher values mean more tickets get options.

Match strength required

Default 0.6, range 0 to 1.

How close to the best match an article must score to be used at all, as a share of the best match’s score. Articles below this share are left out of the reply, even if there is room for them.

This is relative rather than absolute, so it carries between knowledge bases. It stops a ticket about one topic from being answered alongside articles about something else.

Sampling puts the useful range between roughly 0.35 and 0.75. The default sits in the middle of that range. Raise it and articles that genuinely fit stop being used. Lower it and unrelated articles start grounding replies.

Maximum articles used

Default 5, range 1 to 10.

The most articles that can ground a reply and appear beside it.

This is a ceiling, not a target. Only articles close in relevance to the best match are used (see Match strength required), so raising it changes nothing unless several articles genuinely fit. More articles give the AI more to work from, and more to misquote.

Default 3, range 1 to 10.

How many past tickets are shown to the agent when the knowledge base has little to offer. Related tickets are shown as links for the agent. They are never used to ground an answer, because a closed ticket shows that an agent stopped working on it, not that the resolution was right.

Answer verification

Every reply is built from your knowledge base, and every claim in it names the article it came from. Before a reply reaches anyone, Flexivity checks that every claim actually cites a retrieved article. A claim that cites nothing, or cites an article that was not in the retrieved set, causes the reply to be withheld. This check is always on and has no setting — it is part of how replies are built.

Answer verification adds a deeper check: a second AI call that reads each cited article and judges whether it actually supports the claim. It catches claims that sound right but are not in your articles. A common case is a correct general instruction that your article does not actually say, such as a menu path for a different version of a product.

There are two settings, one for tickets and one for chat, and both are off until you turn them on.

Answer verification is an add-on, charged per ticket. Without it, both settings show off and cannot be changed, and no reply is checked whatever a setting said before. Contact us to add it. Once it is on your plan, both settings can be changed as below; if it is later removed, verification stops and your chosen values are kept for when it returns. The always-on citation check above is not part of the add-on.

Answer verification — tickets

Default off. Choices: off, record, enforce.

Covers the replies the AI sends to a customer on a ticket, and the drafts an agent reviews at Assist. Both are checked before anybody reads them, so they are one setting.

Off: replies and drafts are not checked.
Record: each one is checked and the result is recorded, but nothing is held back. Use it to see how often a reply would have failed before you let the check act.
Enforce: the check reads every sentence and acts on the ones that matter.
- A minor issue, such as a general tip or a harmless rewording, is left in and recorded.
- A risky detail the knowledge base does not support, such as a number or a step, is removed or softened.
- A serious statement, such as a promise about a refund, a price or a deadline, is removed. When a reply goes to a customer with something serious removed, the rest is sent with a note that a colleague will pick up the rest, and the ticket goes to an agent with a list of what was left out.
- A reply is held back only when nothing useful is left, or the check could not run. The ticket then goes to an agent, who is told the answer could not be verified.
An agent reviewing a draft at Assist gets the corrected draft, with a line naming anything that was removed or softened.

Answer verification — chat

Default off. Choices: off, record, escalate.

Covers the replies the AI sends in chat. A chat customer is waiting, so a chat reply is checked after it has been sent. A failed check cannot take the reply back.

Off: replies are not checked.
Record: each reply is checked and the result is recorded, and nothing else happens.
Escalate: when the reply contained a serious unsupported statement, such as a promise or a price, the reply the customer has already read stands and the conversation is handed to a person. Lesser issues are recorded and the conversation carries on.

A value you chose before is now the chat value. If a brand was set to record or escalate, it keeps that value here and chat behaves as it did. Its ticket setting starts at off.

Whichever you choose, the outcome is recorded against every reply, and AI Activity can filter on it.

Never suggest

Default: none.

A list of things replies must never tell a customer to do, in your own words, such as “reinstall the app” or “contact your bank”. Up to 25 entries, each 3 to 200 characters.

Entries are matched by meaning, not by exact wording, so “remove the app and install it again” counts as “reinstall the app”, while “restart the app” does not. A statement that matches an entry is removed or softened before the reply is sent.

The list is read by answer verification, so it only applies where Answer verification is set to record, enforce or escalate. With verification off, nothing reads it.

Conversations

These settings decide how long the AI keeps a conversation going, and how it ends one when the customer stops replying. They apply to tickets handled at the Clarify and Resolve levels.

Ticket and chat settings are separate. Ticket waits are measured in hours, chat waits in minutes.

Ticket conversations

Clarifying questions in a row — tickets

Default 1, range 0 to 3.

How many questions the AI may ask in a row on a ticket before it must answer with what it has — conditionally, covering the likely cases, if it cannot be certain — or hand the ticket to an agent.

“In a row” means with no answer to the customer’s problem sent in between. A conversation that asks, answers, and later asks again about something new has asked one question in a row each time, not two.

Reaching the bound does not mean giving up. The AI is told to answer with what it has; it hands over only when it genuinely cannot answer from your knowledge base.

Clarifying questions in a row — chat

Default 2, range 0 to 3.

The same bound in chat. The default is higher because narrowing a problem over a couple of short exchanges is how a chat widget is normally used, while a ticket can carry a longer answer that covers each likely case in one reply.

Messages per auto-resolved conversation — tickets

Default 5, range 2 to 20.

How many messages a customer can exchange with the AI before a person takes the ticket over. This counts every message, not only questions. It does not count the closing note the AI sends when a conversation ends. The default is roughly two and a half exchanges.

Messages per auto-resolved conversation — chat

Default 7, range 2 to 20.

The same bound in chat, and the default is higher on purpose. Chat asks one question per reply and may ask twice in a row, where a ticket asks once and can put several questions in that one reply — so a chat conversation spends more messages reaching the same point. Its replies are shorter too, so an answer that fills one paragraph on a ticket may take a couple of chat turns.

Wait after asking a question

Default 24 hours, range 1 to 672 (28 days).

How long to wait for an answer to a question the AI asked before ending the conversation.

This is shorter than the wait after an answer. A question nobody answered is a conversation going nowhere: holding it open gains nothing and delays the person who now has to deal with it.

Wait after sending an answer

Default 72 hours, range 1 to 672 (28 days).

How long to wait after an answer before treating silence as the end of the conversation.

The default crosses a weekend on purpose. A customer who reads the answer on Friday evening and would have replied on Monday should not find the conversation closed behind them.

Waits are limited to 28 days because your help desk closes a solved ticket on its own schedule, and a longer wait would never be reached.

When a question goes unanswered

Default hand over. Choices: hand over, resolve.

What happens when nobody answers a question the AI asked. Silence after a question tells you nothing, so by default the ticket goes to an agent. Marking it solved would claim something the customer never said.

When an answer draws no reply

Default resolve. Choices: hand over, resolve.

What happens when a customer does not reply to an answer. Silence after an answer is weak evidence that it worked, so by default the ticket is marked solved.

It is recorded as an assumed resolution rather than a confirmed one. If the customer comes back, the resolution is withdrawn. The AI marks the ticket solved and never closes it, so the customer can always reopen it.

Internal notes on tickets

Default Off. Choices: On, Off. Set per brand.

Whether the AI adds internal notes to your tickets. Today that is one note: when the AI hands a conversation to a person, it can leave an internal note on the ticket saying why, so whoever picks the ticket up — and anyone who exports it later — can see the reason without opening the AI’s activity log. Internal notes are seen by your agents and never by the customer.

This is off unless you turn it on, which is a change from before. The note used to be written on every hand-over, which added an entry to every escalated ticket. Now no brand writes it until an administrator switches it on for that brand, or for a parent brand it inherits from.

Turning it off changes only the note. The hand-over still happens, the customer still receives the same message, the agent still sees the AI’s card on the ticket, and the hand-over is still recorded in the activity report. The conversation also still ends its automation exactly as before.

The setting covers every internal note the AI writes on a ticket, so any note added in future is governed by the same switch. Chat conversations have no ticket to annotate and are not affected. Everyone in your account can see the setting; only an administrator can change it.

Chat conversations

Chat has waits and endings but no message limit.

Chat: wait after asking a question

Default 20 minutes, range 3 to 5760 (4 days).

How long a chat customer may be silent after the AI asked them something, before the conversation ends.

This is longer than the wait after an answer, because ending a chat with a question left hanging is the worse mistake. Very short values measure how long somebody takes to type, not whether they have gone.

Chat: wait after answering

Default 10 minutes, range 3 to 5760 (4 days).

How long a chat customer may be silent after an answer before the conversation ends. The default matches the value Zendesk uses for the same decision on its own bots.

Chat: when a question goes unanswered

Default resolve. Choices: hand over, resolve.

What happens when a chat customer stops replying to a question. On chat the wait is minutes, not days, and somebody who has closed the tab has gone rather than paused to think. So by default the conversation is marked resolved.

It is recorded as an assumed resolution and withdrawn if the customer comes back. Choose hand over to put a person on every abandoned chat.

Chat: when an answer draws no reply

Default resolve. Choices: hand over, resolve.

What happens when a chat customer does not reply to an answer. Silence after an answer is the ordinary end of a chat that worked.

Default Off. Choices: On, Off. Set per brand.

Turn this on if your chat lets customers sign in. It matters only where you have enabled data connector lookups: the AI looks up a chat customer’s record only when they are signed in, and a customer who is not signed in is never looked up. When they ask about their own order or account, the AI says it cannot see their account in this chat and offers a person. With this setting on, it also tells them they can sign in and ask again.

It is off by default because nothing tells Flexivity whether your chat has sign-in set up, and offering a sign-in your chat does not have would leave the customer nowhere to go. Everyone in your account can see the setting; only an administrator can change it.

Procedures

A procedure tells the AI how to handle a kind of request in your own words: what to look up, what to tell the customer and when to hand over. See Procedures.

Brand voice

Brand voice sets the tone of the standard messages your customers receive: the welcome, the messages sent when a conversation is resolved, and the messages sent when a conversation is handed to a person. Choose it per brand. The tone chosen for all brands applies to any brand that has not chosen its own.

Tones

Friendly: relaxed and approachable. This is the default, and the wording customers already receive.
Professional: courteous and polished, in complete sentences.
Matter-of-fact: short and direct. Says what is happening and nothing more.
Warm: caring and reassuring, with a little more empathy in every message.

Each tone shows a sample of its messages; View all messages shows the full set. A tone marked Also shapes AI-written replies also adds a short instruction to how the AI writes its answers. We add that only for a tone whose instruction we have measured and found not to make answers less accurate; the others change the standard messages only.

Standard messages

The page lists every standard message your customers can receive, grouped into Welcome, Resolution and Hand-overs, with when each is sent and whether it is sent on tickets, chat or both. There are two welcomes, one for tickets and one for chat: the chat welcome arrives on its own while the AI works, and the ticket welcome opens the first reply.

Hand-overs are grouped by what the customer needs to hear, so one message covers several reasons. For example, General hand-over is sent whenever one of your rules applies, a conversation reaches a limit or stops moving forward, or AI processing is paused. Your agents still see the precise reason on the ticket.

Two chat hand-overs are for a customer who is not signed in and asks about their own account, which cannot be looked up in the chat: one says a colleague will help, and the other, sent when Customers can sign in to chat is on, also says that signing in and asking again lets their account be looked up.

To use your own words, choose Edit on a message. Messages are sent as plain text, up to 500 characters. Revert goes back to the standard wording of the tone selected now. Changing the tone changes every message you have not edited, and keeps your wording for the ones you have; the page says how many before you confirm.

Messages are edited per brand. When you are viewing all brands, you can set the tone but not edit messages.

Disclosures

Automatic messages end with a short disclosure saying they were sent automatically and, where it applies, how to reach a person. Disclosures follow the tone and cannot be edited, so customers can always tell when they are talking to the AI.

Summary

The summary is what an agent reads about a ticket, including when it is handed to them.

Prompt: the instructions the summary is written from. Until you change it, the built-in prompt is used and shown on the page.
Temperature: how much the wording may vary, from 0.1 to 0.5. Lower values keep summaries of similar tickets consistently worded.

Both apply per brand, like every other setting on AI Configuration.

History lists the brand’s earlier prompts, newest first, with who saved each one. Restore loads a version into the editor so you can review it; nothing changes until you save. To go back to the built-in prompt, use Remove override, or restore a version marked went back to the built-in default and save it.

Replies to customers

The Replies to customers card, at the bottom of the AI Configuration page, stops the AI from sending anything directly to a customer. It covers every help desk you have connected, whichever brand you are viewing.

Use it when you need replies to stop now, for example while you investigate a reply that should not have gone out.

Your rules are not changed. The card only decides whether what they allow is sent. While replies are stopped:

Reply, Clarify and Resolve tickets are handled as Assist: the AI drafts a reply and an agent decides whether to send it.
Everything else, including summaries, drafts for agents and classification, carries on.

Resume replies on the same card turns replies back on, and your rules apply again to new messages.

While replies are stopped, nothing runs with Execute either: those tickets are handled as Assist, so an agent approves every change. Beside this card, Unattended changes says whether Flexivity has paused unattended changes for every account during an incident; only Flexivity can lift that.