Appearance
Rules
Rules pick out the conversations your team should handle with extra care: a refund request, a legal letter, a customer asking for a person, a draft that does not fully answer. Before anyone on your team sees a new reply draft, Overdesk's reviewer reads it beside the conversation and the sources it was written from, and your rules act on what the reviewer finds. Every workspace starts with a set of starter rules, and admins can adjust them.
What the reviewer reads
For every reply draft, the reviewer reports:
- whether the customer wants information, or wants something done, such as a refund, a cancellation or a change to their account;
- which of your topic rules the conversation is about;
- how the customer sounds, and whether they asked for a person;
- whether the draft fully answers the customer, using only what your knowledge sources and the conversation support. A draft saying a feature, setting or guide doesn't exist counts as unsupported unless your sources or the conversation say so. A draft that answers in full and also asks the customer something, or suggests a next step, still fully answers; one that can't answer until the customer sends more information, or that answers for a different plan or version than the customer's, does not. When all a draft leaves unanswered waits on something only the customer can tell you that they haven't given anywhere in the conversation, such as their site's address, a version or an error message, and the draft asks them for it, the reviewer says so;
- whether the draft says someone tested something, such as "I set up WordPress 6.8 with your plugin and the gallery opened", and, when the draft was tried on a test site, what that test showed;
- whether the draft says a person wrote it, or denies that AI did;
- whether the draft talks about its own sources or caution instead of the customer's question, such as what the documentation it was given covers, what it couldn't find, what it can't see, or that it won't guess;
- whether the customer's latest message is a request for support at all, or marketing, spam, an automatic notice from a website or service, a bounce or an out-of-office reply;
- whether the draft promises something your team has to do, such as a date, a refund, passing the question to your team, getting back to the customer later, fixing something or logging into the customer's site, or says something was done that the conversation doesn't show, and whether passing the question on or getting back to the customer is all it promises. Saying what happens once the customer replies, such as "if that doesn't fix it, send me the link and we'll look further", is not a promise;
- whether the draft repeats something only your team's internal notes say;
- whether the draft contradicts what the customer was already told, by someone on your team or in a reply Overdesk sent;
- when Overdesk handed the conversation to a person, whether that issue has since been dealt with (see Escalation).
It reads the customer's messages, the replies your team sent and the replies Overdesk sent, as far back as the draft read them, and everything the draft was written from, in full: your product description, the knowledge base articles with their links, the customer's earlier conversations, their account in your connected store, past answers and saved replies, and the pictures, PDFs, linked pages and text files the customer sent. A picture, page or file supports what it shows about the customer's own setup, never a claim about your product; a linked page of your own documentation site counts only as Overdesk stored it from your site (Pictures and links). Only your team's own words count as support: in a forum thread or an issue that is your team's reply or status, never what a customer wrote there, and a customer's earlier subject lines show what they asked, not what was done. When the draft opened files from your code, the reviewer reads those files too and judges the reply against them: what the code does and the comments in it are your team's words, text a file only stores or quotes (a test's sample message, for example) is not, and a change in your code you haven't released yet doesn't answer a customer on an earlier version. When that material, or the customer's messages, is too long to read in full, when the draft relied on your list of files without opening one, when it names a file from your code (Overdesk is told never to), or when the check of your code doesn't finish in time, the reviewer can't approve the draft, and it waits for a person. It recognises the replies Overdesk sent itself and reads them as Overdesk's: one never counts as your team answering the customer, as support for what a draft says, or as proof that something was done. Your team's internal notes are read for two questions only: whether the draft repeats one, and, after a hand-over, whether the issue has been dealt with. Customer names and email addresses are redacted before it reads anything, exactly as they are for drafting, and it runs on your workspace's own Anthropic API key.
What a rule does when it matches
While sending is off, the draft is written and placed in your help desk as usual, whatever the rules say. Once your workspace turns sending on, a reply a rule under Answer automatically when approves is sent instead, and every other reply is placed as usual; see Sending. On top of that:
- The decision is recorded on Decisions with the rule's name and its reason.
- The draft carries a badge on Agent drafts: Held for a topic rule, Escalated for an escalation rule.
- A rule under Escalate when raises an alert for your team.
- A rule with someone chosen under Hand over to assigns the conversation to them in your help desk and raises an alert naming them. See Escalation.
Questions and requests
A topic rule tells a question about its topic from a request on it. "What is your refund policy?" is a question, and it is answered like any other. "Please refund my order" is a request, and it waits for your team. Asking how to make a change yourself counts as a request too, so "How do I cancel my subscription?" waits for your team, and so does a customer disputing a charge.
Legal and privacy holds questions as well, because an answer about data protection law is a statement a person should make.
A few phrases wait for your team even when they come as a question, because they only ever mean the customer wants their money or their account: "refund me", "my money back", "chargeback" and "cancel my" under Refunds and cancellations, and "delete my account" under Account security. They count in the message being answered, so a customer who asked for a refund last week and asks something else today gets an answer. "Is there a money back guarantee?" is still answered. Turning the rule off turns its phrases off with it.
A topic rule's keywords catch a request the reviewer did not tie to the topic, and they do the whole job whenever a review is not available. They never hold a customer the reviewer found was only asking a question.
One decision per draft
Each draft gets a single decision, and the first matching rule makes it. Overdesk checks Always hand to a person first, then Escalate when, then Never answer automatically when, and Answer automatically when last. The tab lists the groups in that order, and the rules inside each group in the order they are checked.
That order matters when you hand rules to people. A customer who asks for a real person about a refund matches Refunds and cancellations first, so Human agent requested does not fire for that draft. If refund conversations should reach someone, choose a target under Hand over to on Refunds and cancellations too.
It matters for your own rules too. A rule you add is checked after the starters, so a rule about chargebacks sits behind Refunds and cancellations, which matches the same conversations first. The New rule form warns you when a keyword you type is already claimed by a rule checked earlier. Give each rule a distinct topic: overlapping topics also make it harder for the reviewer to tell which one a conversation is about.
How matching works
- Keywords match anywhere in the subject or in the customer's own messages, in any case, including inside longer words.
refundalso matches "refunds". - Tag conditions match the tags on the conversation, including the ones Triage adds, and the type and urgency the reviewer reads from the conversation itself.
- The customer list matches a full email address exactly, or a whole domain:
acme.comcovers every address at acme.com and its subdomains.
The starter rules
Under Always hand to a person:
| Rule | Starts | Holds |
|---|---|---|
| Billing | On | Requests about charges, invoices, payments, pricing, plan changes and upgrades, and buying, renewing, moving or freeing a license or subscription. Trouble installing or activating a license key is a product question, not a billing one |
| Refunds and cancellations | On | Requests for a refund or money back, chargebacks, payment disputes, and cancelling a subscription or an order |
| Legal and privacy | On | Legal threats, formal complaints, data protection law such as GDPR, and requests to see, export or delete personal data. Questions included |
| Account security | On | Requests about access to the customer's account with you: signing in, being locked out, two-factor codes, a hacked account, fraud and identity checks |
| Excluded customers | On | The customer's address or domain is on your list, which starts empty |
Under Escalate when:
| Rule | Starts | Matches when |
|---|---|---|
| Human agent requested | On | The customer asks for a person or asks whether a person or an AI is writing to them, or writes "speak to a human", "real person", "talk to someone" or similar |
| Frustrated customer | Off | The customer sounds frustrated |
| Customer keeps writing back | Off | The customer has written back 3 times or more since the first reply that reached them |
| Weak answer | Off | The draft only partly answers the customer, does not answer, or says something your knowledge sources do not support |
| No knowledge-base match | Off | Nothing in your knowledge sources clearly matches the question |
| Test site could not finish | Off | The draft was tried on a test site that failed or ran out of time before it finished |
Under Never answer automatically when:
| Rule | Starts | Matches when |
|---|---|---|
| High urgency | On | The conversation is urgent, by its tags or by the reviewer's reading |
| Urgency unknown | On | Neither Triage nor the reviewer could tell how urgent it is |
| Bug reports | On | The conversation is a bug report, or mentions error, broken, crash or "not working" |
| Automatic replies and bounces | On | The message reads like an out-of-office reply, a no-reply sender or a bounce |
Under Answer automatically when:
| Rule | Starts | Matches when |
|---|---|---|
| Answers the reviewer approves | On | The reviewer finds the draft fully answers the customer and says nothing your knowledge sources do not support, your knowledge sources clearly match the question, and the conversation is not urgent |
| Approved answers the test site checked | Off | Everything Answers the reviewer approves needs, the test site checked the draft and finished, and what it showed is the problem the customer reported or that it works as designed |
| Also send follow-ups | Off | The message is a follow-up nudge Overdesk wrote on its own. It never matches a reply, and it sends a nudge only where Overdesk itself answered the customer's last message and nothing else holds it |
Two kinds of draft that don't fully answer count as answering for Weak answer, Answers the reviewer approves and Approved answers the test site checked, when they are sent. One that waits for your team for any other reason is still a weak answer, and Weak answer escalates it. One answers all it can, promises nothing, and asks the customer only for information they haven't given yet, never a way into their site or accounts, and never a secret such as a token, a config file's contents, a cookie or a connection string. The reviewer is told to hold those in any language, and for a draft written in English Overdesk also checks for them itself, holding one that names such a secret anywhere, even in a how-to step, though not a word that names no secret there, such as clearing cookies, a cookie banner or blurring a token out of a screenshot. It is sent only when each question in it is worded the way Overdesk knows a plain ask, such as "Which version are you on?" or "Could you send a screenshot?", and asks about something the customer's message or your knowledge sources mention, such as their version, a screenshot, a link, the error message or the steps they tried. The questions go out in Overdesk's words: the draft is sent exactly as written up to its first question, and that part may promise nothing itself ("We'll know more once we see it" keeps the draft for your team); in place of the rest Overdesk asks for each thing the draft asked for, from a fixed list: a screenshot of the error, the page, the settings page or what the customer sees, the link to the page where it happens, the address of their site, the exact error message, the steps they took, what they see, and which version or browser they are using. A question for anything else, one that names what it is about ("Which version of FooGallery are you on?"), a screenshot "of it" or "of that", a yes-or-no question, and anything after the questions but thanks, an apology, a hope that it helps, being glad to help, an offer to answer more and a plain sign-off, keep the draft for your team. It is held when it asks for someone's personal details, such as a full name or a phone number, or for their customers' data or a copy of their site, when it names a secret or mentions money anywhere, when it mentions a fix, workaround or solution, and when Overdesk spots another language in it. Overdesk knows each of these by its common wordings, not every wording. Overdesk reads it as written and again as the customer will see it. Such a draft is sent like any approved answer. The other can't answer, and tells the customer their question was passed on: it is sent only when it can be handed to someone, and Overdesk then hands the conversation over (how). Both still need everything else those rules need, including nothing your knowledge sources do not support.
A draft Overdesk tried a test site on waits for a person, so Approved answers the test site checked sends nothing yet, even when it is on (see Holds on what a reply says). Bug reports is checked before either automatic answer, so while it is on a bug report is never answered automatically.
In Draft assist, nothing is sent. A draft a rule under Answer automatically when approves shows on Agent drafts as Would answer automatically, and it waits for a person like every other draft. With Send replies automatically, a reply such a rule approves is sent, as long as none of the checks on Sending holds it, and shows as Answered automatically.
The escalation rules start off because each raises an alert on every draft it matches. Turn one on and watch Decisions for a few days before you hand it to anyone.
Under each rule that is on, the tab shows how many drafts it decided in the last 30 days.
Holds on what a reply says
Overdesk holds these drafts for a person itself, for what the reply says, for what the customer's message is, because Overdesk tried a test site on the draft, because the reply was allowed to ask the customer for their Site Health info, because the reply was cut off before it finished, or because the customer posted a link that logs in to their site. They are not rules, and nothing in your workspace switches them off:
- A draft that says a person wrote it, or denies that AI did, shows on Decisions as The reply says a person wrote it. It would tell your customer something false about who is answering. A draft that says AI helped write it, and that a person on your team can take over, is not held.
- A draft that says someone tested something, when no finished test on the test site backs it, shows as The reply says it was tested, but no finished test backs it up. A reply claiming a test that never happened is a false statement about what your team did. Starting the test site without checking anything on it is not a finished test. Overdesk also reads the reply's own words for the plainest claims, such as "I couldn't reproduce it" or "I tried it on a fresh install", so the draft is held even when the review misses the claim.
- A draft Overdesk tried a test site on that your rules would otherwise answer automatically shows as Overdesk tried a test site on this draft, so a person checks it first. A test site that was not available does not count. Replies that rely on a test site are not sent automatically yet, which is why Approved answers the test site checked sends nothing for now.
- A draft for a problem with one of your products, where the test site did not finish and no code was read, shows as The test site did not finish and no code was read, so a person checks it first. Its reason says what happened to the test site: it failed, timed out, was not available, checked nothing, or was not used. The investigation is incomplete, so a person decides what to check next. This applies where Overdesk starts the test site itself.
- A draft that was allowed to ask the customer for their Site Health info shows as Overdesk let this reply ask for the customer's Site Health info, so a person checks it first. It is held whether or not the reply made the ask, and no rule can send it. Replies like this are not sent automatically yet.
- A reply that was cut off before it finished shows as The reply was cut off before it finished, so a person checks it first. Its writing stopped before the end, so part of it may be missing. It is never sent automatically, whether it is a reply or a follow-up nudge, and no rule can send it. It waits for a person, with this reason on Decisions.
- A draft answering a customer who posted a link that logs in to their site, such as a temporary login link, a WordPress reset link or a magic login link, shows as Customer posted a login link. Overdesk hides the link from its records (see How customer data is handled), and a person has to handle it, so no rule can send the reply. When it can, the conversation goes to the person you marked as logging in to customer sites (see Escalation). A link posted before the last reply from your team that reached the customer, or only quoted from an earlier message, does not count.
- A draft that talks about its own sources or caution instead of the question shows as The reply talks about its own sources or caution: what the documentation it was given covers, what it couldn't find, what it can't see, or that it won't guess. Only software writes that way, so your customer would know an AI wrote the reply. Saying what isn't known yet and what happens next, linking the guide the customer asked for, or saying where a fact comes from is not held.
- A draft answering a message that isn't a request for support shows as Not a request for support: marketing, spam, a scam, an automatic notice from a website, plugin or payment service, a bounce, an out-of-office reply, or another help desk confirming it received a message. The language a customer writes in is never a reason: a question in Spanish or Turkish is support. With Screen out non-support mail on, most of these messages never get a draft at all: see Mail that is not support. This hold covers the rest.
- A draft answering a message a machine sent shows as An automatic message, like an out-of-office reply. Overdesk finds these without the reviewer, from who sent the message and how it opens: a delivery report, an out-of-office reply arriving in the customer's own conversation, or a help desk's acknowledgement naming a ticket number or a wait. Answering one is how two automated systems end up writing to each other.
- A draft for a customer who asked for a person shows as The customer asked for a person, even if you turn off Human agent requested, which otherwise escalates it first.
- A draft that promises something your team has to do shows as The reply promises something the team has to do: a date, a refund, a credit or an extension, passing the question to your team, getting back to the customer later, fixing or changing something, logging into or checking the customer's site or account, or saying the question was passed on when the conversation doesn't show it. Someone on your team has to keep that promise, so they see it first. In a reply written in English, a second check, Overdesk's own and not the reviewer's, also looks for these promises, whether or not you have set up escalation targets: a promise to fix or change something, to refund or credit, to send something or to look at the customer's site, and saying your team did one of these or that it is on its way. A reply that names an amount, or speaks of the customer's own refund, charge, credit, invoice or payment other than to say it can't be done, is held the same way, whatever words it uses; your refund policy and the steps to ask for a refund are not. The exception is a reply whose only promise is passing the question on, or that the customer will hear back: with sending on, it can be sent, and Overdesk then hands the conversation to the target it names, or to the person you marked as keeping the promises replies make. See Escalation. A reply asking the customer for access to their site, such as their login details, a temporary login link, a temporary admin user, an account or a collaborator request, or saying the writer will log in to their site, is held the same way, whatever else it says, and goes to the person you marked as logging in to customer sites (see Escalation). Offering more help, or saying what you'll look at once the customer replies, is not a promise; a refund, a credit or a fix promised for then still is.
- A draft that tells the customer something only your internal notes say, such as what a developer said or a discount your team discussed, shows as The reply repeats something from a teammate's note. A note asking for something to be passed on to the customer is fine to pass on.
- A draft that corrects or goes against what the customer was already told, by someone on your team or in a reply Overdesk sent, shows as The reply contradicts an earlier reply, so a person on your team can put it right.
- A draft that asks the customer again for information, access or a file that they already gave in the conversation, without saying why, shows as The reply asks again for something the customer gave. Making a customer repeat themselves is a poor answer, and a person can check whether the first one was missed. A reply that gives a reason is not held: the earlier one expired, didn't work, was incomplete or wrong, or the customer asked to change it. When the reviewer was shown only part of a long conversation, it cannot tell and holds nothing on this account.
- A draft with a link your team's own material didn't give it shows as The reply has a link not found in your team's material. Overdesk checks this itself, without the reviewer. Every link in the reply, as your customer would receive it, has to appear in what the draft was given from your side: your instructions and review link, your product description, your knowledge base articles' links and the text of your own articles, your saved replies, the code files the draft opened, and your team's replies in the conversation. A past answer to another customer counts only for a link to a site those already link to, because it can repeat that customer's own address. A link that only the customer's messages, a page or file they sent, a picture, what a customer wrote in a forum thread or an issue, or a note gave it is not your team's, so the reply waits for a person. This keeps a link someone planted in a page or screenshot your customer shared from reaching your customer from your address. Every form counts: a link, a web address written out (
https://…orwww.…), and an email address, written out or as a link. Two links still match when one useshttpand the otherhttps, the site name's capital letters differ, or one ends in/or a#part; a different page or query, an email link whose subject, body or anything else after?differs from your team's, or a site name that only looks the same, holds the reply. A link's title, the words of its#part, and an email link's subject and body reach your customer too, so the checks Overdesk makes itself on the reply's words, such as a claim that a person wrote it, a test that never ran, or a request for access, read them as well. A link to a phone number always waits for a person. A site name written withouthttps://orwww.(such asexample.com/page) and a phone number written out are not read as links, though your customer's email app may make them clickable.
On Agent drafts, each of these carries a red Held badge and shows in the list.
The reviewer also has to answer each of these, how the customer sounds and whether they asked for a person, before its approval counts toward an automatic answer. A review that leaves any of them unanswered approves nothing.
Change the rules
- Open Guardrails → Rules. Each rule is one line: its switch, its name, who picks it up, and how many drafts it decided.
- Use the switch beside a rule to turn it on or off.
- Click a rule to open it. You can change its Name, its Keywords as a comma-separated list, and for Excluded customers the addresses or domains under Customers / domains. On a topic rule, Questions decides whether questions about the topic wait as well as requests. On Customer keeps writing back, Times they write back sets how many messages it takes. On Test site could not finish, Test site result chooses which results it matches. A test site that started but checked nothing is its own result, Started, nothing checked, and the rule leaves it out unless you tick it. On a rule under Answer automatically when, Only when the test site finished its check makes it answer only drafts the test site finished checking. Approved answers the test site checked also needs the test site to have shown the problem the customer reported, or shown it works as designed (what the result means). Those drafts still wait for a person, so a rule with it ticked sends nothing yet.
- Under Hand over to, choose an escalation target, or one of the two ways of choosing nobody. See Escalation. Targets are set up on the Escalation tab first.
- Press Save changes in the bar at the bottom of the screen. It appears as soon as something is unsaved, and Discard puts everything back.
Overdesk warns you if you follow a link with unsaved changes. It cannot catch the browser's back button or signing out, so save before you leave the page.
A rule's Name is part of how it matches: the reviewer is given the name with the topic, so naming Billing "Payments and invoices" changes what the reviewer reads.
Only admins can change rules. Everyone else sees them read-only. The conditions a rule reads from the reviewer and from tags show as a line of text under the rule and cannot be edited here.
If another admin saves while you are editing, Overdesk refuses your save rather than overwriting theirs, and offers to reload. Your own unsaved changes are lost in that case, so save before you walk away.
What you change is kept when Overdesk improves the starter rules. A workspace owns each rule's switch, name, keywords, customer list, hand-over target, reply threshold, questions setting, test site results and finished-check setting; everything else about a starter follows the current version.
Your own rules
New rule adds a rule about a topic your team cares about, in your own words.
- Press New rule.
- Give it a Name, and under What it is about describe the conversations the way you would tell a new teammate. The reviewer reads this, so it is the rule.
- Choose what happens when it matches: Always hand to a person, or Escalate and alert the team.
- Keywords are optional. They catch a request the reviewer did not tie to your topic, and they do the whole job whenever a review is not available.
- Choose Hand over to, then Add rule, then Save changes.
Your rules are checked after the starter rules in their group. You can have 20 topic rules switched on at once, starter rules included, because the reviewer reads them all for every draft.
Open one of your rules to edit or delete it. Deleting takes effect when you save, and the rule's past decisions stay in the log.
Excluded customers
Excluded customers marks conversations from people you name, such as a key account or a customer your account manager handles personally. Their drafts still arrive as usual, carry the Held badge, and show in the decision log with the reason. To send those conversations straight to one person, choose that person under Hand over to on this rule.