The Elbuz helpdesk for online stores lets AI answer customers and handle simple order actions within your store policy. Every decision can be undone.
When you need this:
- Half the emails: “where's my order?”, “cancel it” — AI answers and cancels unpaid orders after a “Yes”.
- Nobody answers nights and weekends — customers get instant messenger replies; complex cases wait for staff.
- No answer in the knowledge base — AI asks the warehouse or a manager and names a reply time.
- Afraid to let AI act — “Test” only records what AI would do.
Where to start: “Tickets” gear → “Autonomous support” → turn it on in “Test” mode.
What autonomous support can do
- Order actions — cancellation before shipping, changing the Nova Poshta branch, returns within the return period, reshipment — after the customer confirms with “Yes”.
- Auto-replies in chats and messengers — replies to the customer using the same engine as for emails.
- Auto-replies to emails — the same “Pulse” suggestions, but with a test mode, review and a breaker.
- Proactive notices — “your parcel is waiting at the branch”, reminders before a return, review requests.
- “Ask an employee” — AI asks a colleague and continues the reply to the customer once they answer.
- Control — the “Review” queue, automatic trust downgrade, the breaker, test mode and backtests on past tickets.
- Conversation quality — AutoQA scores closed tickets, and a spike in a single topic shows up right away.
The difference from the AI assistant for agents is who clicks “Send”.
| AI assistant | Autonomous support | |
|---|---|---|
| Who replies to the customer | an employee, from an AI draft | AI itself, within the policy and trust |
| Order actions | “AI suggests” buttons, a person clicks them | AI offers them to the customer and performs them after “Yes” |
| What you need to turn on | nothing, works out of the box | the main switch, the policy, “Live” mode, trust |
| Default | on | everything off |
AI never moves money: cancelling a paid order creates an unposted draft cash payment order for a person, and a return creates an unposted document.
How to turn on autonomous support safely
Nothing reaches the customer until three conditions are met at once: the main switch is on, the policy is applied, and the action type and channel are in “Live” mode. Here is an order in which AI never acts blindly:
- On the “Policy” tab, set up the action types and click “Apply” — version 1 appears.
- Set “Mode for all types” and the action types you need to “Test”.
- Turn on “AI may act by itself under the policy” and confirm with “Turn on”.
- For a week, watch the “Review” queue: what AI would have done and whether you agree.
- For action types that move money, run a backtest on past tickets on the “Test and replay” tab.
- Switch one action type in one channel to “Live”, then the rest.
The main switch “AI may act by itself under the policy” is also the emergency button. Turning it off takes effect instantly: no new actions are offered, and those awaiting customer confirmation are paused and handed to a person. Turning it back on doesn't lift the pause. You can see next to it who switched it and when.
AI signature and quiet hours
Messages AI sends to the customer on its own — an auto-reply in chat or by email, a “Yes/No” question, an action summary, “I am checking with my colleagues” — are signed with the line “Answered by the store's AI assistant”. Replies sent by an employee aren't signed.
- The “Sign AI messages to customers” checkbox is on the “Channels and proactive” tab and is on by default.
- The signature text: “Chat dialogs” → gear → “Chat settings” → “Automatic message texts” → “Tickets: AI signature under messages to the customer”. Empty text means no signature.
Quiet hours (21:00 to 09:00 by default) apply to messages AI writes to the customer first. Proactive notices and confirmation reminders wait until quiet hours end. The “confirmation expired” message and apologies for a colleague's delayed answer aren't sent at all during this time. A reply to a customer who wrote at night isn't held back by quiet hours.
How the “Off”, “Test” and “Live” modes work
The lowest of three modes applies: the global “Mode for all types”, the action type and the channel. An action type in “Live” with the global mode at “Test” works as a test. A main switch that is off or a policy that isn't applied means “Off” for everything.
| Mode | What AI does | What the customer gets |
|---|---|---|
| “Off” | offers nothing; chats are handled by the widget's AI assistant as before | the same as without autonomous support |
| “Test” | runs all checks and records “would do” in the ticket feed and in the “Review” queue | nothing from AI |
| “Live” | acts on its own if trust and the “Controller” allow it; otherwise waits for a person in “Pulse” | a reply, a “Yes/No” question, an action summary |
A test shows exactly what AI would do without a person. That's why for auto-replies the test records “would do” only with “Act on its own” trust: without it, the reply would have waited for an employee anyway.
| What | Where | Note |
|---|---|---|
| All action types together | “Policy” → “Mode for all types” | the ceiling for all other modes |
| Order action type | “Policy” → action type card, or “Actions and trust” | a money-moving type goes “Live” only after a backtest (review) |
| “Auto-reply to an email” | “Actions and trust” | “As now” — sent based on “Pulse” trust, as before, with no test and no breaker |
| “Messengers”, “Website chat” | “Channels and proactive” → “Auto-replies in chats” | work together with the channel's “By support policy” mode (details) |
| Proactive notices, questions to staff | “Channels and proactive” | follow the main switch and the global mode |
| Reviews, spikes, AutoQA | “Channels and proactive” | send nothing to customers, so they work even without the switch |
A mode changed on the “Actions and trust” tab is saved right away as a new policy version: “Mode saved — version N”. Next to it you can see which mode actually applies, taking the switch and the global mode into account.
1
Mode chips2
Act on its own3
Verification4
Breaker- Mode chips — switch a kind between off, test and live; the email auto-reply also has “As now”.
- Act on its own — the trust switch for the kind; not touched in this shot.
- Verification — shows whether the kind has passed a backtest on past tickets.
- Breaker — shows whether the kind has been tripped.
How to describe your store rules as text
You can write the rules in your own words, and AI fills in the form. The system executes the form after “Apply”, not the text.
- “Policy” → “Policy as text”: describe the rules, for example “We cancel unpaid orders up to UAH 3,000, no more than 5 per hour. Returns — 14 days after receipt, up to UAH 2,000.”
- Click “Parse”. AI fills in the fields of the action type cards and shows the “How I understood it” block.
- Check the “Not recognized — will not be applied” and “Contradictions in the text” blocks, and fix the fields by hand.
- Click “Apply”. Nothing has changed until then.
Modes can't be turned on through text: “Test” and “Live” are set only with buttons. Numbers above the system ceiling are cut down when the policy is applied, and a toast shows what was cut. The ceiling is shown in the field's tooltip.
| Field | Default | Maximum |
|---|---|---|
| “Order total up to, UAH” | cancellation 3,000, return 2,000, reshipment 1,500 | 50,000 |
| “Per hour, at most” | cancellation 5 | 100 |
| “Per day, at most” | cancellation 30, branch change 20 | 500 |
| “Days after receipt” | 14 | 60 |
| “Per customer in 90 days” | return 2, reshipment 1 | 10 |
Limits are counted by the system from the action log, not by AI. The fields of each action type and how AI performs the action are covered in the article on order actions.
Conversation procedures and processes from text
Some rules don't fit the form — they're conversation steps or actions triggered by an event.
- “Conversation procedures” — steps the AI responder follows with the customer, one per line: up to 20 steps of 500 characters each. For example, “If a customer reports a defect, ask for a photo first”. They take effect only after “Apply” and don't change limits.
- “Events → processes” — parts of the text such as “notify the warehouse” or “escalation”. The “Create process (disabled)” button builds a Flows process; turn it on in the “Processes from the policy” table once you've checked it.
Versions and rollback
Every “Apply” and every mode change on “Actions and trust” creates a new version in the “Policy versions” table: number, when, who. “Roll back” makes the selected version current; the current settings stay in the history, and nothing is deleted. The last 50 versions are kept.
When AI acts on its own and when it waits for a person
“Live” mode lets AI act, but each decision still goes through a series of checks. The first one that fails stops the automation, and the decision waits for an employee in “Pulse” or is handed to a person in the ticket.
- The action type and channel are in “Live”, and the breaker hasn't tripped.
- The “AI · Support” AI employee is on and within budget.
- The order belongs to the ticket's customer and isn't from a marketplace, unless that marketplace is in “Order sources”.
- The customer's identity is confirmed, or there is a channel to send a code.
- Action conditions: no , stock not written off, the status allows it.
- Policy limits aren't used up, and the same action hasn't been performed yet.
- The reason isn't “human only”: legal matters, threats, court.
- A money-moving action type has passed a backtest or has the “Enable without verification” checkbox ticked.
- In “Pulse”, “Act on its own” trust is on for the action type, and the “Controller” let the decision through.
You turn on trust with the “Act on its own” toggle in the “Trust” column of the “Actions and trust” tab, or on the “Trust” tab in “Pulse”. For order actions, the toggle can be turned on only when the action type is in “Live” and verified.
Replies to customers have three more “Pulse” restrictions: AI doesn't reply on its own if it lacked information, if its confidence is below 80%, or if the reason is a return, a complaint or another “human only” reason. When AI declines, it gives the customer a short reason without thresholds or limits.
How AI asks an employee
When the answer isn't in the articles, orders or products, AI can ask the warehouse or a manager instead of saying “I'll check”. The customer immediately gets “I am checking with my colleagues and will reply by 15:30”, and the employee gets the question in the ticket card.
Where: “Channels and proactive” → “AI questions to staff”.
| Field | What it does | Default |
|---|---|---|
| “Mode” | “Test” — the question is only recorded, nothing reaches colleagues or the customer | “Off” |
| “Warehouse”, “Manager”: “Employee” or “Department” | whom AI can ask; an employee takes priority over a department, and without an addressee AI doesn't ask that target | not selected |
| “Answer time, working minutes” | counted by the “Tickets” schedule; the customer sees this time in the message | 120 (from 10 to 1440) |
| “Questions per ticket per day” | how many times AI can ask colleagues within one ticket | 2 |
The addressee gets the notification “AI asks in ticket #N”, and the question shows up in “My Day”. The ticket gets an “AI asks” card: the question, to whom, “Reply by”. Pick an option or write your own answer (up to 2,000 characters) and click “Answer”. AI then prepares the customer reply as usual — through trust and the “Controller”.
1
Question2
To whom and deadline3
Answer options4
Answer button- Question — what AI asks the colleague.
- To whom and deadline — the department asked and the time to answer by.
- Answer options — ready-made replies in one click.
- Answer button — sends the reply (not pressed).
- The first to answer wins. A question that has already been answered is closed for everyone else.
- One question at a time. While the ticket is waiting for a colleague, AI doesn't ask a new one.
- An employee replying to the customer closes the question: “AI question closed — an employee replied to the customer”.
If the deadline passes, the question goes to the manager of the person who was asked: the ticket gets “High” priority, and the customer gets “Sorry for the delay — still checking with my colleagues, I will reply by …” (not during quiet hours). If the second deadline passes too, AI stops waiting and an employee handles the ticket.
How to review AI decisions
Control works even before the first real action: test mode, the review queue, automatic trust downgrade and the breaker.
The “Review” queue
The “Review” tab collects AI decisions that a person should look at. In “Tickets”, you open it with the “AI decision review” icon with a counter.
- the first 20 decisions of each action type — all of them, then a 20% sample, but at least one per day;
- always — decisions with confidence below 90%, test decisions, decisions undone by a person;
- always — when the customer replied negatively after the action, reopened the ticket or rated it “Bad”.
Open a decision on the left; on the right: “What the customer wanted”, “What AI did”, “Text to the customer”, “Check trace” and “Controller”. Rate it 👍 or 👎; for 👎, pick a reason: “Wrong action”, “Should not have acted”, “Wrong text”, “Tone”, “Other”. Set the “Unsafe — could harm the customer or the shop” flag separately, then “Save rating”.
1
Counters2
Decision list3
Check trace4
Rating- Counters — decisions to review, share of “Good”, “Bad” and unsafe ratings, and test decisions.
- Decision list — decisions worth a human check; clicking a row only opens it.
- Check trace — the checks the decision passed: mode, owner, identity, limits, duplicates.
- Rating — Good and Bad buttons with a comment field; nothing was saved.
Automatic downgrade and the breaker
An action type switches to “Suggest” by itself — AI no longer acts, it only suggests. This happens when:
- among the last 20 rated decisions of the action type, at least 2 are “Bad” or at least one is “Unsafe”;
- 3 automatic executions of the action type in a row within 10 minutes ended with an error;
- AI performed more than 40 actions of all types on its own within an hour.
“Management” gets the notification “Breaker: “…” now suggests only”. To restore the action type, click “Resume” in the “Breaker” column of the “Actions and trust” tab. Trust then stays at “Suggest”: “Act on its own” is turned on separately.
Testing and backtests on past tickets
A backtest shows what AI would have done with orders in tickets that are already closed, and whether that matches what people did. Nothing is sent to customers, and orders aren't changed.
- “Test and replay”: set “Period, days” (up to 90), “Reason” and “Tickets, at most” (50 by default, up to 200).
- Click “Run”. The “Backtest estimate” window shows the number of tickets, AI responder calls and the estimated cost. Confirm.
- The backtest runs in the background, and you can follow its progress in the Operations center; “Stop” cancels it.
Result: “Match”, “AI would act, human did not”, “Human acted, AI did not”, “Gate refusals”, refusal reasons and the “Differences from people” table.
1
Parameters2
Run button3
Match4
Differences- Parameters — period, reason and the maximum number of tickets.
- Run button — starts a new backtest (not pressed).
- Match — share of tickets where AI and people agreed.
- Differences — tickets where AI would act and the human did not.
Money-moving action types (returns, reshipment, cancelling a paid order with a draft payment order) go “Live” only after a passed backtest: 30+ decisions, 90%+ match, no unsafe ones. Otherwise only with the “Enable without verification — I understand the risk” checkbox. The latest backtest result is in the “Verification” column of “Actions and trust”.
How to undo an AI action or take over a ticket
Everything AI did in a ticket is shown in the “AI actions” block on the right: the action, its state, and the reason if AI handed the request to a person. Each step is also written to the ticket feed.
- “Undo action” reverts a completed action: restores the previous status and reservation and deletes the unposted draft. The undone decision goes to “Review” as “Bad”.
- “Take over” ends AI's waiting in the ticket and pauses AI in the customer's chats — an employee takes it from there.
- A regular employee reply to the customer also ends AI's waiting: the customer won't get a parallel question.
Action states and when undo isn't possible are covered in the article on AI order actions.
Scenarios
The first week: see what AI would do
Apply the policy, set the global mode and all action types to “Test”, turn on the switch and “Act on its own” trust for email auto-replies. Go through “Review” every day: 👍 where you'd agree, 👎 with a reason where you wouldn't. Customers get nothing from AI during this week.
Cancelling unpaid orders without a manager
“Order cancellation” is in “Live”, “If paid” is “Hand over to a person”. The customer writes “please cancel”, AI asks “Yes/No” and, after “Yes”, cancels the order if it has no waybill yet. Paid orders, orders picked by the warehouse or with a waybill go to a person, with the reason in “AI actions”. More: order actions.
AI doesn't know when an item will arrive
The “Warehouse” addressee is the stock keeper. The customer asks about a delivery date that isn't in the articles or on the product. AI writes “I am checking with my colleagues and will reply by 15:30”, the stock keeper picks an answer in the “AI asks” card — and the customer gets it without a manager getting involved.
Something went wrong
Turn off “AI may act by itself under the policy”: new actions stop immediately, and pending ones are paused and handed to people. Find the decisions in “Review”, undo the actions in the tickets, fix the policy or roll back a version — and turn it on again starting with “Test”.
What to do when AI doesn't act on its own
| What you see | Why | What to do |
|---|---|---|
| The mode is “Live”, but AI does nothing | The main switch is off, the policy isn't applied, or the global mode is lower | Check the “Mode” column on “Actions and trust”: it shows the mode that actually applies (modes) |
| AI decisions are waiting in “Pulse” | No “Act on its own” trust, the “Controller” stopped the decision, confidence is below 80%, or the reason is “human only” | Turn on trust for the action type; handle the remaining decisions manually (when AI acts on its own) |
| “First switch this action to Live in Autonomous support.” | The action type isn't in “Live” mode | Set the action type to “Live”, then turn on trust |
| “Live cannot be enabled without a passed backtest” | A money-moving action type without a backtest on past tickets | Run a backtest on past tickets or tick “Enable without verification” (backtest) |
| The notification “Breaker: … now suggests only” | Errors in a row, the hourly limit, or “Bad” or “Unsafe” ratings | Look at “Review”, fix the cause, click “Resume” on “Actions and trust” (breaker) |
| “The “AI · Support” employee is disabled” or “The daily budget of “AI · Support” is exhausted” | The AI employee is turned off or has used up its daily budget | Turn it on or raise the budget in “Pulse” (AI employees) |
| The “AI actions” block shows “Paused: autonomy is off” | The main switch was turned off while AI was waiting for the customer's reply | Finish the customer's request yourself: turning the switch back on doesn't lift the pause |
FAQ
Can AI refund a customer?
No. For a paid order, AI creates an unposted draft cash payment order, and a return as an unposted document. An employee posts them.
Can I let AI act only in messengers and answer emails myself?
Yes. The mode is set separately for emails, messengers and website chat; the lowest of the modes applies (modes).
I set a limit of UAH 100,000 in the policy — why does the field show 50,000?
That's the system ceiling: a number above it is cut down when the policy is applied, and a toast tells you what was cut (ceiling).
Does the customer see that AI is replying?
Yes, if “Sign AI messages to customers” is on and the signature text isn't empty: the message is signed “Answered by the store's AI assistant”.
What happens to requests waiting for “Yes” if I turn off autonomous mode?
They are paused and handed to a person; the customer's “Yes” then goes to an employee instead of being executed automatically.
Who can change the policy?
Users who have access to the “Autonomous support” window — only “Management” by default. “Act on its own” trust is turned on by someone with the right to the “Trust” tab in “Pulse”.
Will AI reply to a customer at night?
Yes, if the customer wrote first: quiet hours apply only to messages AI writes first (quiet hours).
How do I restore previous policy settings?
“Policy” → “Policy versions” → “Roll back” next to the version you need. The current version stays in the history.
Where to find it and who has access
The window has no menu item: you open it from “Tickets”.
- From “Tickets”
- the gear above the queue list → “Autonomous support”
- Review queue
- the “AI decision review” icon with a counter next to the gear
- In a ticket
- the “AI actions” and “AI asks” blocks on the right
- Access right
- the “Autonomous support” window — only “Management” by default
- “Act on its own” trust
- the right to the “Trust” tab in “Pulse”
- Prepare
- the “AI · Support” AI employee turned on, the company profile, articles about delivery and returns
The window has five tabs: “Policy”, “Actions and trust”, “Review”, “Test and replay”, “Channels and proactive”.
1
Tab row2
Main switch3
Mode for all types4
Policy as text- Tab row — Policy, Actions and trust, Review, Test and replay, Channels and proactive.
- Main switch — lets AI act by itself under the policy.
- Mode for all types — one chip sets the mode for every action type at once.
- Policy as text — the rules in plain words that AI follows.
Related topics
- To set up cancellations, returns or branch changes — AI order actions.
- If AI should reply in Telegram, Viber or website chat — Auto-replies in chats and messengers.
- If customers should learn about their parcel without asking — Proactive notices.
- If you need to score all conversations, not just AI decisions — Conversation quality and topic spikes.
- If AI should only assist employees — AI assistant in tickets.
