YundaDeskThe same line — “I noticed you were looking at this one, anything I can help with?” — is a thoughtful follow-up at the right moment and a nuisance at the wrong one. Proactive outreach (letting the AI open the conversation at the right time instead of waiting for the customer to ask first) is a great tool for lift on both conversion and experience. It is also the feature most likely to backfire. Get the timing wrong and customers won’t complain — they’ll quietly block you, then leave a review that says “keeps popping up, annoying.”
So making proactive outreach work is never about “can it send” — it is about “how do we send with restraint.” The path below is the one we recommend every new rule walks through: teach the rule in one sentence, roll it out through three modes from observe-only to auto-send, and keep six guardrails you can’t switch off underneath the whole thing.
Teach a rule in one sentence: spell out the timing and the line
You don’t need to open any configuration panel or draw a flowchart. To have the AI speak up at some moment, just tell Yuna in plain language, for example:
If a customer hasn’t ordered 30 minutes after asking, check in with a simple “anything else I can help with?”
Yuna breaks that sentence into a structured outreach rule: the trigger (30 minutes of silence after an inquiry), the action (send a caring line), and a default run mode. It doesn’t go live right away — it produces a learning suggestion you confirm that lands in the owner’s review desk for your sign-off. This is the same controlled loop as every other way of teaching in “gets smarter over time”: learning never takes effect on its own, and every entry is traceable end to end, testable, and revertible in one click. For the full logic of teaching, see Gets smarter over time: how to teach your AI agent.
The point of teaching in one sentence is that you describe business intent, not technical plumbing. “If they add to cart but don’t check out within 10 minutes, remind them about free shipping over the threshold” — you just state the judgment you’d make on the shop floor, and the AI translates it into a rule.
Proactive outreach without annoying customers: turn timing into a measurable signal
Three modes, rolled out gradually: watch, then confirm, then let go
Between the moment a rule is written and the moment it actually auto-sends, there should be two buffers. We built that into three run modes, and we suggest running each for a week before moving on:
| Mode | What the AI does | What you see | When to use it |
|---|---|---|---|
| Observe only | Logs what it would send — sends nothing | A rehearsal log: the moment, the target, the intended line | The default start for any new rule; learn its judgment |
| Confirm each one | Drafts the line, waits for you to hit “send” | Every pending message surfaced for you to edit or block | Once the timing and wording both look right |
| Auto-send | Fires the moment conditions match, no human | Send records and reply rates you can audit afterward | After two clean weeks of rehearsal |
A new rule always starts in observe-only. It behaves like an intern taking notes beside you: it records “if this rule were live, I’d send this customer this line right now” exactly as it would happen, but doesn’t send a single word. Run it a week and the rehearsal log tells you everything — whether its timing is right, whether the wording flows, whether it wants to message someone at 3 a.m. Once you trust its judgment, promote it to “confirm each one,” and only then to “auto-send.”
Skipping the buffers and going straight to auto-send is putting an untested rule in front of real customers. Two extra weeks of rehearsal buys you out of firefighting in the reviews later.
Six guardrails: tune the thresholds, but you can’t turn them off
The real danger with proactive outreach is when “each rule is fine, but stacked together they carpet-bomb the customer.” That’s why six guardrails are always on and cannot be switched off — you can tune the thresholds, but you cannot remove the protection itself. This is a design red line, not an option.
- Cooldown: the same customer and the same rule fire at most once per 24 hours by default. Refreshing the page repeatedly doesn’t get them nagged repeatedly.
- Frequency cap: at most 2 proactive messages to the same customer within 7 days by default, counted across all rules together, so no single rule bombards on its own.
- Quiet hours: nothing sends between 22:00 and 08:00 in the customer’s local time zone. Cross-border customers are scattered across time zones, so this one matters — no one wants to be woken at 3 a.m. by “hey, you there?”
- Don’t interrupt an active chat: when a customer is mid-conversation with you (or a human agent), proactive rules step aside instead of cutting in.
- Do-not-disturb list: once a customer is on this list, every proactive rule skips them permanently. When a customer says “stop messaging me,” that means permanently.
- Sensitive actions need a human: any outreach that touches money (payment nudges, anything compensation-related) must go through human approval — the AI doesn’t send it on its own.
Write to help, not to chase
At the same trigger moment, the posture of the wording decides whether the customer feels grateful or resentful. The core difference is one thing: are you helping them, or chasing them?
- Chasing: “You haven’t paid yet~” — the subtext is “I’m watching your wallet,” and it just feels like surveillance.
- Helping: “Want me to confirm stock and delivery time for you?” — the subtext is “I don’t want you worrying, let me line up the facts.” Same moment, entirely different feeling.
A simple self-check when writing a line: imagine saying it to the customer’s face — does it sound like a sales pitch? If it does, rewrite it. The value of proactive outreach is to clear the customer’s hesitation (shipping cost, availability, whether they can change the address), not to remind them you’re waiting for their money. The former gets an order; the latter gets a closed tab.
One practical tip: for add-to-cart-without-checkout situations, instead of chasing payment, proactively put the information the customer might be stuck on (free-shipping threshold, stock, return policy) in front of them. Move the decision blockers out of the way, and conversion follows as a byproduct.
How to read the results: three numbers are enough
Whether a proactive rule is worth keeping doesn’t need fancy metrics — three numbers say it all:
- Would send: how many messages this rule wanted to send in observe-only mode. It reflects how often it fires — too high means the condition is too broad and may over-nudge; too low means it barely gets a chance and isn’t worth keeping.
- Actually sent: how many actually went out. The gap between this and “would send” is exactly what the six guardrails held back — that gap tells you how hard the rails are catching and whether a threshold should loosen.
- How many replied: how many customers wrote back after getting the message. This is the most direct signal of value — a reply means your outreach genuinely answered a doubt; messages that vanish into silence mean it’s time to rethink the timing or the wording.
Put those three side by side and it’s obvious whether a rule should stay, tighten, or be switched off. Don’t treat “how many we sent” as a scorecard — sending more isn’t the same as being useful. Reply rate is.
Triggers supported today: an honest boundary
The moments you can use to trigger proactive outreach today are all signals YundaDesk can sense on its own, without external data:
- Silence after an inquiry: the customer asks a few things and then goes quiet — follow up once they pass the duration you set.
- Page browsing / dwell / add-to-cart: a customer browsing a page that hosts your chat widget, lingering on it, or adding an item to the cart — these page behaviors can all serve as triggers.
To be clear: payment-nudge outreach for unpaid orders requires connecting your store platform (such as Shopify) first — because “who ordered, and whether they paid” lives on the store side, not the chat side, and has to sync over. That’s the real product boundary, not something we’re withholding: before you connect a store, you can start with page signals like “added to cart, no checkout” for a thoughtful nudge, and unlock precise order-status payment reminders once the store is connected.
How well proactive outreach works isn’t measured by how proactive the AI is, but by whether customers felt intruded on. Teaching in one sentence lets you state the business judgment; three modes give you two weeks of buffer; six guardrails hold the line for you. Put all three together, and when the AI speaks up first, it reads like a veteran employee who knows the boundaries — not a pop-up you can’t close.