YundaDeskThe easiest number to report after launching proactive outreach is “how many messages we sent.” It is easy to pull and easy to misread. A high send count might mean the rule catches the right moment. It might also mean the condition is too broad, the guardrails are not doing enough, and customers are being nudged too often.
The better question is: why should this message go out, did it actually reach the customer, did the customer engage, and can a later order, lead, or human conversation reasonably give this outreach some credit? Measuring proactive outreach should start with delivery proof, not a conversion screenshot.
Start with an identity for every outreach event
Every proactive message needs a traceable identity. At minimum, capture rule ID, customer ID, channel, trigger reason, run mode, guardrail result, and message version.
Measuring Outreach Impact: turn timing into a measurable signal
| Field | Why it matters |
|---|---|
| Rule ID | Ties impact back to a specific rule |
| Customer ID | Connects identities across channels |
| Channel | Compares website widget, email, WhatsApp, Instagram, and other entries |
| Run mode | Separates observe-only, confirm-each-one, and auto-send |
| Guardrail result | Shows what cooldown, caps, quiet hours, and do-not-disturb blocked |
YundaDesk outreach is not a standalone pop-up. Website widget, custom API, email, WhatsApp, Telegram, Messenger, Instagram, TikTok, LINE, WeChat, VKontakte, Zalo, and YouTube all roll into one workspace and one customer profile, so attribution does not get split apart by channel. For more on context, see What is an omnichannel inbox?.
Delivery proof: do not treat sent as delivered
Break the outreach path into four steps:
- Would trigger: the rule matched and would have sent.
- Eligible to send: the six guardrails did not block it.
- Sent: the system sent it, or a human approved and sent it.
- Delivered / accepted by channel: the channel returned a delivery, accepted, or handoff status; if a channel has no complete receipt, mark it as “no delivery receipt.”
Do not treat “sent” as “delivered.” Email may land in spam, social messages may be limited by platform behavior, and quiet hours may correctly prevent sending. Delivery proof tells you whether the issue is trigger quality, guardrail pressure, channel delivery, or a message that arrived but did not earn a reply.
Read the funnel, not just the conversion
A proactive message should be measured through at least five layers:
| Layer | Metric | What it shows |
|---|---|---|
| Trigger | Would-trigger count | Whether the condition is too broad or narrow |
| Send | Actual sends / would-trigger count | Whether guardrails look reasonable |
| Delivery | Delivery-proof count / actual sends | Whether the channel is stable |
| Engagement | Replies / delivery-proof count | Whether timing and copy helped |
| Conversion | Orders, leads, continued cart activity, or follow-up conversations | Whether it moved a business action |
The engagement layer matters. In cross-border ecommerce, a reply like “which size should I choose,” “when will it arrive,” or “can I change the address” often proves the outreach caught a real hesitation. The AI agent can answer first from the knowledge base, then hand off when it cannot answer, when the customer asks for a person, or when the issue is high risk. That is AI answers first, humans back up.
Attribution windows: do not let credit roll forever
If a customer receives a message today and orders three weeks later, should the rule get credit? Without a window, every rule eventually looks impressive.
Use simple windows:
- Immediate inquiry: reply or continued conversation within 24 hours counts as contribution.
- Browse / add-to-cart: return visit, more cart activity, or order within 24 to 72 hours counts as weak contribution.
- Higher-consideration purchase: up to 7 days can count, but only as an assisting touch.
Use levels too: direct attribution, assisted attribution, and no attribution. This avoids claiming natural demand as outreach impact while still counting messages that helped a customer clear a real doubt before a human took over.
Control groups: use observe-only as a baseline
A heavy experiment is not always necessary. Observe-only mode is a clean baseline: the rule still decides what it would send, but sends nothing. You can see:
- how often the rule would have triggered;
- which countries, languages, time zones, and channels were included;
- what the message would have said;
- how those customers behaved naturally without receiving it.
Then move the rule to “confirm each one” or “auto-send” and compare similar customers, periods, and channels. This is more reliable than a before-and-after chart, because campaigns, discounts, seasonality, and traffic mix all move conversion rates.
Count guardrail blocks as part of impact
Outreach quality is partly measured by what did not get sent. Review each rule with a guardrail table:
| Guardrail | Heavy blocking may mean |
|---|---|
| Cooldown | The same customer keeps matching; the condition may be broad |
| Frequency cap | Multiple rules compete for the same customers |
| Quiet hours | Target-market time zones or send windows need work |
| Active chat protection | The customer already has context; do not interrupt |
| Do-not-disturb list | Segmentation or copy may cross a boundary |
| Human approval for sensitive actions | Refunds, compensation, and price changes need approval |
Hold the last line firmly: refunds, compensation, and price changes always require human approval and audit. AI can remind, collect context, and prepare the conversation for an agent. It should not execute money-moving actions by itself.
Review actions: keep good rules and stop noise
Weekly outreach review should not become a leaderboard for send volume. For each rule, choose one action:
- Keep: delivery is stable, engagement is healthy, and direct or assisted conversion is visible.
- Tighten: would-trigger volume is too high, guardrails block too much, or replies show irritation.
- Rewrite: delivery is fine, but customers do not reply or keep asking the same unclear question.
- Pause: no delivery proof, no replies, no downstream action, and the rule consumes frequency allowance.
This is where YundaDesk’s “gets smarter over time” loop matters. Questions the AI missed, agent follow-ups, and agent corrections can become learning suggestions you confirm. Only after the owner approves them do they become skills, knowledge, or customer memory. Every entry is traceable end to end, testable, and revertible in one click. Learning never takes effect automatically. For the full loop, see How to teach AI support that gets smarter.
Outreach impact is not “how proactive the AI was.” It is whether the customer was picked up at the right moment. Start with delivery proof, attribute through trigger, send, delivery, engagement, and conversion, then count the interruptions the guardrails prevented. That is how you tell whether a rule is helping or just adding noise.