Chatbot Launch Checklist

The Week Before the Switch, the Hour After It, and What to Have Ready to Turn It Back

Quick answer: A chatbot launch is a week, not a click. Seven days out, you are waiting on gates you do not control: Meta's business verification decision, which its documentation says may take up to 14 business days; a new WhatsApp business portfolio's limit of 250 unique recipients per moving 24 hours for messages sent outside a customer service window; an Instagram professional account with Allow access to messages switched on in the Instagram app, which Manychat's troubleshooting page lists among the first checks when an account will not connect; in the US, SMS campaign registration, which our 10DLC guide calls the gate on any automated text reaching a phone; and WhatsApp template review, which Meta says can take up to 24 hours. The day before, you run a twelve-item switch list, and the last item is the one that matters: you write down where the off switch is and who holds it. Launch day, you open the bot to a first phase small enough to read every transcript, on the controls the platform gives you (Intercom's audience rules, Tidio's page and country visibility, Manychat's per-automation Set Live). In the first hour you read four numbers; in the first week you read them daily; and you pull the bot back to the previous phase, not to zero, on any of three conditions.

This guide starts where testing ends. The ten tests and the 3.5-hour baseline are the QA testing protocol; if you have not run them, run them first, because nothing below substitutes for them. What a phase is, and which platforms can cut one by page, audience, channel or hour, is the same-day phased rollout entry. The build itself is the how to build a chatbot guide.

Why launch is a separate job from testing

Testing answers one question: does the bot do what the flow says, for the inputs we thought of? Launch answers three others. Will the channel let the bot speak at all (Meta, Instagram and US carriers each have a gate in front of that)? Will a person be there when the bot hands off? And if the answer to anything is no at 9:40 on launch morning, how fast does the bot disappear again? The QA protocol cannot answer any of the three, because they are about the world outside the builder. That is the checklist below.

T-7: the gates outside your control

Start these a week out because the slowest of them takes longer than a week. Each row's lead time is the one the vendor publishes, cited in Sources or in the linked page that read it.

Two things about the table are commonly read backwards. WhatsApp verification is not a gate in front of sending; on our reading Meta lets a new portfolio message from day one, capped at 250 unique recipients a day for messages delivered outside a customer service window, and messages inside that window are not counted, so on our reading an inbound-only bot can launch while verification runs. And the Instagram toggle is the cheapest item on the list and the one that produces the most confusing failure, because Manychat's own troubleshooting page puts it in the first two checks for an account that fails to connect, right after the professional-account requirement: "toggle Allow access to messages on."

T-1: the switch list

Run the day before, in this order, and initial each line. Twelve items; our estimate is roughly ninety minutes.

1. The default answer is live. On Manychat, the Default Reply is set per channel and the automation behind it must be published; the help page says to look for "a small yellow dot next to the automation title, which indicates there are unpublished changes that need to be addressed." On any platform, send the bot a message it cannot possibly match and confirm the fallback answer, not silence, comes back. The wording of that answer is the fallback copywriting guide.
2. The greeting matches the page it is on. A greeting written for the pricing page reads oddly on the returns page. Check it on every page in the first phase.
3. Handoff has a named person and a promised time. The handoff rules route to a queue; the queue needs someone in it during the hours the bot runs. Write the name and the target (say, under five minutes during business hours) on the checklist. Intercom's help center notes that a team-only phase "will also allow you to check that your handover rules are running as you intended."
4. After-hours behavior is decided. If nobody is in the inbox at 22:00, the bot must say so and take a contact, not promise a person. Our handoff design guide records how thin operating-hours features are across our corpus, so this is often a flow branch you build, not a setting you flip.
5. The first phase is set on the platform's own controls. Intercom: the Who will see Fin audience rules. Tidio: Visibility and position under Settings, Live Chat, Appearance (pages by Exact or Contains, trusted domains, countries by IP; page rules and the Display toggle per device). Manychat: one automation Set Live, other channels unconnected. The phased rollout entry has the mechanics; pick the phase now, not on the morning.
6. The golden run is fresh. Re-run the golden dataset against the exact build you are launching, and keep the result; it is the baseline every later number is read against.
7. Analytics fire. Open a test conversation and find it in the platform's reporting within the hour. A launch you cannot measure is a launch you cannot pull back on evidence. The metrics guide covers fallback and containment; the other two numbers in the first-hour table are this guide's own.
8. The privacy line is on the page. Whatever your data privacy guide checklist says the widget must disclose, it is disclosed on every page in the phase, and the retention setting is the one you decided on, not the vendor default.
9. Templates are approved, not pending. Every WhatsApp template the launch flow sends shows an approved status, in the language the flow sends it in.
10. Team notifications reach a phone. Whoever holds the handoff queue gets the ping on a device they carry, tested with a real handoff, not assumed from the setting.
11. The rollback is written down. Where is the off switch, who can reach it, and how long does it take? Intercom: narrow the Who will see Fin audience back to your team. Tidio: the Display toggle in Visibility and position, per device, or narrow the page rules. Manychat: toggle the Default Reply off in its settings bar (the help page's own off-and-on switch), or replace it with the x in Settings, selecting the automation that ran before. Website script tags: remove the snippet, and know how long your CDN or tag manager takes to propagate. Write the answer on the checklist in one line per channel.
12. The previous state is one switch away, not a rebuild. If the launch replaces an old bot, a canned autoresponder, or a human-only inbox, that older state still exists and can be restored without rebuilding it. Our vendor migration playbook runs the two in parallel for exactly this reason.

Day 0: the first phase and the first hour

Open the phase you set on item 5, at a time when the handoff person from item 3 is at their desk and has nothing else booked for two hours. Not Friday afternoon; not the morning of a promotion. Then read every transcript for one hour. At a few hundred conversations a day site-wide, an hour is a dozen or two conversations, and a narrow phase is a fraction of that, which is readable, and reading them all is the point of a narrow phase.

Four numbers at the end of the hour, against the golden baseline from item 6:

The last row is the one platform reporting rarely names, and it is the launch-specific failure: the gate table's failures (toggle off, template pending, campaign unregistered) all present as a customer who wrote and got nothing. Find it by reading transcripts; dashboards rarely surface it.

Days 1 to 7: widen or pull back

Read the same four numbers once a day for a week before widening anything, because the question mix changes by weekday and a Tuesday launch has not met Saturday's questions until Saturday. Widen one dial at a time (pages, then audience, then channel, then hours) so a changed number has one cause. Add answers for the gaps the transcripts show, and re-run the golden set after each change; that discipline is the regression testing guide.

Pull back, using the rollback from item 11, on any of three conditions, and fix the cause before repeating the same phase: fallback more than a third above baseline for a full day, a handoff that went unanswered past the promised time, or any silent conversation you cannot explain. The pull-back goes to the previous phase, not to zero. A bot that worked for the support page still works for the support page; the failure is in what you widened to. And never widen to fix a number: a wider audience multiplies a broken flow, it does not dilute it.

At day 7, with the numbers holding, the launch is over and the operating cadence begins: the QA protocol's weekly check of failed conversations, its monthly re-run of the three core tests, and its quarterly full re-run.

How three reviewed platforms expose the switch

As read on the vendors' help pages on 21 September 2026; none of it was exercised under live traffic for this guide.

Intercom puts the phase and the switch in the same place. Fin's audience is set in the Who will see Fin section of the chat and email deploy screens, and the help center describes a team-only phase filtered on your email domain and, separately, recommends "rolling Fin AI Agent out in stages by customer segment," starting with lower-tiered customers or a region. It also notes that a live test is billed per resolution, with an exception for conversations where the tester asked to "Get more help" or to speak to the team. Our Intercom review records the per-resolution meter.

Tidio's visibility settings carry no audience rule, so the phase is the widget's visibility: pages by Exact or Contains string, trusted domains, countries by IP, and a Display toggle, with the page rules and the Display toggle set separately for desktop and mobile. Showing the widget only on named pages, rather than hiding it on named ones, takes a script from Tidio's second help article. Below the widget, a Flow can be scoped on its own trigger: our review's hands-on session recorded "On certain days and time," "Visitor opens a specific page" and "First visit on site" in the Flows trigger picker, which is how an hours-limited or page-limited first phase is built on Tidio. The Tidio review measured 9 minutes from signup to a working widget and 14 with the Lyro AI agent active, and recorded that Lyro cannot be switched on until the widget is live on at least one site, which fixes the order: narrow widget first, AI second.

Manychat switches per automation. The Default Reply help page has you publish the automation with a Set Live button; the Default Reply is chosen per channel and can be replaced with an x in Settings; and the once-per-24-hours option on the Default Reply trigger throttles one person's frequency rather than limiting the audience. The Manychat review measured about 30 seconds from a blank canvas to a two-node skeleton and lists "live published flow execution to real production users" among the items its session did not exercise, which is a fair description of the gap between building and launching.

What our fifteen reviews record

Searched 21 September 2026, case-insensitively, across the fifteen files matched by sample-reviews/*-review.md: launch checklist appears in 0; go live or go-live in 0; set live in 0; soft launch in 0; publish in 15, mostly in the sense of what a vendor does or does not publish (pricing, compliance pages), with the builder's publish action and a video's publish date behind it. No review protocol launched a flow to real customers and watched the first hour; the reviews measure time to a working skeleton or a working widget, which is the moment before this guide starts. That gap is in our protocol, not in any platform, and the cheapest experiment that closes it is one platform, one narrow phase, one Tuesday morning, and the transcript log at noon. Corrections to editorial@chatbotscape.com.

FAQ

What should a chatbot launch checklist include?

The external gates and their lead times (WhatsApp verification and template review, Instagram's message-access toggle, US SMS registration), a switch list run the day before (default answer live, greeting per page, a named handoff person and promised time, after-hours behavior, the first phase set, a fresh golden run, analytics firing, the privacy line, templates approved, notifications tested, the rollback written down, the previous state one switch away), the first-hour reads, and the conditions that mean pull it back.

How long does it take to launch a chatbot?

The build can be minutes; the launch is a week. The bound is the slowest external gate: Meta says a WhatsApp business verification decision may take up to 14 business days, though a new portfolio can deliver to 250 unique recipients a day outside a customer service window while it runs; US SMS registration is the launch date for a text bot; Instagram and website widgets have no published waiting period. Plan the switch list for the day before and a week of daily reads after.

Should a chatbot be soft-launched?

Yes, to a phase small enough to read every transcript. Intercom's help center recommends staged rollout by customer segment; Tidio phases by page, domain and country; Manychat by automation and channel. The phased rollout entry covers what each can and cannot cut.

What are the signs a chatbot launch is going wrong?

Three, in the first hour or the first week: fallback more than a third above the pre-launch baseline for a full day (a breach in hour one is a warning), a handoff unanswered past the promised time, and any conversation where a customer wrote and the bot said nothing. The third is the launch-specific one, because every gate failure (toggle off, template pending, campaign unregistered) looks like silence to the customer.

How do you roll back a chatbot launch?

To the previous phase, not to zero, using the switch you wrote down the day before: Intercom's audience rules, Tidio's Display toggle or page rules, Manychat's Default Reply toggle or replacement, or the site's script tag, allowing for tag-manager propagation. Fix the cause, re-run the golden set, and repeat the same phase before widening again.

Related guides

• Phased rollout — what a phase is, the four dials, and which platforms turn them.
• Chatbot QA testing protocol — the ten tests that come before this checklist.
• Chatbot regression testing — the re-run after every change made during the launch week.
• How to build a chatbot — the build this checklist assumes is finished.
• Verify your WhatsApp Business account — the slowest gate, documents and timings.
• SMS chatbot setup and 10DLC — why registration is the launch date for a US text bot.
• Chatbot handoff design — the after-hours branch you will probably have to build.
• Migrating vendors in 14 days — keeping the old state alive while the new one launches.

Sources

• Intercom Help, Live testing Fin with your team or customer segments (intercom.com/help/en/articles/7837527, dated 14 May 2026), read 21 September 2026: "It's a good idea to introduce your AI agent in stages and monitor the impact as you go"; the Who will see Fin section; the team-only example ("select 'Users' and use the filter 'Email contains…'"), which "will also allow you to check that your handover rules are running as you intended"; "We recommend rolling Fin AI Agent out in stages by customer segment"; "You'll need to have a live Messenger installed before you can test Fin on your website"; "If you test Fin in your live environment, Fin will be billed per resolution," with the "Get more help" exception.
• Tidio Help, Disable Tidio on specific pages (help.tidio.com/hc/en-us/articles/5454505438620, updated 17 July 2026), read 21 September 2026: Settings > Live Chat > Appearance, Visibility and position; Trusted domains; Country visibility "based on their IP addresses"; the Display toggle "separately for the desktop and the mobile version"; Page visibility with Exact and Contains rules.
• Tidio Help, Show Tidio Widget Only on Specific Pages (help.tidio.com/hc/en-us/articles/5463603462172, dated 25 October 2024), read 21 September 2026: "this process requires access to your website's code"; the whitelist script.
• Manychat Help, Default Reply in Manychat (help.manychat.com/hc/en-us/articles/14281159586588, updated 3 December 2025), read 21 September 2026: set per channel under Settings; "every time a user sends a message or limit it to once every 24 hours for the same user"; "click the Set Live button to publish the automation"; "a small yellow dot next to the automation title, which indicates there are unpublished changes that need to be addressed"; "click the x icon next to it in the Settings section" to replace it.
• Manychat Help, How to connect Instagram to Manychat (help.manychat.com/hc/en-us/articles/14281290924444, updated 27 August 2026) and Can't connect Instagram to Manychat (help.manychat.com/hc/en-us/articles/14281291094428, updated 4 December 2025), read 21 September 2026: "your Instagram account must be set to a professional account (Business or Creator)"; "Settings and activity → Messages and story replies → Message controls → Connected tools, and toggle Allow access to messages on."
• Meta for Developers, WhatsApp Business Platform, Messaging Limits (developers.facebook.com/documentation/business-messaging/whatsapp/messaging-limits), read 21 September 2026: "Newly created business portfolios have a messaging limit of 250"; the limit counts unique recipients "outside of a customer service window, within a moving 24-hour period," and limits "are calculated and set at the business portfolio level and are shared by all business phone numbers within a portfolio"; the scaling paths, after which Meta "analyzes your message quality" and approves or denies the raise to 2,000.
• Chatbotscape, WhatsApp chatbot tutorial (/academy/whatsapp-chatbot-tutorial) and WhatsApp message template (/glossary/whatsapp-message-template): the "up to 14 business days" verification figure and the "up to 24 hours" template-review figure, each read from Meta's pages on the dates those entries record.
• Chatbotscape review corpus (the fifteen platform reviews at /reviews), searched 21 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE 'launch checklist' sample-reviews/*-review.md returns 0; grep -liE 'go live|go-live' sample-reviews/*-review.md returns 0; grep -liE 'set live' sample-reviews/*-review.md returns 0; grep -liE 'soft launch' sample-reviews/*-review.md returns 0; grep -liE 'publish' sample-reviews/*-review.md returns 15. Passages cited: tidio-review.md line 194 (9 and 14 minutes; Lyro after the widget); manychat-review.md lines 536 and 749 (the 30-second skeleton; the items-not-exercised list); tidio-review.md lines 566 and 580 (the Flows trigger picker, captured during the 28–29 May 2026 hands-on session).
• Chatbotscape evaluation methodology. /methodology (continuously updated).

About this guide

Chatbotscape launched in 2026 as an independent review site for chatbot platforms. This guide is part of our SMB chatbot Academy and is written for the owner or operations lead of a small business whose chatbot has passed testing and is about to meet customers. It reads Intercom's, Tidio's, Manychat's and Meta's published help and developer pages as of 21 September 2026 and draws on our fifteen platform reviews for what each platform's builder measured and did not.

Methodology

Every platform mechanism and quotation was read on the vendor page or the review passage named in Sources and is quoted with its own wording; Meta's verification and template-review times are cited to the Chatbotscape pages that read Meta's documentation on the dates they record. The gate table, the twelve-item switch list, the four first-hour numbers, the thresholds and the pull-back conditions are editorial working recommendations derived from those documented rules and from the deployment patterns recorded across our reviews; they are not the output of a controlled study, and no bot was launched to real customers for this guide. The corpus counts are reproducible with the commands printed in Sources. Manychat, Tidio and Intercom carry affiliate links elsewhere on this site; the checklist does not depend on which platform a reader chooses.

Last updated

22 September 2026 — first published.