15 min read

Chatbot Maintenance Cadence

What to Check Weekly, Monthly and Quarterly After Launch — and the Failures That Never Announce Themselves

Quick answer: A live chatbot needs a maintenance calendar, not just a test suite. Weekly, skim the channels for one specific failure mode: a channel that used to carry traffic and now carries none, with no error anywhere in the platform's UI — our own hands-on Manychat review found exactly this after a bot sat idle for roughly four months, the Telegram-side token registration drifted, and "Manychat surfaces no warning, no banner, no 'webhook health' indicator anywhere in the channel settings UI," while "the channel toggle continues to read 'Enabled.'" Monthly, check three things that decay quietly: whether the pages your bot answers from are still accurate, whether any WhatsApp template has slipped in quality rating, and whether the vendor changed the model underneath you without asking. Quarterly, read the drift, not just the day: has resolution rate trended down over ninety days even though no single day looked alarming. None of this is testing. Testing asks whether the bot answers correctly against a fixed set of questions; maintenance asks whether the world the bot operates in has moved out from under it.

What this guide is and is not

This guide owns the operating calendar for a live chatbot: what to check, how often, and why each check exists. It does not own testing procedure. The QA testing protocol owns the ten pre-launch tests and its own weekly failed-conversation read plus monthly and quarterly test re-runs. The regression testing guide owns the golden-set diff you run before and after every change, plus its own monthly full run and quarterly set review. Both of those are about whether the bot still passes a fixed test. This guide is about the things a passing test does not reveal: a channel gone silent, a page the bot quotes that the business changed six weeks ago, a WhatsApp template that quietly lost delivery, and drift in the underlying model or the questions people ask it. The same-day model drift entry owns the definition of that last one and what does and does not count as it. Intercom, Manychat, Chatbase and Tidio have Chatbotscape reviews that carry affiliate links; Meta and IBM do not and we have no commercial relationship with either; see our affiliate disclosure.

Why maintenance needs its own calendar

A test suite, however good, only asks questions you already thought to ask. The golden dataset behind the regression guide catches the moment a known question gets a new, wrong answer. It cannot catch a question nobody is asking anymore because the channel that used to receive it went quiet, and it cannot catch a page the bot correctly quotes verbatim that the business quietly changed. Those are not test failures. They are operational decay, and they need a calendar with an owner, not a test run with a pass or fail.

The Manychat finding above is the clearest example in our own corpus of why this matters: nothing failed loudly. The automation stayed published, the toggle stayed on, and the dashboard showed nothing except zero contacts and an empty inbox, which on a quiet week looks identical to a channel that is simply working and receiving no messages. A weekly test run would not have caught it, because there is no failing assertion to write against silence that looks like success.

The maintenance calendar

When	Check	Why it is not covered elsewhere	Owner
Weekly	Every connected channel still shows recent inbound traffic, or a documented reason it should not (a channel intentionally paused, a genuinely slow week)	The QA protocol's weekly read is failed conversations, which assumes conversations are still arriving. This check is upstream of that: are they arriving at all, on every channel	The bot owner
Weekly	Work off any vendor content-suggestion queue, if the platform has one	Intercom's own guidance calls this a weekly action; most platforms don't surface it, so on those, skip to the monthly pass	The bot owner
Monthly	Spot-check that the pages, prices and policies the bot quotes from still match what the business actually offers	Testing checks the bot against its own test set, not the test set against reality. If the business changed and nobody updated the source content, every test still passes and every answer is wrong	The bot owner or whoever owns the source pages
Monthly	Every WhatsApp template used by the bot still carries a Medium or High quality rating and is not paused	A template can fail silently between sends; nothing in a conversational test catches a delivery problem that only a WhatsApp Manager dashboard shows	Whoever manages the WhatsApp Business account
Monthly	Check the vendor's changelog or release notes for a model swap or feature change made without your action	The regression guide's monthly full run catches the symptom (a new wrong answer); this check looks for the cause before it shows up as one	The bot owner
Quarterly	Read resolution rate, fallback rate and escalation rate as a ninety-day trend, not a single day	A single bad day is noise; a quarter of gradual decline with no single alarming day is the signature of drift, and it is invisible to a dashboard built to flag daily spikes	The bot owner
Off-cycle, triggered	Any of: a product line launches, retires or reprices; the vendor announces a model change; a template gets paused; resolution rate on any content item crosses below 90 percent	These are the events that make the calendar above insufficient on its own — they need a check the moment they happen, not at the next scheduled date	The bot owner

The four kinds of decay a testing protocol does not catch

Content going stale. The bot answers exactly what it was built to answer, correctly, against source content the business no longer stands behind. A pricing page changes, a product is discontinued, a policy is updated on the website but not in the bot's knowledge source. Nothing about the bot broke; the ground under it moved. This is the reason the monthly content spot-check exists, and it is also the argument for keeping the knowledge base synced to a single source of truth rather than a copy the bot answers from and the website answers from separately.

Silent technical failures. A channel connection can die without any of the usual signals — an error banner, a failed-status badge, a support ticket from an angry customer, because the customer who would have complained never got a bot to complain about. Our Manychat review's finding is the concrete case: after roughly four months idle, "the Telegram-side token registration drifts, the Manychat-side webhook receives no events," and the operator's entire signal is "zero contacts, empty Inbox, no error message." A webhook that goes quiet looks, from the dashboard, identical to a channel with no traffic that week. The only defense is a weekly habit of checking that every connected channel is still receiving something, not trusting that the absence of an error means the absence of a problem.

Template quality decay. WhatsApp grades every message template on a rolling quality rating, and the grading is not a one-time approval. Meta's own documentation defines the three tiers plainly: a High (Green) template "has received little to no negative feedback from WhatsApp users"; Medium (Yellow) and Low (Red) both mean the template "has received negative feedback from multiple WhatsApp users, or low read-rates." A template that slides to Red is automatically paused, and Meta's pausing schedule escalates: "1st Instance: Paused for 3 hours," "2nd Instance: Paused for 6 hours," "3rd Instance: Disabled." Since Meta introduced template pacing, a paused template no longer resumes on its own when the clock runs out; it needs to be manually unpaused. A template that was fine at launch can be Yellow by the third month if its content stops matching what customers expect, and nothing about the chatbot itself changed. This is the reason for the monthly WhatsApp Manager check, and it is a different failure than the template approval process that the launch checklist and verification guide already cover, which is about getting a template approved once, not about it staying healthy afterward.

Model drift. The bot's classifier or language model gets quietly worse at understanding what customers mean, either because the vendor changed the model underneath the product or because the questions themselves shifted (new product lines, new slang, a new customer segment) faster than the bot's training or prompting kept up. This is the one form of decay that has a name in the wider machine-learning field, and it is covered in full, including what does and does not count as it, in the same-day model drift entry. The signal worth watching here, distinct from the individual content-item resolution rate mentioned below, is the trend: read fallback rate and intent confidence over a full quarter, not a single week, because drift by definition is gradual and a one-week read looks like noise.

What three platforms actually give you (and one that gives you nothing)

As read on the vendors' own pages on 22 September 2026.

Intercom hands you a weekly and monthly checklist, with a number attached. Its knowledge-management guidance lays out a cadence in the vendor's own words: weekly, "Review and action AI content suggestions" and "Review and act on teammate suggestions"; monthly, "Optimize content for AI." It also gives a concrete threshold for what counts as content worth fixing: "Flag items with a resolution rate below 90% as candidates for optimization," focused on "the top 20% of content by Fin involvement rate," meaning the content Fin leans on most, not the content that happens to be easiest to check. Fin itself contributes to the queue: "Fin will surface recommendations when it identifies content gaps based on conversations that required escalation." That is a maintenance workflow built into the product, and it is the most complete vendor-provided answer to this guide's question that we found among the platforms we cover.

Manychat gives you a troubleshooting checklist, not a health check. Its automation troubleshooting article is thorough about what to verify once something is already suspected broken: "The automation is published," "Make sure the trigger is turned on," "Try refreshing your permissions," and confirming "the automation does not contain empty cells, such as incomplete action blocks, data collection fields, or sequences." Every one of those is a real, useful check. None of them is proactive; the article exists to be read after a customer or an operator has already noticed a problem, not to surface one on its own. That gap is exactly what our Manychat review's Telegram finding fell into: none of the documented troubleshooting steps would have triggered, because nothing inside Manychat flagged that a problem existed to troubleshoot. This is the practical argument for treating "every channel still shows recent traffic" as a standing weekly check rather than something to look at only when a customer complains.

Chatbase gives retraining a dedicated button, not a schedule. Our review's screenshots of the Data Sources screen show a Retrain agent action sitting next to the file and text-snippet library, letting an operator push updated source content into the live agent after an edit. That answers the mechanical half of content maintenance (getting a change live) but not the calendar half (when to make one); Chatbase's own interface gives no cadence recommendation, so the monthly content check in the table above is the operator's to schedule, on Chatbase as on any platform without Intercom's built-in suggestion queue.

Tidio, on the pages we read for the same-day phased rollout and launch checklist entries, documents visibility and Flow-trigger settings but nothing framed as an ongoing maintenance or content-freshness workflow, so the calendar above applies to it exactly as written, with no platform-specific shortcut to note.

What our fifteen reviews record

Searched 22 September 2026, case-insensitively, across the fifteen files matched by sample-reviews/*-review.md: maintenance appears in 2; drift in 5, of which four are the reviews' own standard disclosure sentence about third-party rating aggregators ("Aggregator counts are dynamic and may drift between this review's publish date and reader's read date") and one is the Manychat finding quoted throughout this guide; stale in 3; resolution rate in 3; retrain in 1 (Chatbase, the Retrain agent button captured in a screenshot caption). No review protocol tracked a live bot's performance over months to observe drift directly; our reviews are point-in-time hands-on sessions, typically under an hour, which is the same evidence-class gap the launch checklist and phased rollout entries note about their own claims. The Manychat webhook finding is the one exception: it surfaced because the reviewer returned to a bot left running unattended for months, which is closer to real operating conditions than any single-session test, and it is the strongest evidence in our own corpus for why a maintenance calendar, not just a launch checklist, belongs on this site. Corrections to editorial@chatbotscape.com.

FAQ

How often should a chatbot be checked after launch?

Weekly for channel silence and any vendor content-suggestion queue; monthly for content accuracy, WhatsApp template quality, and vendor changelogs; quarterly for trend reads on resolution, fallback and escalation rates; and immediately after any product change, pricing change, vendor model swap, or template pause. The QA testing protocol and regression testing guide run their own, separate testing cadences alongside this one.

What is the difference between chatbot maintenance and chatbot testing?

Testing checks the bot against a fixed set of questions with known-correct answers. Maintenance checks whether the world the bot operates in — the content it quotes, the channels it runs on, the templates it sends, the model underneath it — has changed in a way no fixed test would catch. A bot can pass every regression test and still be maintenance-broken, quoting a discontinued product or sending on a channel that silently stopped connecting.

Can a chatbot fail without any error message?

Yes. Our hands-on Manychat review found a Telegram channel that had been disconnected for roughly four months with the automation still marked published, the channel toggle still reading "Enabled," and no warning anywhere in the settings UI — only zero contacts and an empty inbox, which looks identical to a quiet week. This is the specific failure mode the weekly channel-traffic check in this guide's calendar is built to catch.

How do I know if a WhatsApp template needs attention?

Check its quality rating in WhatsApp Manager. Meta rates every template High (Green), Medium (Yellow) or Low (Red) based on user feedback and read rates, and a template that reaches Low is automatically paused, for 3 hours on the first instance, 6 hours on the second, and disabled on the third; paused templates no longer self-resume and must be manually unpaused. A monthly check catches a sliding rating before it reaches a pause.

What is model drift and how is it different from a broken flow?

Model drift is the gradual decline in a chatbot's understanding as the model or the questions it faces change over time, not a single flow breaking. The full definition, what does and does not count, and how to spot it is covered in the model drift glossary entry.

Model drift — the definition, and what does and does not count as it.
Chatbot QA testing protocol — the ten pre-launch tests and their own testing cadence.
Chatbot regression testing — the golden-set diff run before and after every change.
Build a chatbot knowledge base — constructing the source content this guide's monthly check keeps honest.
Chatbot launch checklist — the week the bot goes live, before this calendar begins.
Verify your WhatsApp Business account — the one-time approval process behind the templates this guide checks monthly.
Reduce chatbot hallucinations — the weekly production read for made-up answers, a close cousin of the content-staleness check above.

Sources

IBM, What is model drift? (ibm.com/think/topics/model-drift), read 22 September 2026: "Model drift refers to the degradation of machine learning model performance due to changes in data or in the relationships between input and output variables"; "the accuracy of an AI model can degrade within days of deployment."
Intercom Help, Mastering knowledge management for great AI support (intercom.com/help/en/articles/11782981), read 22 September 2026: "Review and action AI content suggestions" and "Review and act on teammate suggestions" (weekly); "Optimize content for AI" (monthly); "Flag items with a resolution rate below 90% as candidates for optimization"; focus on "the top 20% of content by Fin involvement rate"; "Fin will surface recommendations when it identifies content gaps based on conversations that required escalation."
Manychat Help, General automation troubleshooting (help.manychat.com/hc/en-us/articles/14281479971996, updated 27 August 2026), read 22 September 2026: "The automation is published"; "Make sure the trigger is turned on"; "Try refreshing your permissions"; "The automation does not contain empty cells, such as incomplete action blocks, data collection fields, or sequences."
Meta for Developers, WhatsApp Business Platform, Template quality rating (developers.facebook.com/documentation/business-messaging/whatsapp/templates/template-quality), read 22 September 2026: High quality "has received little to no negative feedback from WhatsApp users"; Medium and Low quality both defined as having "received negative feedback from multiple WhatsApp users, or low read-rates."
Meta for Developers, WhatsApp Business Platform, Template pausing (developers.facebook.com/documentation/business-messaging/whatsapp/templates/template-pausing), read 22 September 2026: a template at Low quality "will automatically be paused"; "1st Instance: Paused for 3 hours"; "2nd Instance: Paused for 6 hours"; "3rd Instance: Disabled"; paused templates require manual unpausing since the introduction of Template Pacing.
Chatbotscape, Manychat review (/reviews/manychat-review, sample-reviews/manychat-review.md line 738), hands-on finding: "the Telegram-side token registration drifts, the Manychat-side webhook receives no events — but Manychat surfaces no warning, no banner, no 'webhook health' indicator anywhere in the channel settings UI. The channel toggle continues to read 'Enabled.' The operator sees: zero contacts, empty Inbox, no error message."
Chatbotscape, Chatbase review (/reviews/chatbase-review, sample-reviews/chatbase-review.md lines 492–500), screenshot captions: the Data Sources screen's Retrain agent action beside the file and text-snippet library, and the per-workspace storage quota shown alongside it.
Chatbotscape review corpus (the fifteen platform reviews at /reviews), searched 22 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE 'maintenance' sample-reviews/*-review.md returns 2; grep -liE 'drift' sample-reviews/*-review.md returns 5; grep -liE 'stale' sample-reviews/*-review.md returns 3; grep -liE 'resolution rate' sample-reviews/*-review.md returns 3; grep -liE 'retrain' sample-reviews/*-review.md returns 1.
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 business whose chatbot is already live. It reads Intercom's, Manychat's, Meta's and IBM's published pages as of 22 September 2026 and draws on our fifteen platform reviews, including a multi-month-old finding from live-account testing, for what a maintenance calendar needs to watch for that a launch-week checklist does not.

Methodology

Every vendor mechanism and quotation was read on the page named in Sources on 22 September 2026 and is quoted with its own wording. The maintenance calendar, the four kinds of decay, and the off-cycle triggers are editorial working recommendations derived from those documented rules, from the general machine-learning literature on model drift, and from the deployment patterns and one live-account finding recorded across our reviews; they are not the output of a controlled study, and we did not run a multi-month maintenance program ourselves for this guide. The corpus counts are reproducible with the commands printed in Sources. Intercom, Manychat, Chatbase and Tidio carry affiliate links elsewhere on this site; this guide's calendar does not depend on which platform a reader uses.

Last updated

23 September 2026 — first published.