Playbooks · 6 min read
Who Checks the Automation on Tuesday: An Agency Rota
An agency's automations do not fail loudly. They fail on a Tuesday, in one workspace, quietly, and the client discovers it six weeks later when somebody asks why a comment from March never got a reply.
That is the shape of the problem, and it is why the interesting question for an agency is not which tool. It is who opens which screen, on which day, and what they write down afterwards. What actually breaks between one client and ten is a separate argument and worth reading first. This post is the rota that comes after you have accepted it.
Nothing pages you, so the calendar is the only mechanism
Say the honest part out loud before you design anything around it. There is no alerting here. No email when a token expires, no aggregate view across fifteen workspaces, no health score on a wallboard. A connection that has gone stale still shows the account name, and the account simply stops sending.
So the routine cannot be "we monitor it". It has to be "a named person opens this screen on this day". Monitoring implies a watcher. What you actually have is a schedule, and a schedule only works if the intervals are chosen deliberately rather than inherited from whatever your project tool defaults to.
Coverage is a promise you cannot keep. A cadence is a promise you can, and clients can tell the difference within a month.
The four cadences, and why they do not line up
The mistake is one weekly meeting where everything gets looked at. The four things that need checking move at genuinely different speeds, and forcing them onto one clock means three of them are checked at the wrong time.
- The publish cadence — triggered, not scheduled. Every time a client publishes a post that carries a comment-this-word line, somebody attaches an automation to it before the post goes up, and retires the automation for whatever that post replaced. This is not a weekly job. It happens at publish time or it happens never.
- The workspace pass — weekly, one afternoon. Open each workspace, read the blocked activity, check the connection reads healthy rather than merely present. A few minutes an account once you know the shape.
- The template read — monthly, out loud. Read every live reply end to end against current reality. Offers move, hours move, a service gets discontinued, the client hires a new number. Nothing in the product notices. A person reading the sentence does.
- The access review — quarterly, boring. Who has a seat in which workspace, and does every one of those people still work here. Removing a seat is nobody's job by default, which is exactly why it needs a date.
The publish cadence is the one agencies invent last and need first. An automation is attached to a post, so the moment a post exists is the moment the automation decision exists. Doing it a week later means the post ran for a week with a trigger word in the caption and nothing behind it.
What "checked" actually means
A rota with a vague verb decays into a tick. Define the pass as three questions with three answers, and the pass takes minutes instead of an afternoon.
Expected, write nothing
Write it down and act
The ten checks run in a fixed order before every send and a blocked reply is recorded with its reason, which means the blocked list is not a fault log. It is the only honest status report you have. What each check does and why it exists is worth one read by whoever owns the pass, once, properly.
One line per workspace, dated
The artefact matters more than the check. Without it, two account managers each assume the other did last week, and neither can prove otherwise.
The format that survives contact with a busy Thursday is one line per workspace in a shared doc: date, initials, and the one thing you would say to the client if they asked. "18 Sep, RS, all healthy, retired the Ganpati automation." That is enough. It gives you a history, it makes a gap visible, and when a client asks what you have been doing it is the answer.
Resist building a dashboard for this. There is no API to feed it, so the dashboard becomes a second thing to update by hand, and the second thing is the one that stops getting updated.
Credits are a diagnostic, not a bill
Templated replies are free and unlimited. A credit is spent only when the AI writes a new reply, one credit per reply. The free tier is 100 credits with no card, and packs start at ₹499.
The operational use of that number is not cost control. It is a smell test. A client whose credit use climbs month over month is usually a client whose automation has drifted into answering questions it should be routing to a human — or whose templates have stopped covering the questions people actually ask. Either way, the number moved before anybody complained, which makes it the closest thing to an early warning in the whole setup.
What the rota does not turn into
It does not become an SLA. You are reviewing on a schedule, not watching, and selling the second one will end badly the first time a token dies on a Friday night.
It does not become a report the tool generates. There is no analytics product here, no client-facing dashboard, no scheduled export. Leads leave as CSV and the summary is written by a person.
And it does not extend past Instagram. Official Graph API only, seven days to reply to a comment, twenty-four hours inside a DM thread, restarted only when they message again. No WhatsApp, no CRM sync, no publishing.
Start with the publish cadence, because it is the one that prevents work rather than finding it. The commercial side of all this is a different conversation with a different answer.
