Customer.io Routines: Three To Turn On This Week, One To Leave Paused

Customer.io shipped Routines on 8 April 2026 as part of its biggest AI release. The docs got another pass on 12 May 2026, when the plan limits and the auto-pause threshold settled. Seven weeks in, most accounts on Premium or Enterprise have either turned on every suggestion card and then stopped opening the daily emails, or treated the feature as a beta and ignored it.

A Routine is a recurring agent session. You write a prompt, set a schedule, and the agent runs it in safe mode (read-only) and emails you a summary each time. When you open Routines in the sidebar, the agent offers six suggestion cards.

Three of them earn their schedule on day one. One is dangerous on the default prompt because it floods you with a report nobody can act on. Two should sit paused until you've built a habit around the rest.

The three to turn on: deliverability watchdog daily, segment hygiene check weekly, and goal conversion report weekly. The one to pause: subscriber churn early warning. The two to skip for now: broadcast recap and audience growth tracker.

Here's the reasoning behind each call.

What a Routine is (and what it isn't)

A Routine is a scheduled agent session. Each run starts with no memory of previous runs, executes against your prompt using the same tools the agent has in a normal conversation, and emails you the result. It runs in safe mode, which is read-only—the agent can query and analyse, but it cannot create, edit, or delete campaigns, segments, newsletters, or workspace settings. The same guardrails you'd put on the agent itself still apply, just with the write paths closed.

Plan limits are per user, per workspace. Essentials gets 1 active routine with a weekly minimum interval. Premium and Enterprise get 5 active each with a daily minimum. Total routines, active or paused, are capped at 10 across all plans. A routine that fails 3 times in a row auto-pauses by default, and that threshold is configurable between 1 and 5.

Routine #1: Deliverability watchdog, daily

Deliverability earns a daily slot because it moves in hours, not weeks. Bounce rate, complaint rate, and domain reputation can shift inside a single send, and by the time you spot it on a Monday-morning dashboard, the bad send has already gone through.

The default suggestion prompt is broad. Rewrite it before saving and put specific numbers in. Gmail's postmaster guidance shifted in late 2025 and 0.10% is the new working ceiling for spam complaints, not the 0.30% headline number most teams still anchor to.

A tight prompt looks like this:

Each morning, check yesterday's bounce rate per sending domain. Flag anything above 2%. Check yesterday's spam complaint rate. Flag anything above 0.08%. Check our Microsoft and Gmail sender reputation scores. Flag any change from the prior day. Email me a one-paragraph summary. If nothing is flagged, say so in one line.

The signal-to-noise rule: if the routine emails "nothing to flag" three days in a row, the prompt is doing its job. If it sends a wall of text every day, your thresholds are too loose and you'll stop opening the email by week two.

Routine #2: Segment hygiene check, weekly

Segment hygiene earns a weekly slot because every workspace accumulates a segment graveyard. Empty segments, segments with duplicate condition logic, segments built for a campaign that launched six months ago and got archived. Nobody runs that audit by hand, which is exactly why it's the right shape for a Routine.

What "empty, unused, or redundant" means in practice: a segment with zero people, two segments with the same condition logic and different names, or a segment that hasn't been referenced by any active campaign or broadcast in 60 days.

A worked example. A campaign trigger watches "Trial Users - Week 2 No Activity". Someone renames the segment to "Trial - Wk2 - No Event" because they're building a new variant. The original segment still exists, still has people in it, and the campaign trigger still fires against it. But the new variant is the one being maintained. Two months later, the original segment is stale, the trigger is firing into a stale audience, and nobody knows. The hygiene routine catches it.

A tight prompt:

Each Monday, list segments that meet any of: (a) zero members; (b) share condition logic with another segment; (c) have not been referenced by an active campaign or broadcast in the last 60 days. Cap the list at the worst 10. For each, tell me which campaigns reference it.

Routine #3: Goal conversion report, weekly

This routine only works if you've set Goals up properly. And Goals are not conversion events—they're an independent workspace-level object that multiple campaigns connect to. If you've configured the same outcome as both a Goal and a per-campaign conversion event, you'll get reports that don't reconcile.

Weekly is the right cadence. Goal conversion moves slower than deliverability, and a daily report on goals trains you to overreact to noise.

Write the prompt to compare week-over-week, not just totals

The default suggestion gives you absolute numbers. Tell the routine to compare against the prior week and flag movement. That's the version you'll open and read.

Each Monday morning, list every Goal in this workspace. For each: current week conversion count, prior week count, week-over-week percentage change. Flag any Goal where the change is more than ±15 percentage points and give me a one-line note on which campaigns connect to that Goal. Headline number at the top of the email.

If you don't yet have Goals defined, skip this routine until you do. A goal conversion report against zero goals is a wasted run every week.

The routine to pause: Subscriber churn early warning

This one looks useful and isn't, on the default prompt.

The card's pre-filled prompt asks the agent to "spot at-risk subscribers before they disengage or unsubscribe." Too broad. On a daily schedule, each run reads across your active list, classifies signals against the agent's own definition of "at risk," and emails you a number. Routines don't draw on your AI credit bundle—only LLM Actions do (AI credits docs)—so the cost here isn't money, it's attention. A prompt that reads broadly across the whole active list every day produces the longest, vaguest email of any card, and it lands every morning whether or not there's anything to do.

There's a second problem: actionability. If the routine flags 4,000 at-risk subscribers on Monday morning, what does Monday-morning-you do with that? Nothing scales from that output, and the unsubscribe click is the receipt for a decision made weeks earlier anyway. A list of 4,000 doesn't tell you which 50 to intervene on.

The constrained alternative: pick one segment, one timeframe, one threshold.

Each Monday, list trial users in week 3 who have not fired key_event_activation in the last 7 days. For each, give me signup date, last recorded event, and which onboarding campaign they're currently in. Cap the list at 50.

That produces a list a human can act on between coffee and the first standup. If you can't write a constrained version of the prompt for your business, pause the routine. Don't run it on the default.

Two routines worth thinking about, but not yet

Broadcast recap. Useful if you broadcast at least once a week. Below that cadence the recap is empty most weeks and slightly distracting on the weeks it isn't. Run a one-off agent conversation after each broadcast and review it manually until your cadence justifies the schedule.

Audience growth tracker. Useful in principle, but most of what you'd ask it to surface is already in the five-minute Monday Workspace Performance scan. Opening that page is instant. A scheduled routine emails you the same data on a delay and spends one of your limited active-routine slots. The dashboard wins until you've outgrown what it shows.

How to write a Routine prompt the agent can execute on a schedule

"Check my campaigns" is the canonical bad prompt. The docs themselves call it out. The agent has no thresholds to evaluate against, no time window to scope to, and no idea what a finding looks like. So it does a vague pass over everything and emails you a vague paragraph. That's a wasted run that emails you nothing usable.

The good-prompt pattern has four components. A threshold (what counts as a flag, in specific numbers). A time window (last 24 hours, last 7 days). A finding format (what the email should contain when something is flagged). And an empty case (what to say when nothing is flagged).

Specificity also keeps the routine worth reading. A vague prompt makes the agent search broadly and reason about what's relevant, so it emails you a longer, vaguer summary you'll stop opening by week two. A specific prompt scopes the session to a known shape with a clear finding. One thing specificity doesn't affect is your AI credit balance: routines run the agent, and the agent doesn't draw on AI credits—that bundle is reserved for LLM Actions inside campaigns. The cost of a sloppy routine is your attention, not your balance.

The 30-minute setup

Open the agent sidebar, click Routines. Pick the suggestion card you want and read the pre-filled prompt—assume you'll rewrite it before saving. Set the schedule: daily for the deliverability watchdog, weekly for the others. Hit save. Wait for the first run.

When the first email lands, ask yourself one question. Would I have spent five minutes reading this if it had arrived on a Tuesday? If yes, the prompt is doing its job. If no, edit the prompt or pause the routine. The point is to surface things you'd otherwise miss, not to manufacture a Monday-morning chore.

Failure threshold, plan limits, and what to do when a routine auto-pauses

A routine auto-pauses after 3 consecutive failures by default. The threshold is configurable between 1 and 5 when you create the routine. If a routine pauses, read the error banner on the Routines page, edit the prompt to fix what broke, then toggle it back on. Toggling resets the failure counter.

The active routine limits are 1 on Essentials, 5 on Premium, 5 on Enterprise—per user, per workspace. Five is not a target. Three well-written routines beats five vague ones every time, and you'll actually read the output.

Frequently asked questions

Do Customer.io Routines cost AI credits?

No. Routines are recurring agent sessions, and the agent doesn't consume AI credits—only LLM Actions inside campaigns do (AI credits docs). Routines are limited by plan instead: one active routine per user per workspace on Essentials, five on Premium and Enterprise. The real constraint on a routine is your active-routine slots and your attention, not a credit balance.

Can a Routine modify my campaigns or segments?

No. Routines run in safe mode, which is read-only. The agent can query data, analyse it, and email a summary, but it cannot create, edit, or delete campaigns, segments, newsletters, or workspace settings.

What's the difference between a Routine and a scheduled report in Customer.io?

A scheduled report runs a predefined query on a fixed cadence and emails the same shape of output every time. A Routine runs an agent session against your workspace and produces a written summary in plain English, with whatever findings it considers relevant under your prompt. Reports work when the question is fixed. Routines work when you want the agent to reason about what to flag.

How many routines can I have on a Premium plan?

Five active per user, per workspace, with a daily minimum interval. You can create up to 10 total per user per workspace, including paused ones. Two users on Premium each get their own 5-active limit independently.

Can I run a Routine manually outside its schedule?

Yes. Ask the agent in a chat conversation to run a specific routine immediately. Handy for testing a prompt before you commit to the schedule, or for pulling an on-demand report.

What happens if a Routine fails three times in a row?

It auto-pauses by default and the Routines page shows an error banner with the last failure. Edit the prompt to fix the issue and toggle the routine back on, which resets the failure counter. The threshold is configurable between 1 and 5 consecutive failures when you create the routine.

Can different users in the same workspace each have their own routines?

Yes. Routine limits are per user, per workspace. Your colleagues' routines don't count against your active or total limits, and you don't share each other's prompts.

Why does the agent run a fresh session each time instead of remembering previous runs?

Each routine run is a separate agent session with no conversation history from previous runs. That keeps routines independent from each other and from your active chats, so a routine running in the background won't interrupt a conversation you're in the middle of. It also means the agent can't accumulate context across runs, so anything you want compared across time (week-over-week, day-over-day) has to be queried fresh from workspace data each run.

Are Routines available on Essentials?

Yes, with a smaller limit: one active routine per user per workspace, and a weekly minimum interval. Routines don't draw on AI credits on any plan—that's only LLM Actions. If you need daily routines, you'll need Premium or Enterprise.

How do I pause a Routine without deleting it?

Toggle it off on the Routines page. The schedule pauses but the routine stays in your workspace. Paused routines don't count against your active limit, but they do count against your 10-routine total per user per workspace.

What's the smallest schedule interval I can set?

Weekly on Essentials, daily on Premium and Enterprise. Sub-daily intervals (hourly, every six hours) aren't supported.

Can a Routine email someone other than me?

No. Results go to the email address on your Customer.io user account. If you want a routine's findings shared, you forward the email manually.

How is a Routine different from an LLM Action inside a campaign?

A Routine runs the agent against your workspace on a schedule and emails a summary to you. An LLM Action fires inside a campaign at runtime for each individual customer entering that step, and stores the model's output as a journey attribute. Routines are for the marketer. LLM Actions are for the customer journey. And the billing differs: LLM Actions draw on your AI credit bundle, Routines don't.

Sources

• Routines, Customer.io Documentation, last updated 8 June 2026
• AI credits, Customer.io Documentation, last updated 8 June 2026
• Some launches add features. This one adds up to something., Naomi West, Customer.io Announcements, 8 April 2026
• Why we built one agent, not many, Matthew Newhook, Customer.io Learn, 10 April 2026
• How the agent works, Customer.io Documentation, last updated 8 April 2026
• What shipping an MCP server taught us about our customers, Paul Senechko, Customer.io Learn, 23 April 2026