Your automation is live. Now the real work starts.
Automation projects end at go-live. But every workflow depends on APIs that change, data formats that shift, and models that update. Nobody plans for that until something breaks.
Your automation is live. Now the real work starts.
Most automation projects have a clear end point: go-live. Project closed. Team moves on.
Six months later, something breaks. An API endpoint was deprecated. A field was renamed. A model version was updated. Nobody noticed because nobody was watching.
This isn't a story about careless teams. It's about projects that defined success as "launched" and never thought past that.
Automation is a dependency, not a deliverable
Every automated workflow depends on things outside your control: APIs that get versioned, software that updates, data formats that shift over time. These changes happen whether your workflow accounts for them or not.
The typical failure here isn't a crash. It's a workflow that quietly produces wrong data for weeks. CRM records with missing fields. Invoices misrouted. Orders that arrived but ended up in the wrong system. The error compounds until someone notices a downstream consequence, usually weeks after the actual incident.
AI-powered workflows aren't more stable
They add another dimension, actually. Model versions change. Prompts that worked reliably six months ago may produce differently formatted outputs with an updated model. Downstream steps that parse those outputs break quietly.
Not a reason to avoid AI in workflows. Just a reason to maintain them the same way you'd maintain anything else.
What maintenance actually means
I'm not talking about a dedicated ops team. For most workflows the overhead is small.
Every workflow that runs daily and causes real damage if it fails needs one owner: a person who gets alerted when something breaks and knows the workflow well enough to debug it. Basic monitoring: an alert when the workflow hasn't produced results in 24 hours. Most platforms offer this natively. Almost nobody turns it on. And a protocol for upstream changes: when a connected system announces an update, someone checks whether the workflow is affected before it lands.
For simple three-step flows on stable integrations: two to four hours per year. For complex multi-system flows with AI components: more. Never zero.
The planning implication
A workflow saving 300 hours of manual work per year is worth maintaining even if it takes 20 hours per year to keep it healthy. But if those 20 hours aren't in the plan, the ROI numbers are wrong.
The question worth asking at the start: what does this workflow need to still be running reliably in two years? Answering that early is cheaper than answering it after an incident.
If you want to check which of your existing automations have proper monitoring and clear ownership, the free Automations Check takes about 30 minutes.