Blog · 20 November 2026
Automations that survive month three
If you search for “automation maintenance”, the top results promise flashy percentages. 70%, 85%, and variations. Anyone can check. The wins pile up in month one. Almost none explain what happens when a price changes, a provider shifts, or a file format moves.
I design agents and flows that stay alive after that. I do it because reality moves. What fails is not “the model”. The edges fail: integrations, data, schedules, and permissions.
What breaks an automation after month three
Pages change their HTML and the selector stops finding the value. An API adds a field and deprecates another. A CSV arrives with columns in a new order. The provider tightens the per‑minute limit. A credential expires. A holiday hits the rota and no one reviews an alert. All that happens quietly if I do not prepare for it.
Growth also breaks things. Ten rows become a hundred thousand and the same job blows memory or times out. A price changes and rounding drifts create mismatches that no one checks. Legal asks to anonymise a field that used to store personal data.
Early signals instead of silence
I prefer a task that stops itself and alerts to one that keeps producing rubbish. I place signals at entry and exit. At entry, I validate format, types, and ranges. At exit, I check volumes, distributions, and unique keys. If the input fails, I cut. If the output strays from reasonable, I cut and notify.
The alert is not a shout every five minutes. It is a clear summary with the minimum: which stage failed, what it expected, what it got, and where to see the batch. And a link to a log with context. No black screens. No cryptic traces with no explanation.
Data contracts and light tests
I declare a contract. Which fields I expect, which keys must be unique, and which nulls are allowed. With that, a quick preflight stops errors from cascading. If the source is a flaky site, I add a simple visual check: a screenshot or a checksum of key sections to spot layout changes.
I test in three tiers: a tiny valid file, one with typical errors, and one with edge data. No lab needed. With declared scenarios, for example, “a file with only one row” or “a number with a comma instead of a dot”, I get strong coverage with little friction.
Idempotency, retries, and when to stop
I retry with exponential backoff, but only when it makes sense. Without idempotency, retries duplicate orders or emails. So I assign stable identifiers per operation and check before writing. If a source misbehaves three times in a row, I trip a circuit breaker. It is marked out of service and I do not burn the quota.
I declare tolerances as well. Small drift can be normal. When I design the flow, we agree cut thresholds with the team. A realistic example: “if the daily counter of valid records falls below half of the average of the last week, stop and alert”. It is explainable and adjustable, not a magic number.
Heartbeats and controlled changes
A daily heartbeat confirms that everything lives. Even with no data, the job reports an “empty run” with duration and versions. That lets me spot broken dependencies early. When I update a library or a connector, I do staged rollouts. First a small subset. I watch metrics, then expand. If it goes wrong, I roll back without losing the batch.
Logs that help you decide
A useful log is not a dump. It is structured. It includes batch IDs, stages, timings, and counts. With that I can answer in minutes: “what did it affect?”, “since when?”, “how much is left to restore?”. When people are involved, I add readable context: the original filename, the source link, and the template used.
Humans in the loop without friction
Some steps should not be fully automated. An accounting exception review, for example. I design it as a review tray. A task arrives with a summary, evidence, and a button to decide. If no one responds within the agreed window, I escalate to another person. Better that than blocking the whole queue over a small doubt.
Maintenance cost from the start
Maintenance is not an add‑on. It is part of the design. In a starting scenario, a team that spends twenty hours a month on this cuts repeat fires and frees attention to improve. The key is to assign those hours from day one and measure them. If you cut them, technical debt grows and month three collects with interest.
If you wonder what is worth automating, I have already explained how I size an AI automation that pays off and why maintenance is there from the start. It is not about promising numbers that do not exist. It is about committing to useful life.
How I design so it warns before failing
I start with a dependency map. Which APIs I touch, which sites I scrape, which files arrive and how often. Each edge gets a control: schema validation, rate limits, credential expiry, temporary storage, and safe resume. Then I set minimal metrics: counts, timings, and categorised errors.
I define alert routes: one channel for informatives, one for errors, one for stops. The goal is for the team to see health, not only emergencies. A daily “all good” lowers anxiety and makes trend shifts visible. When something changes, the first alert arrives with a diagnosis and a suggested next step.
What I do when something really changes
If a provider shifts the format, I do not patch blind. I reproduce with samples, update the contract, expand tests, and redeploy in stages. If response times rise, I adjust concurrency and batch sizes. If a page is redesigned, I renew selectors and add a specific alert for that case. I always document the why.
The result is not an immortal automation. It is one that breaks in a controlled way, with clear alarms and ways back. That is what gets you to month three without shocks, and to month six with measurable gains that do not depend on inflated percentages, but on fewer incidents and faster, better decisions.
I post working automations on @WhatsMarketing_es.
If you want me to look at your case, get in touch. I work from Buenos Aires, originally from Málaga, with clients in Mexico City, Argentina, the rest of Latin America and Spain.