Why automation projects look successful until they are not
RPA and zap-style connectors create fast wins on repetitive tasks: data entry, notifications, lead routing. The failure arrives later — when the UI changes, the vendor rate-limits, or the exception path was never designed.
API-first automation is more durable than UI scraping, but only if schemas are versioned, errors are observable, and humans have a queue for the cases machines should not decide.
Half of AI projects delivering measurable outcomes is not a victory lap. It is a reminder that automation without measurement is theater.
The edge problem
Automation lives at the edges between systems. Those edges are where consent, identity, and commercial state change. If marketing automation can write to CRM without validation, you will invent customers. If support automation can refund without limits, you will invent losses.
Design automation as a face with contracts: allowed actions, idempotency keys, retries, dead-letter queues, and clear ownership. Then connect it to product and web events so the machine is not guessing.
A production checklist before the next bot ships
Can you replay yesterday’s runs? Can you explain why an action fired? Can you stop the flow in under a minute? Can a non-engineer see the health of the system? If any answer is no, you own a liability, not an asset.