When automation comes up, most people think of “set it up, then it runs on its own.” Systems that prepare reports, collect data and monitor prices look attractive because they save time. But it is exactly this “on its own” idea that is the source of the biggest risk.
The invisible risk of automation
If a system makes errors silently, nobody may notice: a report is produced with old data, a measurement does not run for days, a wrong figure goes to the client. Automation speeds up errors too. That is why my approach is this: automation speeds up the work, it does not take over responsibility.
Three rules of supervised automation
- The system prepares a draft, a human delivers. Every report, every decision and every visual is read before it goes to the client.
- Supervision is written down. What will be checked, by whom, and against what is determined in advance (a supervision protocol).
- That it works is tracked with evidence. “It did not error” is not enough; evidence that the work was really done (a file, a record, a figure) is required, and a warning comes if it is missing.
Example: ad double verification
Ad panels count by their own attribution rules, and this is often higher than the store record. For an e-commerce brand I compared the sales declared by the panel with the store's order records: the panel had shown 33 sales, while the store records had 14 orders from ads. I now make this comparison regularly. The system does the comparison quickly; keeping a mismatched figure out of the report is a decision, and that decision belongs to a human.
When should automation not be set up?
- If the process is not yet clear: process first, then automation.
- If the cost of an error is very high and irreversible: a human does that step, and the system only supports.
- If there is nobody to supervise the work: setting up unsupervised automation means producing silent errors.
