In short: An error your team sees gets corrected. It costs time. An error nobody sees costs trust, or stays forever. Anyone planning AI automation should therefore not ask how often an error happens, but whether it gets noticed.
Two kinds of errors
Picture two situations.
In the first, an AI writes a draft reply to a customer email. The draft is too formal and names the wrong contact person. Your employee reads it, corrects two sentences and sends it off. The error cost one minute.
In the second, an AI automatically filters out incoming inquiries that do not fit the core business. A good inquiry accidentally lands in the reject pile. Nobody ever sees it. The customer waits, hears nothing and goes to a competitor.
The second error may happen less often. It is still the more expensive one. And it shows up in no error statistics, because nobody can report an error they never saw.
Priority by visibility
In the sales tool I am currently building with a partner, we listed every point where a decision is made about the quality of a company. Then we sorted them not by probability, but by visibility.
The guiding question was: If something goes wrong here, does the user see it?
- Visible: A phone number is missing. The user sees the empty field and looks it up themselves. Annoying, but harmless.
- Semi-visible: A number is wrong. The user only notices when calling. Costs time and a bit of trust.
- Invisible: A good company is filtered out as a duplicate or an aggregator page. It never appears in the list. Nobody misses it.
The invisible points get the strictest thresholds. For us, that means: no good lead may be wrongly filtered out. Not few, none.
Where no human is looking, the indicator becomes the decision
Especially dangerous are workflows in which no human looks anymore. As long as an employee checks every traffic light herself, the traffic light is an indicator. She can question it.
As soon as a process continues automatically, the traffic light is no longer an indicator but the decision. Red then means: will not be processed. Without anyone noticing.
That is the moment the type of error changes. An error that was visible before becomes invisible. That is why, for every step you let run fully automatically: the quality requirement goes up.
When in doubt, do not filter out
From this follows a rule that runs through all my projects: When in doubt, do not filter out, but flag.
If the system is unsure whether a company fits, it goes into the list with a warning. Better one riding stable too many in the list than one sauna rental too few. The riding stable costs a glance and a swipe. The missing sauna rental costs an order nobody knows about.
The same principle applies to email automation. When I build a system for clients that sorts inboxes, there is always a category “Unclear”. Without it, the model produces confident wrong decisions instead of honest gaps. An honest “Unclear” is worth more than a confident wrong classification.
Too little is not experienced as an error
There is a second form of invisibility that is often overlooked. A tool that delivers too little is not experienced as faulty, but as mediocre.
If your new system finds only 30 matching companies instead of 80, nobody complains. The user thinks: well, that is probably all there is. Or they go back to their old way without ever reporting an error. In both cases, you do not learn that something is wrong.
That means: you also have to measure what is missing. Not just whether what was found is correct.
A visible error is bad enough
To avoid a false impression: visible errors cost too. Just in a different way.
In a demo of the sales tool in front of experienced salespeople, the AI enrichment live wrote the legal notice of an aggregator page into several companies. One participant said immediately: “But then I will just delete all of them.”
A single visible data error cost more trust in that moment than ten missing features. Data quality is the trust killer, not missing features.
The lesson: visible errors must be rare, invisible errors must be ruled out. Both at the same time.
What should never happen automatically
From these considerations, a few firm principles follow for me that I apply to every automation project:
- Nothing leaves the building without approval. An agent that answers emails writes drafts. It does not send.
- No invented facts in drafts. If a price or date does not come from a source, the draft shows a visible gap that a human has to fill.
- Undo instead of asking. Bulk changes like an import show beforehand what will happen and can be fully reversed.
- What is filtered out stays viewable. What the system discards does not vanish into nothing, but lands in a list that can be spot-checked.
What this means for you
Go through your planned or running AI process step by step and ask at each point:
- If something goes wrong here, does anyone see it? If not, this is your most critical point.
- What happens in case of doubt? Filter out or flag?
- Is there a category for “I don’t know”?
- What leaves your company without a human having seen it?
- Do you measure what is missing, or only what is there?
The best automation does not make errors impossible. It makes sure they get noticed.
If you would like to know where the invisible points are in your process, we go through it together in a first call.
