What automation should never decide for you

Automation is good at one thing above all others: repeating a decision you have already made well. It takes a rule you have settled on and applies it a thousand times without tiring, forgetting, or arguing. That is genuinely valuable. The trouble starts when a business quietly lets the machine cross from repeating a decision to making one, and nobody notices the line being crossed.

The standard reference definition captures the boundary neatly: automation covers "a wide range of technologies that reduce human intervention in processes, mainly by predetermining decision criteria, subprocess relationships, and related actions, as well as embodying those predeterminations in machines." [3] Read that twice. The criteria are *predetermined*. Someone decided them, in advance, once. The machine's job is to embody that decision, not to originate it. When a tool starts originating decisions you never actually made, you have not saved time. You have handed over judgement, usually without meaning to.

This article is about where that line sits, and how to keep it in view. We build software that automates a great deal, so we have a direct interest in you drawing the line clearly rather than vaguely. A tool that oversteps its remit costs its owner trust, and trust is the only reason anyone lets software touch their customers.

The test: can you write the rule down completely

Here is a practical test for whether something is safe to automate. Can you write the rule down completely, in advance, so that a stranger following it would reach the same answer you would?

If you can, automate it. "Send the receipt the moment payment clears." "Move a lead to *cold* if nobody has replied in thirty days." "Flag any invoice over the approval limit." These are decisions you have already made. The rule is the decision. Running it by hand is just slower.

If you cannot write the rule down completely, you are not automating a decision. You are automating a *guess* about a decision, and dressing the guess as certainty. "Decide which complaint deserves a refund." "Decide whether this customer is worth keeping." "Decide the tone to take with someone who is upset." You cannot reduce these to a clean rule because the answer depends on things the machine cannot see: context, history, the specific human on the other end, and your own sense of what your business is for. We wrote more about applying this test in how to tell whether a task should be automated.

The honest position is that most work is a mix. A follow-up sequence can be fully automated in its *timing* and its *sending*, while the *decision to keep chasing a particular person* stays with you. Splitting a task this way is usually better than automating all of it or none of it.

Four decisions to keep for yourself

Some categories are worth naming, because they are the ones businesses most often surrender by accident.

Decisions that end a relationship. Closing an account, cancelling a customer, writing someone off as lost, refusing service. Automation is fine for *flagging* that a relationship looks dead. It should not be the thing that pronounces it dead. The cost of a false positive here is a person, and people do not come back easily.

Decisions that override a rule you set. The point of an exception is that a human looked and judged. If your system quietly grants exceptions on its own, you no longer have a rule, you have a suggestion. Keep the granting of exceptions with a person, even when the flagging is automatic.

Decisions about money that a customer will feel. Charging a card, applying a late fee, pricing a quote, issuing or refusing a refund. Predetermined amounts on predetermined triggers are fine. Novel judgements about what is fair are not, because "fair" is exactly the thing you cannot write down completely in advance.

Decisions that carry legal or reputational weight. This one is not only good sense; in much of the world it is law. Under European data protection rules, a person "shall have the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her." [1] The same law spells out the remedy where automated decisions are allowed at all: the person keeps "the right to obtain human intervention on the part of the controller, to express his or her point of view and to contest the decision." [4] Even if that regime does not bind you directly, it encodes a principle worth borrowing: a decision that materially affects a person should have a named human who can be asked about it.

Why the line drifts without you noticing

The reason this matters more than it seems is a well-documented quirk of how people work alongside machines. It is called automation bias: "the propensity for humans to favor suggestions from automated decision-making systems and to ignore contradictory information made without automation, even if it is correct." [2] Once a system produces a confident-looking answer, people stop checking it. The screen said cold, so the lead is cold. The system approved it, so it must be fine. The output becomes the truth, and the human becomes a rubber stamp who signs faster and faster.

This is how judgement leaks away. Not in one deliberate handover, but in a hundred small moments where checking felt unnecessary because the machine seemed sure. The remedy is not to distrust every output. It is to design the work so that the decisions in your four kept categories *require* a human to actually look, not merely to click. A number on a dashboard is only useful if someone reads it and can act against it, a point we made in numbers that change a decision.

What good automation feels like

When the line is drawn well, automation feels like a diligent assistant, not a replacement. The assistant does the gathering, the timing, the repeating, and the flagging. It says: *here is what I noticed, here is what the rule suggests, here is what needs your call.* You make the calls that need a person, and you make them faster because the groundwork is done.

When the line is drawn badly, automation feels like a colleague who has quietly started signing your name to things. Everything is faster, right up to the day it sends the wrong customer the wrong message, and you discover you had stopped reading months ago.

The difference is not the technology. Both cases run the same software. The difference is whether a person decided, in advance and on purpose, exactly which decisions the machine was allowed to make, and kept the rest. There is a related judgement about when to add people rather than tools, covered in when to hire and when to automate.

A short checklist

Before you automate anything that touches a customer, ask:

• Can I write the rule down completely, so a stranger would reach the same answer?
• If it goes wrong, what is the worst outcome, and can I live with it happening unattended?
• Is there a named human who can explain and reverse the decision if a customer asks?
• Have I automated the *doing* while keeping the *deciding*, where the two can be split?

If a task passes all four, automate it and stop doing it by hand. If it fails the first, you are not saving time. You are outsourcing a judgement, and one day it will make the call you would never have made.