I spend a lot of my time helping organizations automate work, so it may surprise you that I spend just as much time telling them where to stop. The failures I have seen almost never come from automating too little. They come from automating a decision that needed a human, discovering the problem too late, and cleaning up a mess that costs more than the automation ever saved.
Knowing where to draw the line is a skill, and it is one you can develop. Here is how I think about which decisions stay human and how to build review points that hold up under pressure.
Three Decisions That Should Keep a Human in the Loop
Over the years, the same three categories keep showing up as places where full automation goes wrong. If a step touches any of these, put a person on it.
-
Money. Anything that moves funds, commits your organization to a cost, sets a price, or approves a payment deserves human eyes. Automation is great at preparing the transaction and terrible at noticing when something is off by a decimal point or heading to the wrong recipient. Let the system draft the payment; let a person release it.
-
Relationships. Communication that shapes how a client, donor, employee, or partner feels about you carries risk that a machine cannot assess. An automated message that lands wrong at a sensitive moment can damage a relationship you spent years building. The tool can draft and schedule; a person should decide whether this particular message, to this particular person, at this particular time, is right.
-
Judgment calls and exceptions. Automation follows rules. The moment a situation falls outside the rules, an unattended system will either force a bad fit or fail silently. Anything involving nuance, fairness, or a genuine gray area needs a human who can weigh context. The presence of exceptions is the clearest signal that a step is not ready to run alone.
Designing Review Checkpoints That Work
Saying "a human should review it" is easy. Building a checkpoint that people actually use, rather than rubber-stamp, takes a little design. A few principles I rely on:
- Put the human at the decision, not the busywork. The point of the checkpoint is judgment. Let automation gather the information, lay out the options, and prepare the draft, so the person spends their attention on deciding rather than assembling.
- Make the review meaningful, not constant. If you ask someone to approve every routine item, they will start clicking approve without looking, and the checkpoint becomes theater. Route only the items that genuinely need judgment to a human, and let the clearly-fine ones flow.
- Give the reviewer enough context to catch a problem. A checkpoint is only as good as the information at it. Show the reviewer what the automation is about to do and why, so a wrong decision is visible before it goes out, not after.
- Make it easy to stop. The reviewer needs a real ability to halt or change the action. If saying no is hard or slow, people stop saying it.
The Failure Modes of Over-Automation
When organizations automate past the point they should, the problems tend to look the same. Recognizing these early saves a lot of pain.
The first is silent errors at scale. A human making a mistake makes one mistake. An automated process making a mistake makes the same mistake hundreds of times before anyone notices, and the cleanup is proportionally larger.
The second is automation drift. The world changes, the rules the automation was built on quietly go stale, and the system keeps confidently doing the now-wrong thing. Without a human periodically watching, drift is invisible until it causes damage.
The third is skill atrophy. When people stop doing a task entirely, they lose the ability to notice when it goes wrong or to step in when the automation fails. Keeping a human in the loop is partly about keeping the human capable.
The fourth is the eroded relationship. Over-automated communication starts to feel like it, and the people on the receiving end can tell they are being processed rather than served. For a small business or nonprofit whose advantage is being personal, that is a costly trade.
Automation and Judgment, Working Together
The goal is not to automate everything or to distrust automation. It is to let machines do what they are genuinely good at, which is the consistent, high-volume, rule-bound work, while keeping human judgment exactly where judgment matters. Done well, this pairing is more reliable than either a fully manual process or a fully automated one, because each covers the other's weakness.
If you are weighing an automation and are not sure where the human should stay in the loop, that is a conversation worth having before you build. Take a look at how I approach AI automation, or book a free consultation and I will help you map out which steps stay human in your specific process.