When Not to Automate: Tasks That Should Stay Human

We teach automation. We build automations. And we are here to tell you that some tasks should not be automated, not because the technology cannot handle them, but because automating them creates problems that cost more than the time they save. Knowing when not to automate is as important as knowing how.

The automation bias problem

When you learn a new tool, everything starts looking like a use case for that tool. This is normal. We have seen it in every cohort of our course. A student learns how to set up a Zapier workflow, and suddenly every process in their life looks like it needs a zap. We call this "automation bias," the tendency to automate tasks simply because you can, without asking whether you should.

The cost of bad automation is not always obvious upfront. It shows up later as damaged relationships, compliance violations, missed context, or decisions that nobody can explain or defend. Here is where we draw the line.

Tasks involving empathy and emotional judgment

Delivering difficult news

Telling a customer their claim has been denied. Informing an employee their contract will not be renewed. Notifying a patient about a concerning test result. These conversations require emotional awareness, the ability to read how the other person is receiving the information, and the flexibility to adjust your approach in real time.

An automated email or chatbot message delivering bad news feels cold and dismissive. It communicates that the organization did not consider the recipient important enough to warrant a human conversation. In Kenyan professional culture, where personal relationships underpin business relationships, this can destroy trust that took years to build.

The rule: If the message might cause the recipient distress, a human should deliver it.

Handling complaints that involve frustration or anger

Chatbots can handle routine inquiries. They should not handle a customer who is genuinely angry about a mistake your company made. When someone calls because their M-Pesa payment went through but the service was not delivered, they need to feel heard. A chatbot's "I understand your frustration" followed by a scripted response does the opposite.

The most damaging customer service experiences we see in Kenya happen when frustrated people are trapped in automated loops, pressing numbers, repeating their issue to a bot, and never reaching someone who can actually help. The time "saved" by automation is lost many times over in customer churn and negative word-of-mouth.

The rule: If the person is emotionally invested in the outcome, route them to a human as quickly as possible.

Tasks requiring contextual judgment

Hiring decisions

AI and automation can help screen resumes for basic qualifications, schedule interviews, and send follow-up communications. They should not make the final hiring decision.

A CV tells you what a person has done. It does not tell you how they will work with your specific team, whether they share your organization's values, or whether they will grow in the role. These assessments require judgment that accounts for context no algorithm can fully capture.

In Kenya's job market, where personal networks and referrals carry significant weight, a purely automated screening process can filter out strong candidates who do not have conventional career paths but bring valuable experience. The university graduate who worked in their family's hardware shop in Kisumu for three years might be exactly the resourceful, practical thinker your business needs, but an automated screen might reject them for not having "formal" experience.

The rule: Use automation to support the hiring process, not to make the final call.

Performance evaluations

We have seen tools that claim to automate performance reviews by analysing productivity metrics, communication patterns, and task completion rates. This is a dangerous road.

Performance is not purely quantitative. An employee who handled fewer tickets but resolved the most complex customer situations is not underperforming. A team member who spent time mentoring new hires may show lower individual output but increased team capability. These nuances require a human manager who understands the context.

The rule: Gather data with tools, but assess performance through human conversation.

Negotiation and relationship-dependent decisions

Whether you are negotiating a contract with a supplier, discussing payment terms with a client who is going through a difficult period, or mediating between team members, these situations depend on reading the room, understanding unspoken concerns, and making judgment calls that account for the relationship beyond the immediate transaction.

In Kenyan business culture, many important agreements are built on personal trust that develops through face-to-face interaction and repeated good faith. Automating the communication layer of these relationships signals that the relationship is transactional, not personal. That signal has consequences.

The rule: If the outcome depends on the quality of a relationship, keep the interaction human.

Tasks with serious consequences for errors

Financial approvals above a meaningful threshold

Automating the approval of routine, low-value transactions is sensible. Automating the approval of a a reasonable cost vendor payment without human review is not. The potential cost of an error (paying the wrong vendor, duplicating a payment, approving a fraudulent invoice) is too high relative to the time saved.

The rule: Set a clear threshold. Below it, automate. Above it, require human review and approval.

Legal and regulatory compliance decisions

Should this transaction be flagged for anti-money laundering review? Does this marketing claim comply with the Kenya Consumer Protection Act? Is this data processing activity compliant with the Data Protection Act? These questions involve interpreting regulations in context, and regulations in Kenya change frequently. A workflow that was compliant last quarter may not be compliant today.

Automating compliance checklists and reminders is helpful. Automating the compliance decision itself is risky, because when something goes wrong, "the system approved it" is not a defence that regulators accept.

The rule: Automate compliance tracking and alerts. Keep compliance decisions with qualified humans.

Medical and health-related assessments

AI tools can assist with preliminary screening, appointment scheduling, and information gathering. They should not make diagnostic or treatment decisions without physician oversight. A chatbot that tells a patient their symptoms "are probably nothing to worry about" when those symptoms actually indicate a serious condition creates liability and endangers health.

The rule: AI can gather and organise health information. Diagnosis and treatment recommendations stay with licensed professionals.

Tasks where accountability matters

Here is a question that clarifies when not to automate: if this automated process makes a mistake, who is responsible?

If the answer is clear (a specific person reviews and approves the output, and that person takes responsibility), automation is working as a tool. If the answer is vague ("the system did it" or "nobody checked"), you have an accountability gap.

Accountability gaps cause real damage. When a Kenyan bank's automated system incorrectly flags a customer's account and freezes it, the customer does not blame "the algorithm." They blame the bank. The bank then needs a human to review the case, communicate with the customer, and resolve the issue. The automation that caused the problem creates more human work than it saved.

Before automating any process with meaningful consequences, define who reviews the output, who is accountable for errors, and how errors are corrected. If you cannot answer these questions clearly, the process is not ready for full automation.

The right mental model

The best automation practitioners we know think of automation as removing tedious steps from a process, not removing humans from it. The human stays in the loop for judgment, empathy, and accountability. The machine handles the repetitive, structured, and time-consuming parts.

Our AI and Automation for Beginners course teaches this approach from the first module. We want our students to build automations that make them more effective, not automations that create problems nobody noticed until it was too late.

FAQ

Does this mean automation is limited to simple tasks?

No. Automation can handle complex, multi-step processes. The key distinction is not complexity but consequence. A complex data pipeline that transforms and routes information can be fully automated if errors are detectable and correctable. A simple task like approving a refund request may need human involvement if the amounts are significant or the context matters.

How do I convince my boss that some tasks should not be automated?

Frame it in terms of risk and cost. Calculate the potential cost of an automation error (lost customers, compliance fines, reputational damage) and compare it to the time saved. Most managers respond well to concrete scenarios: "If this automated approval process sends a reasonable cost to the wrong vendor, here is what recovery looks like."

What if a competitor automates everything and becomes faster?

Speed is one factor among many. A competitor who automates customer complaint handling might process complaints faster, but if their customers feel unheard and switch to your business where a real person listens, their speed advantage becomes a retention problem. Compete on the quality of outcomes, not just the speed of processing.