What to automate with AI — and what to leave to rules

The mistake teams make with AI automation is not being too cautious. It is automating the wrong things — pointing an expensive, probabilistic model at work that simple rules did better, while leaving the genuinely hard judgment work manual. Here is a framework for deciding what AI should automate, what rules should automate, and what should stay with a person.

Three buckets, not two

Every task in a workflow falls into one of three buckets: rules can do it, AI should do it, or a human must do it. Most failed AI automation projects come from putting tasks in the wrong bucket — usually from over-reaching, handing AI work that belonged to rules or to people. Getting the sorting right is most of the battle, and it happens before a line of code is written.

What rules should automate

Anything deterministic and structured belongs to rules. If a task is “when this, do that” with no judgment involved, rule-based automation is faster, cheaper, perfectly reliable, and fully auditable. Moving records between systems, sending scheduled reminders, applying a fixed calculation, routing on a clear field — these are rules jobs.

Using AI here is not just wasteful, it is actively worse. You are paying a real per-run cost for something that was free, adding seconds of latency to something that was instant, and introducing a probability of error to a task that previously had none. The first question of any automation project is not “how do we use AI” — it is “which of this is just rules”.

What AI should automate

AI earns its place on tasks that need reading, interpretation, or judgment over messy input — and where being right most of the time is genuinely valuable. Classifying and routing unstructured requests, extracting data from inconsistent documents, summarising long material, drafting first versions, triaging by reading and deciding.

The test is two-part. First: does a person currently do this by reading something and making a judgment? Second: is an occasional error acceptable, or cheap to catch and correct? If both are true, it is a strong AI automation candidate. If the task is mechanical, rules win. If an error is expensive and invisible, it is not ready to be automated by anything.

What should stay human

High-stakes, irreversible decisions

Anything that cannot be undone and carries real consequence — sending money, making a legal commitment, deleting data, a final hiring or firing call — should stay with a person. AI can prepare the decision; it should not make it.

Errors that are expensive and hard to catch

Some mistakes are cheap and obvious; others are costly and silent. Where an error would do real damage and would not be noticed until much later, the task needs a human in the loop — not because AI cannot attempt it, but because nothing is checking it.

Work that needs accountability or empathy

Some work needs a named person who is answerable for the outcome, or genuine human empathy in a difficult moment. AI can draft, suggest, and surface options for this work. It should not own it.

The four questions to ask of any task

For each step in a workflow, ask four questions in order. Is it fully deterministic? If yes, it is a rules job. Does it need judgment over messy input? If yes, it might be an AI job. Is an occasional error acceptable or easily caught? If no, a human owns it regardless. And is the volume high enough that automating it is worth the build? If no, leave it manual and spend the effort elsewhere. Four questions, asked honestly, sort almost any workflow.

A worked example: a support inbox

Take a common workflow — handling an incoming support inbox — and run it through the three buckets. A message arrives. Logging it, timestamping it, and acknowledging receipt are deterministic steps with no judgment: rules. Reading the message, working out what it is actually about, and routing it to the right team is judgment over messy, unstructured text: AI. Drafting a first-pass reply from the relevant help articles is also AI — reading and composing, where a mostly-right draft is genuinely useful and a person will review it.

Then the buckets shift. Approving and sending the reply to the customer is a person’s call, because it is the accountable, customer-facing moment. Issuing a refund above a threshold is irreversible and high-stakes: human. Closing the ticket and updating the metrics afterward is deterministic again: rules.

One ordinary workflow, and every bucket appears in it. Notice what the right design looks like — rules carry the structure, AI does the two judgment steps in the middle, a person owns the decisions that carry consequence. Nobody handed the whole inbox to one agent, and nobody left the judgment steps manual. That sorting, step by step, is the actual work of planning an AI workflow.

Why “automate everything” fails

Teams that try to make one AI agent run an entire end-to-end process build something slow, expensive, hard to debug, and impossible to trust. Every step is a model call, every model call can fail in a new way, and a single trace becomes impossible to follow. The workflows that actually hold up in production are mostly rules, with AI dropped in at the specific judgment steps and humans at the irreversible ones. Ambition aimed at the whole process is how AI automation projects quietly fail; ambition aimed at the right two or three steps is how they succeed.

Start with the audit, not the agent

Before building anything, map the workflow step by step and put each step in a bucket. That map — which steps are rules, which are AI, which are human — is the real deliverable of an AI-readiness audit, and it almost always shows that the highest-value AI automation is one or two specific steps rather than the whole process. The audit costs little, takes a couple of weeks, and routinely saves a far larger budget from being spent automating the wrong things.

Common questions