If every input looks the same and every output follows a fixed rule, you probably do not need AI. You need automation.

If the input is messy—emails, notes, PDFs, photos, free-text forms—or the work requires summarizing, classifying, extracting, or drafting, AI may be useful. And if the task depends on consequential judgment, empathy, negotiation, or accountability, a person may still need to own the decision.

The mistake is treating those as competing technologies. A strong workflow often uses all three: normal software for predictable steps, AI for the fuzzy middle, and a person for the decision that matters.

The Minimum Intelligence Principle

Use the least intelligent system that can do the job reliably. Fixed rules are easier to test than AI. AI is useful when rules break down. Human judgment belongs where the cost of being wrong is too high or the decision cannot be reduced to a stable pattern.

The four levels of a business workflow

Before choosing a tool, put each step of the workflow into one of four buckets.

1. Manual work

A person does the task from beginning to end. This is not automatically bad. Manual work is appropriate when the task is rare, changes constantly, or depends on context that is difficult to formalize.

2. Rules-based automation

The system follows a predictable condition: when this happens, do that. Examples include creating a CRM record after a form submission, sending an invoice reminder after 30 days, renaming a file, or notifying a teammate when a status changes.

These steps usually do not benefit from AI. Adding AI can make a deterministic process slower, more expensive, and harder to debug.

3. AI-assisted automation

AI handles the part that is difficult to express as a fixed rule: reading an email, summarizing a conversation, extracting details from messy notes, classifying free text, or creating a first draft.

The surrounding workflow can still be deterministic. A form arrives, the AI extracts the project type, a normal automation creates the record, and a person reviews the result before anything customer-facing happens.

4. Human judgment

Some steps should remain owned by a person because the decision carries meaningful consequences or depends on context that cannot be safely reduced to a model output. Examples may include approving a large refund, making an employment decision, accepting legal obligations, or handling a sensitive customer escalation.

AI may prepare information for those decisions. It should not automatically inherit the authority of the person making them.

The four-question decision test

For each step in a workflow, ask these questions in order.

Question 1: Is the input predictable?

  • Yes: try rules-based automation first.
  • No: AI may help interpret the input.

Question 2: Can the correct output be described clearly?

  • Yes: the step may be testable enough for automation or AI assistance.
  • No: keep a person closer to the work until “good” can be defined.

Question 3: Can a mistake be caught before it matters?

  • Yes: a reviewed AI step may be appropriate.
  • No: avoid using an unproven automated decision.

Question 4: Is the step valuable enough to automate?

  • Yes: test the simplest workable approach.
  • No: leave it alone. Not every manual task deserves software.

That last question is important. A technically automatable task that takes five minutes per month can still be a bad project.

Six examples: automation, AI, or human?

Example 1: Website form → CRM

A customer submits name, phone number, email, ZIP code, and service type. Those fields already have structure.

Best starting point: rules-based automation. Copy the fields directly. AI adds little value.

Example 2: Free-text project description → service category

The same form also asks, “Tell us what you need,” and customers describe the job in their own words.

Best starting point: AI assistance. Use AI to suggest a category or summarize the description, then pass the structured result into the normal workflow.

Example 3: Overdue invoice reminder

If an invoice is unpaid after a known number of days, send an approved reminder or create a task for the account owner.

Best starting point: rules-based automation. The trigger is objective.

Example 4: Customer follow-up after a complicated call

The follow-up needs to reflect what was discussed, organize the next steps, and preserve the company's tone.

Best starting point: AI-assisted draft plus human review. The source material is unstructured, but the final commitment should still be checked.

Example 5: Approving a large refund

The facts can be gathered automatically, but the decision may depend on customer history, policy exceptions, fraud risk, and commercial judgment.

Best starting point: automation and AI can prepare the evidence; a person owns the decision.

Example 6: Scheduling

If customers choose from open calendar slots, normal scheduling software is enough. If they send messages like “any afternoon except Tuesday after the inspection,” AI may help interpret the request before a scheduling rule finds valid options.

Best starting point: hybrid. AI interprets; deterministic software checks the calendar and books.

Why unnecessary AI creates an “AI tax”

AI is powerful precisely because it can handle ambiguity. That flexibility also introduces costs that fixed rules do not have.

The AI tax

  • Uncertainty: the same kind of input can produce different quality across cases.
  • Review: someone may need to verify facts or decisions that a normal rule would execute exactly.
  • Data exposure: another service may receive business information that never needed to leave the original system.
  • Maintenance: prompts, reference material, model behavior, and integrations may need ongoing attention.
  • Cost: usage, subscriptions, and review time can exceed the value of the task.

None of those make AI a bad choice. They are reasons to reserve it for the parts of the workflow where its flexibility creates enough value to justify the extra complexity.

The best architecture is often hybrid

Consider an inbound lead for a home-service company. The workflow might be designed like this:

  1. Automation: receive the form and create a lead record.
  2. AI: summarize the free-text request and suggest a service category.
  3. Automation: route the lead based on territory and service category.
  4. AI: draft an acknowledgment using approved information.
  5. Human: review the message if the request is unusual or high value.
  6. Automation: record the activity and set a follow-up reminder.

Calling that an “AI agent” may sound more exciting. Breaking it into explicit responsibilities makes it easier to test, debug, and control.

Design rule

Let deterministic software enforce facts, states, permissions, and timing. Let AI interpret messy information. Let people own consequential judgment.

How to spot a workflow that is being over-AI'd

Watch for these warning signs:

When those appear, simplify. Sometimes the best AI improvement is removing AI from half the workflow.

A 15-minute workflow split

Take one recurring task and write each step on a separate line. Then label every step with one of four letters:

M / R / A / H

  • M — Manual: keep it manual for now.
  • R — Rule: deterministic automation can handle it.
  • A — AI: unstructured information or language makes AI potentially useful.
  • H — Human: judgment or consequence requires a person to own the decision.

Now look at the sequence. Your first pilot should usually automate a few obvious R steps, add AI to one clearly defined A step, and preserve H where it matters.

If the whole page is marked A, the workflow is probably too broad. If everything is H, the process may not be ready. If most steps are R, you may have found an automation project rather than an AI project—and that is a perfectly good result.

What to do next

If you have a repetitive task but have not chosen the first candidate, use our Automation Fit Test. If you already have a workflow in mind, score it with the 20-point AI readiness checklist. Only after those steps should you start comparing products with the AI tool buying scorecard.

That order—problem, readiness, tool—is intentionally boring. It is also much less likely to leave you with another subscription and no measurable improvement.