AI or automation? How to choose for a business workflow

Three approaches, three different responsibilities

Conventional automation follows explicit rules. An AI workflow uses a model within a process whose steps are defined in code. An agent lets the model choose steps and tools as the task progresses. These can coexist: a model may interpret a request while ordinary code validates fields and controls the final action.

Start with fixed rules when you can

A form with required fields can create a CRM task, assign an owner by region and send an internal notification without a language model. Write the decision rules first. If a person can specify every branch and the inputs are structured, test that implementation before adding AI. Repeated failures caused by missing permissions or inconsistent identifiers need data and integration work.

Add AI where interpretation is the difficult part

An email may describe the same request in many ways. A model can propose a category or extract details for review. Keep an explicit unknown category and validate required fields before acting. If the task needs company knowledge, retrieval can supply approved source material; access control and source freshness still need their own rules.

Use an agent only when the path needs to change

Consider an agent for a bounded task that requires choosing the next investigation from previous findings. Define available tools, stopping conditions and who handles unfinished work. More autonomy creates more paths to evaluate. Do not replace an understood sequence with an open loop simply because the interface looks more flexible.

Worked example: a customer document request

This fictional example shows how the approaches can share one process. A customer asks for the latest approved document. A model interprets the wording, but the record identifier and access checks come from controlled systems. Missing or contradictory information stops the action and creates a review task. No performance or customer results are implied.

• Fixed rules: match the customer record, verify permissions and prevent duplicate sends.
• AI step: interpret the request and prepare a draft with references to the approved sources.
• Human decision: resolve conflicting documents and approve any action outside the agreed scope.

Compare the cost of a completed task

Include integration work, model usage, review time, corrections and ongoing maintenance. A fast draft may still need a slow review. Measure these costs on the same representative task set for each candidate approach. Track useful completions separately from attempts and agree which errors make a release unacceptable before reviewing the results.

A decision checklist for the first release

Keep the decision small enough to reverse. Run the chosen approach on an authorized sample, inspect exceptions and retain a manual path until the operator can manage failures. Expand only after reviewing the complete task against the agreed acceptance criteria.

• Can fixed rules handle the task with the data we already have?
• Which step requires interpretation, and how will we check its output?
• Who owns exceptions, access changes and the decision to stop?