Automation or AI: which do you need?
Fixed rules, a custom workflow, or interpretation? Three questions to help you choose the right technology.
AT A GLANCE
Use automation to execute fixed rules. Consider a custom web system for distinct roles and workflows. Explore AI when varied content needs interpretation. They can work together: AI proposes, rules validate, and a person decides on uncertain cases.
- A fixed rule does not require a language model.
- Keep AI work narrow and checkable.
- Design human decision points in advance.
Can you state the rule in advance?
If a paid invoice needs a “settled” status, a clear rule can do the job. The same applies to deadline notifications and transfers of validated fields. Simple automation is easy to test and can explain why a particular action happened.
Separate collecting data from making a decision. A notification rule may be simple while its source information sits in three different systems. In that case, the first requirement is an integration and an agreed source of truth. AI does not replace a missing connection or unclear data ownership.
Does the workflow need its own interface?
Work involving different roles, approvals, documents, and case states may primarily need a shared system. A customer portal can show what to submit, what is missing, and who owns the next step. This is mostly workflow and interface design.
Automated steps can support the portal. The user still needs to understand what happened, what to do next, and where to get help. Test usability with a real task. If a colleague must keep the old spreadsheet open to finish it, something is probably missing from the proposed workflow.
Does the input require interpretation?
Identifying the topic of a free-text email differs from checking a checkbox. AI may help because people describe similar requests in many ways. Its result can still be wrong or uncertain, so define the expected output and review method before implementation.
Start with a narrow set of categories and representative examples. Include a “cannot determine” route and test difficult cases too. If an incorrect classification affects a customer or an important deadline, add approval. A fluent explanation does not prove a result is factually correct. The review process matters as much as the model choice.
An illustrative combined workflow
Imagine a shared support inbox. The system receives a message, removes information unnecessary for the task, asks AI to suggest a category, and validates that category against an approved list. A colleague sees the original request and the suggestion in an internal workspace. This describes an illustrative design, not a client result.
The model does not receive unlimited authority. Message content is data, not an instruction to the internal system. A suggestion should not automatically trigger payments, permission changes, or sensitive transfers. The application must enforce access rights independently of what the model returns.
Compare before you commit
Try the rule-based and AI-assisted approaches on the same examples. Observe where each fails, how much review it needs, and what operation costs. A few well-designed rules may already cover much of the task. Keep difficult examples as a repeatable check for future changes.
The useful choice might be pure automation, a custom system, or a combination. What matters is clear ownership, visible status, and a checkable outcome. Choose technology according to what makes the work more dependable. A smaller system that colleagues understand is often easier to maintain than an ambitious one whose behaviour nobody can explain.
A PRACTICAL TEMPLATE
Two inputs, two different decisions
A paid status from an exact field can be handled by a rule. The topic of a free-text email can only be suggested: allowed categories and human review constrain the AI output. Input variability and the consequence of error determine the choice.
- Exact field → rule
- Varied text → suggestion
- Validation
- Human decision when uncertain
Sources and further reading
- API Security ProjectOWASP · Checked: 2026-09-15
Borbély Kristóf
The developer behind Kriside, writing practical guides about web systems, automation, and AI solutions.
