Two ways to automate
Automation means letting software do a task that a person would otherwise do by hand. There are two broad ways to build it.
Rules-based automation follows instructions written by a person. If a condition is met, do this; otherwise do that. “If an expense is under a set amount and has a receipt attached, approve it.” “If a form is submitted, send a confirmation email and add a row to the spreadsheet.”
AI automation relies on a system that learned patterns from examples. Rather than being told exactly how to handle every case, it estimates what to do based on similarity to cases it has seen. “Read this email and decide which team should handle it.” “Pull the total from this receipt, whatever its layout.”
Many real workflows combine both: rules for the parts that are clear, and AI for the parts that involve messy input.
How rules-based automation behaves
Rules are explicit, so behavior is predictable. Given the same input, you get the same output every time. If something goes wrong, you can read the rule and see why. Typical features:
- Transparent. Anyone can inspect the logic and explain a result.
- Consistent. No variation between runs.
- Easy to test. You can list the cases and check each one, as described in testing a workflow before you trust it.
- Low running cost. Usually inexpensive to operate.
- Rigid. It only handles what someone anticipated. A new case, an unexpected format or a typo in a key field can cause it to fail or skip.
Much of the everyday tooling in this space, such as triggers and actions between apps and no-code platforms, is rules-based. Robotic process automation mostly follows scripted steps too, though it can be paired with AI components.
How AI automation behaves
Learned systems cope better with variety. They can read an invoice they have never seen before, understand that “can’t log in” and “password not working” mean the same thing, or sort photos into categories. Typical features:
- Flexible. Handles varied, messy input such as free text, scans and audio.
- Probabilistic. The output is a best guess, not a certainty. Some cases will be wrong.
- Harder to explain. It may be unclear why a particular decision was made.
- Variable. Some systems give different answers to the same input, and behavior can shift when the vendor updates the model.
- Dependent on data. Quality hinges on what it learned from, as covered in training data, and may degrade as conditions change.
- More setup and oversight. It needs testing on real examples and ongoing monitoring.
The foundations are explained in the article on machine learning, and the text-generating variety in how language models generate text.
A side-by-side comparison
| Feature | Rules-based | AI-based |
|---|---|---|
| Input it handles best | Structured, predictable | Messy, varied |
| Result for same input | Identical | Usually similar, may vary |
| Explaining a decision | Point to the rule | Often difficult |
| When it meets something new | Fails or skips | Makes a guess |
| Typical error | Missed case or broken step | Plausible but wrong output |
| Maintenance | Update rules by hand | Monitor accuracy, retrain or adjust |
When rules are the better fit
Rules usually win when:
- the logic is clear and stable
- the data is structured, such as fields in a form
- exact, repeatable results are required
- you must be able to explain and audit every decision
- the cost of an error is high and a guess is unacceptable
Examples: calculating a discount, applying a fixed approval threshold, copying values between systems, sending scheduled reminders, or enforcing a compliance checklist. Many cases of scheduling automation and approval workflows fall here.
When AI is the better fit
AI usually makes sense when:
- the input is unstructured, such as free-form text, images or speech
- the rules would be too many or too fuzzy to write
- an occasional error is tolerable, or caught by a person
- volume is high enough that manual handling is costly
Examples: sorting incoming messages by topic, extracting fields from documents with varied layouts, transcribing calls, or drafting first replies. These are the territory of natural language processing and speech recognition.
Combining the two
A common and sensible pattern is a hybrid: use AI to interpret messy input, then rules to act on the result. For instance, an AI step reads an invoice and extracts the amount and supplier. Rules then check the amount against a limit, compare it with the purchase order, and route it for approval. A confidence threshold can send uncertain extractions to a person. This keeps the predictable parts predictable and limits the reach of AI mistakes.
Failure modes to plan for
Rules-based workflows tend to break loudly or silently when something changes: a renamed field, a new product code, an app update. Common causes are listed in common reasons automations break. AI workflows tend to break by being quietly wrong, which can be harder to notice. In both cases, monitoring, a clear owner and automation documentation help.
For AI steps, add human review where the stakes warrant it and decide in advance how mistakes are reported and corrected.
A practical way to decide
- Write out the task step by step. Can you state the rule for every step?
- If yes, build it with rules. It will probably be cheaper and easier to trust.
- If a step involves interpretation of messy input, consider an AI component for just that step.
- Test on real examples, including awkward ones.
- Estimate the total cost, including oversight time, not just the license.
Adding AI does not make an automation better by default. It adds capability and uncertainty together. Match the tool to the task, and keep results checkable.