When not to use AI: rules or AI for a business decision
A practical test for choosing rule-based automation or AI, with hybrid patterns where AI drafts and explains while written rules decide.
6 min readBy AI Guru AcademyFacts last checked
By AI Guru Academy
In short
Use rule-based automation when the criteria for a decision can be written down completely, the outcome must be the same every time, and someone must be able to explain exactly why it happened. Use AI when the input is messy, the criteria cannot be fully listed, or the task is to read, draft or summarise. In many real processes the best design is a hybrid: AI prepares or explains, and written rules decide.
How do you decide between rules and AI?
Ask four questions about the decision, in this order. If the answers point to rules, stop there.
- Are the criteria written down, and are they complete? If a person could hand a new colleague a one-page list and get the right answer every time, a rule fits.
- Is there labelled history? Learning from examples needs past cases with known correct outcomes. Without them, an AI approach has nothing reliable to learn or be checked against.
- Must the result be reproducible? The same input should give the same output on every run.
- Must the reason be explainable? A regulator, auditor or customer may ask why.
Rules win on questions 1, 3 and 4. AI earns its place when question 1 is "no, the criteria are fuzzy or the input is free text, images or speech", and when question 2 is "yes, we can test it against known cases".
What does AI do that rules cannot?
AI handles input that does not fit a fixed format. A rule can check whether an invoice total is above a limit. It struggles to read a rambling customer email and work out whether it is a complaint, a cancellation or a question about billing.
Generative AI is also suited to producing language: summaries, first drafts, translations and plain-English explanations. AWS's overview of generative AI lists uses such as summarising data and generating email content and sales scripts. The trade-off is that a person has to review the output before it matters.
Why is reproducibility the deciding factor in many cases?
AWS explains that a model on Amazon Bedrock generates each token by sampling from a probability distribution, so the same prompt can return different text on different runs. AWS documents inference settings such as temperature, top-K and top-P that influence this. AWS says a lower temperature leads to more deterministic responses, but it does not promise identical output on every run.
That is acceptable for a draft email. It is a problem for a decision such as "approve or decline". If two identical applications can receive different answers, you cannot defend the process, and you cannot test it the way you test a rule. A rule that says "decline if the stated income is below the written threshold" gives the same answer every time and can be audited line by line.
Which decisions should stay with rules?
Keep rules in charge of decisions with a legal, financial or safety consequence, where the criteria already exist in a policy. The table shows how common situations usually split. These are illustrations, not recommendations for any specific business.
| Situation | Better fit | Why |
|---|---|---|
| Refund allowed if the order is under 30 days old and unused | Rules | Complete written criteria, must be identical every time |
| Sorting a free-text inbox by topic | AI | Input varies, mistakes are cheap and reviewable |
| Calculating a discount from a price list | Rules | Exact arithmetic, fully explainable |
| Summarising a long contract for a manager | AI, with human review | Language task, no single correct output |
| Deciding who is eligible for a regulated product | Rules, with AI preparing the case file | Needs a stated, auditable reason |
| Answering staff questions from a policy library | AI grounded in the documents | Wording varies, source must be checkable |
Worked example (an invented illustration): a regional insurer wants faster handling of travel-claim emails. Eligibility is a written policy, so rules decide it. The emails are messy, so AI reads each one, pulls out the dates and amounts, and drafts a reply for the handler. The handler sees the AI's extraction next to the original email and the rule outcome. If the AI misreads a date, the rule check or the handler catches it, and the decision is still explainable because only the rules made it.
What do hybrid patterns look like?
The most useful hybrids keep decisions and language separate. Three patterns cover most cases.
- AI drafts, rules decide. The model reads and structures the input; a rule engine applies the policy.
- AI explains, rules decide. The rule produces the outcome, and the model turns the rule's reasons into a plain-language message. The explanation must be checked against the rule's actual output, not trusted on its own.
- AI answers, guardrails constrain. Amazon Bedrock Guardrails can block denied topics in queries and responses, and its contextual grounding check scores a model response against the source material and user query you supply. AWS describes the grounding and relevance thresholds as configurable. These controls reduce risk. They do not turn a model into a deterministic decision-maker.
Where a system takes actions, AWS documents that Amazon Bedrock agents can use tools and action groups. Treat each tool call as a place to apply a written rule, such as a spending limit, before anything irreversible happens. AWS also notes that you can no longer sign up for Amazon Bedrock Agents (now Agents Classic) and points new builders to Amazon Bedrock AgentCore, so check current documentation before designing around it.
What governance questions should you ask first?
Decide who owns the decision, who reviews AI output and what you will record. ISO/IEC 42001 is the standard for an AI management system; AWS's FAQ describes it as specifying requirements for establishing, implementing, maintaining and continually improving one. ISO/IEC 23053 is a framework for AI systems using machine learning, and ISO/IEC 22989 covers AI concepts and terminology. You do not need certification to borrow the idea: write down who approves a use case, what it may and may not decide, and how often it is reviewed.
A short checklist before approving an AI use case:
- The decision criteria are not fully writable, or the input is unstructured.
- A test set of known cases exists to measure accuracy.
- A named person reviews output that has consequences.
- Any decision with a legal or financial effect has a rule or human behind it.
- You can explain, in one paragraph, why a given outcome happened.
Key takeaways
- If the criteria are written, complete and must be applied identically, use rules.
- Use AI for messy input and language tasks, and measure it against known cases.
- Generated text is sampled, so identical prompts can differ; that matters most for decisions.
- Hybrids work best when AI prepares or explains and rules decide.
- Guardrails and grounding checks lower risk but do not guarantee correct output.
- Name an owner and a reviewer before any AI use case goes live.
FAQ
Is rule-based automation old-fashioned compared with AI?
No. For decisions with complete written criteria it is usually cheaper to test, easier to explain and fully repeatable. Many strong systems use rules for the decision and AI only around it.
Can AI make the same decision every time?
Not reliably. Bedrock models sample each token from a probability distribution, so repeats can differ. AWS says a lower temperature makes responses more deterministic, but it does not present that as a guarantee of identical output.
Can a guardrail make an AI answer correct?
No. Amazon Bedrock Guardrails can block denied topics and check responses against supplied source material, which reduces risk. A person or a rule still needs to own decisions that matter.
Do I need a lot of data before using AI?
For a language task using an existing model, you mainly need good source material and test cases to judge the output. For a model that learns from past outcomes, you need labelled history, and without it you should prefer rules.
Sources
- Amazon Bedrock inference parameters: https://docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html
- Amazon Bedrock Guardrails: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html
- Amazon Bedrock Guardrails contextual grounding check: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html
- AWS, What is generative AI?: https://aws.amazon.com/what-is/generative-ai/
- Amazon Bedrock Agents multi-agent collaboration: https://docs.aws.amazon.com/bedrock/latest/userguide/agents-multi-agent-collaboration.html
- AWS ISO/IEC 42001 FAQs: https://aws.amazon.com/compliance/iso-42001-faqs/
- ISO/IEC 42001 (AI management systems): https://www.iso.org/standard/42001
- ISO/IEC 23053 (Framework for AI systems using machine learning): https://www.iso.org/standard/74438.html
- ISO/IEC 22989 (AI concepts and terminology): https://www.iso.org/standard/74296.html
Last checked 2026-11-16.
AI Guru Academy is an independent training provider and is not affiliated with or endorsed by Amazon Web Services. AWS service names are trademarks of Amazon.com, Inc. or its affiliates.
Next step: take the free AIB-C01 diagnostic at /diagnostic, then see the course at /courses/aib-c01-exam-prep.