Blog · · Jiber Team
Rules or AI? Why the best process agents use both.
Rules-plus-AI design is an approach to process agents where AI models read unstructured inputs, deterministic rules make the decisions and people approve irreversible steps.
TL;DR: The most reliable process agents use both: AI reads unstructured inputs, fixed business rules make the decisions, and people sign off on anything irreversible. Neither rules alone nor AI alone is enough for most real business workflows.
Why is "rules or AI" the wrong question?
When teams first look at automating a process, the question is often framed as a choice: build a rules engine, or use AI. In practice, most business processes contain both kinds of work. Some steps need an exact, repeatable answer. Others start with an input that no rule could parse on its own.
A rules-only system breaks on the first résumé in an unusual format. An AI-only system can give an answer that cannot be explained to an auditor. Jiber is an AI company that builds process agents for businesses, and every agent we design combines the two deliberately.
Where do fixed business rules work best?
Rules are the right tool wherever the outcome must be exact and explainable. A rule either fires or it does not, and anyone can read why. Typical examples:
- Eligibility checks and thresholds, such as whether an applicant meets a minimum requirement
- Mapping figures to the correct line of a tax form
- Approval limits, for example any amount above a set value needs a manager
- Validation, such as totals that must match a source document
Because rules are deterministic, the people who own the policy can test and review them.
Where does AI reasoning work best?
AI models are strongest wherever the input is unstructured. Their job is to turn messy input into structured fields that rules can act on. Common cases include:
- Reading a scanned statement or a photographed receipt
- Extracting skills and experience from a résumé in any layout
- Understanding a free-text enquiry from a property buyer
- Summarising a long email thread into the facts that matter
Treat AI output as a proposal. Low-confidence fields can be routed to a person instead of straight to the rules.
How do rules and AI work together in one agent?
A useful design pattern is "AI to read, rules to decide, people to approve." In a Jiber agent, the flow usually looks like this:
- AI reads each input and extracts structured fields.
- Fixed business rules check those fields and make the routine decisions.
- Exceptions, such as a missing value or an amount over a limit, are flagged.
- A person reviews and signs off at the checkpoints you choose.
- Every step is written to an audit log.
In Form 1120 tax preparation, for example, AI reads client documents and rules map figures to return lines before a preparer reviews the draft. The same split applies to IT consulting recruiting and real estate lead handling.
Why does this split make agents easier to test and trust?
Rules can be unit-tested against known cases, model outputs checked against sample inputs, and approvals act as a safety net while confidence builds.
If a result is wrong, the audit log shows whether the model misread an input or a rule misfired. See our platform page for more.
Frequently asked questions
Can an AI agent work without any fixed rules?
It can, but for business processes that need exact, explainable outcomes, relying on AI alone makes results harder to test and audit. Most reliable agents use rules for decisions and AI for reading inputs.
Are rules-based systems outdated?
No. Rules remain the clearest way to encode policy, thresholds and limits. What has changed is that AI can now prepare messy inputs so that rules can act on them.
Where should a person stay involved?
A person should approve irreversible steps, such as sending external messages, filing documents or committing money, and should review any case the agent flags as an exception.
Want to see how rules and AI would split your own workflow? Book a demo and we will map one of your processes with you.