Contents
Organizing business rules into an ontology alone does not make agent decisions reliable. Business and technical teams need to identify and restore meaning lost when rules become execution conditions.
Key Points#
- An ontology describes business structures; agent instructions define who does what.
- Reducing ambiguous business language to simple Boolean values can cause incorrect decisions in edge cases.
- Add context and precedents to uncertain conditions, and hand cases over to people based on confidence.
- Record human decisions and continually review conditions and explanations.
- Business and technical staff should review the translated instructions together.
Description and Execution Speak Different Languages#
An ontology organizes the entities, relationships, and constraints in a business domain. For example, the rule “Returns are allowed within seven days of purchase if the product has not been substantially used” describes business conditions.
Running an agent also requires defining the actor, actions, and triggering conditions. While CBC (Context, Behavior, Constraint) describes “what is allowed in which circumstances,” ABC (Actor, Behavior, Condition) defines “who performs what under which conditions.”
Meaning can be lost in this conversion. Translating “within seven days” into days_since_purchase <= 7 makes it machine-evaluable. But that expression alone does not determine whether to use the request time or processing time, or how to count the days.
Edge Cases Reveal Mismatches#
Suppose a return request is submitted at 11:58 p.m. on the sixth day after purchase, and the system processes it at 12:02 a.m. the following day. Using the processing time to count the days could make the request ineligible. If the business intends to use the request time, an agent implementing only the expression may evaluate the condition correctly in technical terms yet return a result that is wrong for the business.
“Not substantially used” is also difficult to assess through simple classification. Do light marks from inspection count as use, or are they acceptable as part of a routine check? Such expressions carry context about the rule and how exceptions should be handled.
Preserve Ambiguity During Conversion#
Rather than forcing business language into a single number or Boolean value, make clear where interpretation is needed. The conversion can proceed in three stages.
- Formalize the conditions. Define a “large order” as “an order exceeding three times the historical average for the same category,” and flag uncertainty where the definition is unclear.
- Add the reasoning behind decisions. Provide explanations such as “Assess substantial wear, excluding minor marks from routine inspection,” along with past decisions, as a SemanticHint.
- Route cases based on confidence. For example, if confidence in a decision is below 0.85, send it for human review instead of processing it automatically.
These numbers are design examples, not ready-to-use thresholds. Adjust them for each business process based on the impact of incorrect decisions and the volume of cases people can review.
Use Precedents to Inform Future Decisions#
When sending a case for human review, record both the outcome and the reasoning. When similar requests recur, referring to past decisions can help improve SemanticHints and conditions.
Records should include the input context, the agent’s decision and confidence, the human decision, and its rationale. Review examples regularly to distinguish gaps in rules from insufficient explanations or errors in formalizing conditions. Do not expand automated decision-making simply because more precedents have accumulated; check the impact of incorrect decisions before making changes.
Review Instructions Together#
Business staff explain the intent of rules, exceptions, and past decisions. Technical staff translate these into conditions and explanations for the agent. Leaving this to either side alone makes it easy to overlook cases that “satisfy the written conditions but conflict with business intent.”
Start by choosing one workflow and reviewing the converted conditions and SemanticHints together. Prepare several edge cases and decide which conditions allow automatic processing and when to hand cases over to people.
Next Step#
Choose one small workflow, such as handling returns, and summarize its ambiguous language, decision conditions, and criteria for human handoff on a single page for business and technical staff to review together.
Source: https://www.diginfo.me/ontology-agents-paradigm-shift
Contents
