A rule-based AI system that encodes domain expertise as explicit if-then rules, allowing it to make decisions within a specific domain without requiring training data or statistical learning. For agencies, expert systems remain relevant in content governance, brand safety filtering, and compliance automation where the rules are fully known, stable, and need to be auditable.
Also known as rule-based AI, knowledge-based system, rule engine
An expert system consists of two components: a knowledge base containing a set of domain-specific if-then rules encoded by human experts, and an inference engine that applies those rules to new inputs to derive conclusions. Given an input, the inference engine scans the rule set, finds all rules whose conditions are satisfied by the input, and executes the associated actions, which may include classifying the input, flagging it for review, or chaining to additional rules. The process is transparent: every decision can be traced back to the specific rules that fired and the input conditions that triggered them.
Expert systems represent the pre-machine-learning paradigm of AI. From the 1970s through the 1990s, they were the dominant approach to applied AI in medical diagnosis, financial analysis, and industrial process control. They fell out of favor as statistical machine learning emerged, because machine learning systems could be trained from data rather than requiring laborious manual rule encoding, and because they generalized to inputs that rules had not anticipated. The knowledge acquisition bottleneck, encoding enough expert rules to cover a complex domain, was a fundamental limitation that learning-based systems bypassed entirely.
The return of interest in expert systems is driven by the strengths that machine learning systems lack: complete transparency, deterministic behavior, and immediate auditability. A neural network that classifies content as brand-safe or brand-unsafe cannot explain which specific properties of the content drove the classification. A rule-based system that makes the same classification can list every rule that fired. In regulated environments or anywhere that an AI decision must be explainable and contestable, this distinction matters.
Agencies operate under brand guidelines, content policies, legal constraints, and client-specific approval rules that are often fully articulable as explicit logic. A working ad agency building content governance, brand safety, or compliance automation does not always need a model that learns from data: sometimes it needs a system that reliably applies known rules and can prove it did so. Expert systems are the right tool for that task, and reaching for machine learning where a rule engine will do adds unnecessary opacity and maintenance burden.
Brand safety filtering is a rule-appropriate task in many cases. A client’s brand safety requirements often reduce to explicit rules: do not appear adjacent to content mentioning these topic categories, do not run on domains with this content rating, do not associate with content containing these specific terms. A rule engine applying these criteria is fully auditable, produces no false negatives on explicitly prohibited categories, and can be updated immediately when the client changes the rules. A machine learning classifier that approximates the same criteria may miss edge cases in the explicit categories and cannot be explained when it makes an error.
Compliance automation in regulated industries requires auditability. Financial services and pharmaceutical clients who run advertising under regulatory guidelines need to be able to demonstrate to auditors exactly why specific content was approved or rejected. A machine learning system that cannot produce a rule-based explanation for its decisions is not a viable architecture for this use case. Expert systems that log every rule that fired for every decision provide the audit trail that regulators require.
Hybrid architectures combine the strengths of both approaches. The most effective content governance systems often use an expert system to apply deterministic rules for known, explicit violations and a machine learning model to identify potential violations in the ambiguous middle ground that rules cannot fully specify. Knowing how and when to combine the two approaches is a design judgment that requires understanding what each does well.
An agency manages digital advertising for a pharmaceutical client whose campaigns must comply with FDA guidelines prohibiting the promotion of unapproved uses of prescription drugs. An initial proposal to use a machine learning classifier to screen ad copy for compliance issues is rejected by the client’s regulatory affairs team because the classifier cannot explain its rejections in terms that map to specific regulatory provisions. The agency builds a rule-based compliance screening system that encodes 47 specific regulatory rules provided by the client’s regulatory team, each tied to the relevant FDA guidance document section. Every piece of ad copy that enters the review workflow is evaluated against all 47 rules, and any that fail are flagged with the specific rule text and the provision it references. The system produces a complete audit log for every piece of content reviewed, which satisfies the regulatory team’s documentation requirements and reduces manual compliance review time by 64%.
The automations and agents module of the workshop covers how to build AI workflows that combine rule-based and learning-based approaches, matching the architecture to the requirements rather than defaulting to machine learning everywhere.