Symbolic AI
Symbolic AI is the approach that represents knowledge as explicit symbols — facts, rules, logical relations, ontologies — and reasons by manipulating them: search, inference and planning over a model of the world written by people. It was "artificial intelligence" from the 1950s to the 1990s, produced expert systems and theorem provers, and lost to statistical machine learning on perception and language. It survives where explicit rules are the point — policy engines, compliance logic, knowledge graphs.
- Also known as
- Good old-fashioned AI, GOFAI, Rule-based AI, Logic-based AI
- Used at
- rules and policy engines in fintech (credit policy, AML rules, eligibility), knowledge graphs (entity resolution, ownership structures for KYB), planning and constraint solvers (scheduling, optimisation), formal verification, and as the reasoning component in some agent architectures
- Binding
- no standard; but regulators' preference for decisions that can be traced to rules keeps symbolic components in every regulated decision system
- Typical strengths
- transparency (every conclusion has a derivation), precision on well-defined problems, guarantees (a logic engine does not hallucinate), and easy incorporation of policy; typical weaknesses: brittleness outside the encoded knowledge, the cost of writing and maintaining the rules, and inability to handle perception, natural language and the messy statistics of the world
- Typical history
- Logic Theorist (1956), expert systems (1970s–80s), the AI winter when they failed to scale, the statistical turn (1990s), deep learning's dominance (2012 onward), and the 2020s interest in combining learned models with symbolic reasoning, verification and tool use
- Parties
- the researchers of the first decades, the engineers who maintain rule systems today, and the labs experimenting with hybrids
How symbolic AI works in practice
A symbolic system knows things in a form a person wrote: "a customer is high-risk if resident in a listed jurisdiction or a politically exposed person"; "a company's beneficial owner is any natural person holding 25% or more, directly or through chains of ownership". Given facts — this customer, this ownership graph — an inference engine applies the rules and derives conclusions, each with a trace: high-risk because rule 14, via facts 3 and 7. A KYB system tracing beneficial ownership through five holding companies is doing symbolic reasoning over a graph; a credit policy engine is applying logic to an application; a constraint solver assigning payouts to accounts under limits is search over a symbolic model. None of these learns from data; all of them can be read.
The limits are the reason machine learning won the parts of AI that touch the world. No one can write the rules that recognise a face, parse a sentence or detect a fraud pattern in a billion transactions; a model trained on data can. But the learned model cannot say why, cannot be guaranteed, and cannot be told policy in plain terms — which is why production systems combine them: the model perceives and scores, the symbolic layer decides within rules, and the trace the regulator wants comes from the second. The 2020s' agents make the pairing explicit — a language model that calls a calculator, a database, a rules engine or a verifier rather than guessing — and "neuro-symbolic" is the research name for doing it well.
For a fintech, the question is never symbolic or statistical; it is which decisions must be rules, which can be models, and how the two are joined and audited.
Example
A fintech's onboarding decision engine is symbolic: 200 rules over the applicant's data and the results of vendor checks, producing accept, decline or review with a full trace, reviewed by compliance and read by the regulator. Its fraud engine is statistical: a model scoring device, behaviour and history. The two meet in the decision: the rules require review above a model score, forbid onboarding on certain facts regardless of score, and log which fired. When a new fraud typology appears, the model catches it in days and a rule is written to make the response policy; when the regulator asks why an applicant was declined, the trace answers in the rules' language. An engineer proposes replacing the rules with a model "for accuracy"; compliance asks how the model would explain a decline; the rules stay.
Common mistakes
- Trying to write rules for a problem only data can solve — perception, language, pattern detection at scale. - Replacing policy rules with a model and losing the trace the regulator requires. - Treating symbolic and statistical AI as rivals in a system that needs both. - Rule systems with no ownership, which grow until nobody understands them. - Reading "AI" in a vendor's pitch without asking which kind.