In December 2023, a barrister in England and Wales was fined £2,000 after submitting court arguments citing legal authorities that did not exist. ChatGPT had generated them. The barrister trusted them without checking.
The case illustrates a precise and consequential confusion: treating a tool that predicts text as if it applies explicit rules. The two work differently, produce different kinds of outputs, and carry different levels of risk. For any founder using AI tools in a service business, knowing which one you are dealing with is one of the most practical distinctions you can develop.
What is the difference between AI “thinking” and reasoning?
Pattern-predicting AI tools, including ChatGPT, Microsoft Copilot, and the AI features inside many SaaS platforms, work by predicting the most likely next word or output based on patterns in training data. They do not apply rules, check facts, or follow an explicit logical chain. Rules-based systems, by contrast, apply criteria step by step, and every decision traces back to a specific rule.
TechUK’s paper on AI reasoning describes the difference with useful precision: language models “generate answers by predicting likely sequences of words based on patterns in training data”, which differs fundamentally from symbolic systems that “apply explicit rules and record each reasoning step for audit”. The same paper notes that rules-based systems grounded in verified facts can eliminate hallucinations entirely, because the answer comes from the rules rather than from statistical approximation.
This matters practically. A word processor completing your sentence and a compliance checklist deciding client eligibility are both doing something with information. They are doing different things, and the difference is load-bearing when the output has consequences. The UK Government’s AI Playbook categorises generative models as systems trained to “predict, summarise and generate content”, drawing a clear line between that function and the explicit rule-application work that high-stakes decisions require.
Why does this distinction matter for your business?
The distinction matters most when a decision has consequences. For low-stakes content work, including drafting proposals, summarising meeting notes, or writing job adverts, pattern-predicting AI saves real time and the cost of a near-miss is low. For decisions affecting a client’s money, rights, or regulatory status, a tool that guesses confidently rather than reasons accurately is a liability.
UK regulators have been explicit about this. The ICO’s guidance on AI and data protection requires that, where AI supports significant decisions about individuals, organisations must be able to “provide meaningful information about the logic involved”. The FCA’s Consumer Duty, in force since July 2023, requires firms to evidence fair outcomes from their decisions. The SRA’s 2024 Risk Outlook on AI in legal practice warns specifically about hallucinated citations and fabricated references, naming exactly the failure mode that ended the barrister’s case.
An ONS study published in 2025 found that difficulty identifying which activities to apply AI to was the most common adoption barrier among UK businesses, cited by 39% of respondents, ahead of cost and lack of skills. That confusion is partly a downstream effect of this distinction. When founders cannot tell the tool that completes sentences apart from the system that applies rules, they either overuse the first or dismiss the second entirely.
Where will you actually meet these two types of AI?
Pattern-predicting AI is already inside many of the tools you use. ChatGPT, Microsoft Copilot, Google Gemini, and the AI features embedded in email, CRM, and project management platforms are all pattern-based. Rules-based systems are less visible: they power the compliance logic in regulated sectors, eligibility checks, and the transaction-monitoring engines used by financial services firms.
For an owner-managed business, this means you will encounter pattern-predicting tools constantly, often without a clear label. The AI that suggests your next email reply, summarises a contract, or categorises incoming enquiries is pattern-based. The system that checks client eligibility, flags suspicious transactions, or routes a complaint through a regulatory workflow is typically rules-based, even when it carries AI branding.
In many enterprise tools, both types sit together. TechUK describes hybrid architectures in which a language model handles language and conversation while a rules engine makes any decision requiring traceable logic. For many owner-managed firms, this architecture is invisible from the outside, because vendors wrap both components inside a single product name. The practical skill is knowing to ask which layer a particular feature uses before you rely on it for anything consequential.
When should you use which, and when should you step back?
The working test is straightforward: use pattern-predicting AI for tasks where a near-miss is acceptable and a human reviews the output before it has any effect. Apply rules-based or structured decision logic where the decision must be correct, consistent, and explainable. The practical check is whether, if a client or regulator asked you to trace how a decision was reached, you could do it.
Pattern-predicting AI works well for drafting and content creation, generating options and alternatives, summarising long documents for internal use, and triaging incoming messages before a person reads them. The NCSC’s guidance on secure AI use describes these tools as “untrusted assistants”: their outputs need to be verified, particularly where decisions might affect customers, systems, or access rights.
Rules-based logic is better suited to anything where UK law or sector regulation creates a duty to explain. Eligibility assessments, affordability checks, and complaints handling all fall into this category. The Equality Act 2010 applies regardless of whether a discriminatory outcome came from a human or an algorithm. The FCA’s Consumer Duty requires evidence of fair outcomes, not good intentions.
The UK Government’s AI Playbook is direct on this point: be “open to the conclusion that AI is not the best solution” where a problem can be solved more easily by established technology or straightforward process improvements. A well-designed decision table, testable and version-controlled, often beats a confident language model output.
What related concepts are worth knowing about?
Three developments are narrowing the gap between pattern prediction and explicit reasoning. Chain-of-thought prompting asks a language model to work through a problem step by step, which improves accuracy on structured tasks. Specialised reasoning models are designed to produce verifiable step-by-step logic. Hybrid architectures wrap a language model inside a rules engine to extract the benefits of both in the same product.
None of this closes the gap entirely, at least not yet. The EU AI Act, agreed in March 2024, treats uses including credit scoring, recruitment, and certain public-services decisions as “high-risk” and imposes risk management, data governance, transparency, and audit requirements regardless of the underlying technique. The principle is clear: consequential decisions require explainability, and explainability requires that a trace exists.
UK regulators are taking a contextual, sector-led approach rather than applying a single framework. The FCA, ICO, CMA, and SRA are all extending existing obligations around fairness, accountability, and consumer protection to cover AI-assisted decisions. As a UK SME founder, you sit within this regulatory environment whether you think of yourself as deploying AI or not.
The practical anchor for now: treat pattern-predicting AI as a fast and useful assistant, keep explicit rules in play for anything that must hold up to scrutiny, and ask of every AI feature the same question: “Can I trace how it reached its conclusion?” A tool that cannot answer that question still has real value for content and productivity work. For decisions that must hold up to scrutiny, structured rules with a traceable output are the right choice.
