Machine learning, AI and robotics: formula sheet

Key formulas

AI, ML and deep learning hierarchy

Deep learning ⊂ Machine learning ⊂ Artificial intelligence

Each is a subset of the one before. Use this to answer 'what is the difference' questions.

Rule-based vs machine learning

Rule-based: rules + data → answers. Machine learning: data + answers → rules

In supervised ML the system works out the rules from examples. In rule-based systems people write them.

Supervised learning

Labelled data → predict a known outcome

Used for classification (default or not) and prediction (a value, such as sales).

Unsupervised learning

Unlabelled data → find hidden groups or patterns

Used for segmentation and anomaly detection. No right answer is supplied.

Reinforcement learning

Action → reward or penalty → improved action

Learning by trial and error to maximise reward.

Suitability test for RPA

Rule-based + repetitive + high volume + structured digital data + stable process = good RPA candidate

A checklist, not a calculation. If judgement or unstructured data is needed, consider AI or keep humans.

Simple cost comparison

Annual saving = (Cost of manual process per year) − (Bot running cost per year)

Compare against the one-off set-up cost to judge payback. Include licences, maintenance and oversight, not just licence fees.

RPA vs AI

RPA = follows fixed rules; AI = learns and adapts

Use this to answer 'difference' questions in one line, then add an example of each.

ACCA fundamental principles

Integrity, Objectivity, Professional competence and due care, Confidentiality, Professional behaviour

Use these to frame AI dilemmas for accountants. Name the principle under threat and explain why.

Ethical threat types

Self-interest, Self-review, Advocacy, Familiarity, Intimidation

Over-reliance on an AI tool you built or sold can create self-review or self-interest threats. Pick only those that fit the facts.

Response to a threat

Identify threat → Evaluate significance → Apply safeguards → If not reduced to an acceptable level, decline or withdraw

Gives a safe structure for any ethics requirement.

AI governance checklist

Ownership, Policy, Data quality, Bias testing, Human oversight, Transparency, Monitoring

A memory list for recommendations. Select the items that match the scenario.

Quick revision

• AI performs tasks needing human-like judgement; machine learning learns patterns from data rather than fixed rules.
• Output quality depends on training data: poor or biased data gives poor or biased results.
• RPA mimics rule-based human actions in software; it does not learn unless combined with AI.
• Robotics refers to physical machines; RPA refers to software bots. Do not mix them up.
• Typical benefits: speed, consistency, lower cost per transaction, better forecasting, 24-hour service.
• Typical limits: needs quality data, can be a black box, struggles with unusual situations, and needs human oversight.
• Implementation challenges include cost, skills gaps, system integration, change resistance and unclear benefits.
• Key risks: bias, privacy breaches, cyber attacks, over-reliance and job displacement.
• Humans stay accountable for AI-assisted decisions; management cannot blame the algorithm.
• Explainability and transparency support trust, challenge and good governance.
• Apply ACCA's fundamental principles, especially integrity, objectivity, professional competence and confidentiality.
• Always recommend: weigh benefits against risks, then give a clear, justified conclusion.

Common mistakes

• Treating AI, machine learning and deep learning as the same thing. Fix: State the hierarchy: deep learning is part of machine learning, which is part of AI. Say it in one line when the question asks for the difference.
• Mixing up supervised and unsupervised learning. Fix: Ask one question: does the data include the known answer? If yes, it is supervised. If the system must find groups itself, it is unsupervised.
• Listing generic AI uses with no link to the scenario. Fix: Tie every use to a fact in the scenario, such as the number of transactions or the type of customer.
• Describing what AI is instead of how it creates value. Fix: Spend one line on the definition at most. Spend the rest on applications, benefits and impact on the organisation.
• Treating RPA and AI as the same thing. Fix: State that RPA follows fixed rules, while AI learns and handles judgement or unstructured data. Give one example of each.
• Assuming RPA means physical robots. Fix: Say RPA is software that mimics user actions. Physical robotics is a separate topic.
• Writing a generic list of AI risks with no link to the case. Fix: Quote or use a case detail in every point, such as the data source, the decision affected or the staff involved.
• Confusing bias with poor data quality. Fix: Data quality is about accuracy and completeness. Bias is about unfair or unrepresentative patterns that treat groups differently. Name both separately and give a different control for each.
• Writing a generic essay on AI with no scenario facts. Fix: Quote or paraphrase a scenario fact in every paragraph and say what it means for this company.
• Saying the AI is responsible for a wrong decision. Fix: State that the board and management stay accountable and must keep human oversight and clear lines of responsibility.

Exam tips

• Examiners reward application. Tie every definition to the data, customers or decision in the scenario.
• Give a balanced answer. If you describe benefits, add risks and controls, and finish with a recommendation.
• Use the labelled or unlabelled data test to choose between supervised and unsupervised learning, and show your reasoning in one sentence.
• Link black box models to governance, ethics and accountability, and refer to the ACCA ethical principles where the scenario involves decisions affecting people.
• Keep definitions short. Spend the saved time on analysis and evaluation, where the professional skills marks are earned.
• Always anchor AI uses to scenario facts. Examiners reward application over lists.
• Use clear verbs for the application: detect, predict, automate. It keeps answers structured and quick to write.
• Balance benefits with a brief condition or limit. Many SBL requirements want an evaluation, not a sales pitch.