AI & ML Quick Path

Solve any case: a decision guide

Use this page whenever you meet a new problem at work or in an interview. Follow the nine steps, use the table to pick an approach, then study the worked examples.

Nine steps for any AI or ML case

  1. Step 1: Write the decision

    What will the system decide or produce, for whom, and how often? What does a wrong answer cost? Example: "Each morning, rank customers by risk of cancelling, for the retention team; a missed churner costs about a month of revenue, a wasted call costs ten minutes."

  2. Step 2: Check whether AI is needed

    If clear rules work, write rules. If the pattern is fuzzy, changes often, or involves language, images or sound, AI may help.

  3. Step 3: Find the data

    What data exists, who owns it, is it labelled, is it allowed to be used, and does it contain personal data?

  4. Step 4: Pick the approach

    Use the table below. Start with the simplest approach that could work.

  5. Step 5: Define success before building

    A baseline, a metric tied to the business cost, and an evaluation set that is never used for training or tuning.

  6. Step 6: Build the smallest working version

    A notebook or script for ML; a prompt plus validation for LLM apps. Measure it against the baseline.

  7. Step 7: Plan the risks

    Privacy, fairness, security (prompt injection, excessive agency), cost, and what happens when the model is wrong or down.

  8. Step 8: Ship it safely

    Serve it as an API or batch job, with model version, monitoring, a fallback and a rollback plan.

  9. Step 9: Measure and improve

    Watch drift, quality and cost; turn every confirmed failure into a new test.

Which approach fits the problem?

If the problem is…Start withWhyLearn it in
The rules are clear and stable (tax rate, opening hours)Plain code, no AICheapest, testable and explainable. Use AI only where rules break down.What AI, machine learning, deep learning and generative AI are
Predict a number from a table (price, delivery time, demand)Regression (gradient boosting first)Tables of business data are where classic ML is strongest.Regression: predicting a number
Pick a category from a table or short text (churn, fraud, spam, intent)ClassificationNeeds labelled history; choose the metric from the cost of each mistake.Classification: picking a category
No labels; find groups or unusual casesClustering or anomaly detectionSegments customers or flags fraud and faults for review.Unsupervised learning: groups, compression and anomalies
Predict future values over timeTime-series forecastingSplit by time; beat a seasonal baseline.Time-series forecasting
Suggest items to usersRecommender systemCollaborative filtering plus content features for new items.Recommender systems
Images, audio or videoDeep learning with a pretrained modelTransfer learning needs far less data than training from scratch.Computer vision: classification, detection and segmentation, Speech and audio: recognition and text-to-speech
Write, summarise, translate or extract from textAn LLM with a good prompt and structured outputStart with a hosted or local model; validate every output.Calling an LLM from your application, Prompt engineering that works in production
Answer questions from your own documentsRAGKeeps answers current and cited; no retraining.RAG: answering from your own documents, Advanced RAG: chunking, hybrid search, re-ranking and evaluation
Take actions in other systems (refunds, bookings, tickets)Agent with tools and human approvalLeast privilege and approval in code.Tools and agents with human approval
A specific style, format or narrow skill the prompt cannot reachFine-tuning (for example LoRA)Only after prompting and RAG are measured and fall short.Fine-tuning: SFT, LoRA, QLoRA and preference tuning
Data must never leave the buildingA local modelSmall quantised models run on laptops and servers you control.Running language models on your own machine
Learn by trial and error from a simulator (control, scheduling)Reinforcement learningRarely the first choice in business; needs a safe simulator.Reinforcement learning

Rule of thumb. Most real systems combine several rows. Solve each sub-problem with the simplest approach that works, and connect them with ordinary software (see Production architecture for AI systems).

Worked examples

An insurance company wants to speed up claims

Two separate problems. (1) Read the claim form and photos: an LLM with structured output extracts fields from text, and a vision model sorts photos by damage type. (2) Decide which claims need a human: a classification model on claim history flags likely fraud for review. Every claim decision stays with a person; the AI prepares and prioritises.

Lessons: Multimodal AI apps: images, documents and voice, Calling an LLM from your application, Computer vision: classification, detection and segmentation, Classification: picking a category, Responsible AI: fairness, privacy, transparency and oversight

A hospital IT team wants doctors to search internal guidelines

RAG over approved guideline documents, with permission filtering, citations and "I don’t know" when not covered. Because data is sensitive, a local model or a provider with suitable data agreements. An evaluation set written with clinicians before launch. No diagnosis by the system.

Lessons: RAG: answering from your own documents, Advanced RAG: chunking, hybrid search, re-ranking and evaluation, Running language models on your own machine, Evaluating AI applications, Security for AI applications

For real, documented examples from companies, see the case studies. Practise by applying the nine steps to each one before reading its suggested workflow.