Domain knowledge in practice

What you will be able to do

See concrete cases where AI helps a field, and the knowledge from that field that the AI needs. At the end, you write the same kind of card for your own field. It becomes evidence in your career brief.

Before you startWhat to know and to have readyAbout 40 minutes

What you need to know

• The parts of domain knowledge, from Your experience is a head start: What domain knowledge is.
• The three ways to use your experience: contribute, lead and build, from Your experience is a head start: Three ways to use your experience.
• You do not need to read every field in detail. Read the field that is closest to yours carefully, and at least one other field.

What you need

• Your domain and your “I know …, so I can …” lines, under ## My head start in notes.md. You wrote them in Your experience is a head start: Guided practice: translate your experience. Your domain card uses the same domain.
• This lesson adds the heading ## Domain card.

How to read the domain cards

Each field below has the same four parts:

1. Where AI helps: real kinds of AI use in the field.
2. What the AI team needs to know: the domain knowledge that a technical person without experience in the field does not have.
3. A mistake without it: a short fictional case, built from common real problems.
4. Roles that a domain expert can lead.

The use cases are examples, not a complete list. Every organisation is different, and each idea still needs the problem framing that you learn in Module 3.

Finance: banking and accounting

Where AI helps

• Fraud detection: score each card payment for the risk of fraud, and stop the risky ones for a check.
• Invoice processing: read supplier invoices, and fill in the accounting system.
• Credit decisions: help decide whether to give a loan, and how much.
• Month-end close: find unusual entries before the accounts are closed.

What the AI team needs to know

• Fraud labels arrive late. A customer may report a fraud payment weeks after it happened. So the most recent data looks “clean”, but it is not. A model trained on it learns too little about fraud.
• A false alarm has a cost too. A blocked card on a customer’s holiday is a real harm. Some customers leave the bank after it. The team must balance missed fraud against blocked good customers.
• Explanations can be a legal duty. In many countries, a lender must give the main reasons when it refuses credit. A model that cannot give reasons may not be allowed for that decision.
• Some data must not be used. Laws in many countries ban credit decisions based on protected characteristics such as ethnicity or religion. Other data, such as a postcode, can be a hidden substitute for them.
• Invoice details matter. The invoice date is not the due date. The same supplier can have two names. Tax rules depend on the country and the type of goods.

A mistake without it

A team builds a model to find duplicate invoices. It checks for the same supplier, the same amount and the same invoice number. An accountant sees the problem at once. Real duplicates often have a slightly different invoice number, such as INV-1043 and INV1043. The supplier sent the invoice twice, by two channels. The model misses most real duplicates, and every test that the engineers wrote passes.

Roles that a domain expert can lead: AI product manager for finance tools, evaluation lead for invoice or fraud models, model risk and validation analyst (many banks have this team), AI adoption lead in a finance department.

Health care

Where AI helps

• Clinical notes: turn a conversation between doctor and patient into a draft note, which the doctor checks.
• Medical images: mark areas on an X-ray or a scan that a specialist should look at first.
• Appointments: predict which patients may miss an appointment, so the clinic can call them.
• Referral triage: sort incoming referrals by how urgent they are.

What the AI team needs to know

• Abbreviations have more than one meaning. “MS” can mean multiple sclerosis or mitral stenosis, a heart valve disease. The context decides.
• “No” matters. “No chest pain” and “chest pain” contain the same words. A system that only looks for words gets the meaning backwards.
• Treatment hides risk. As in the pneumonia example from the last lesson, patients who look safe in the data may look safe only because they got good care.
• Too many alerts are ignored. Clinical staff get many alerts already. A new alert that is wrong often trains people to ignore all alerts, including the correct ones.
• The rules are strict. Health data has strict privacy laws in most countries, for example HIPAA in the United States and the GDPR in the European Union. In many countries, software that helps diagnose patients can count as a medical device, and needs approval.
• The workflow decides the value. A good prediction is useless if it reaches the wrong person, or arrives after the decision was made.

A mistake without it

A team builds a model that predicts which patients will miss their appointment. It is accurate. The clinic then gives the predicted “no-show” time slots to other patients. A nurse sees the result. The patients who were predicted to miss often had the most difficulties, such as no car, night shifts or poor health. Their appointments got harder to keep. The model was accurate. The way the clinic used it made care worse for the people who needed it most. The nurse proposes a different action: call those patients and offer help with transport.

Roles that a domain expert can lead: clinical subject-matter expert or clinical lead for an AI product, evaluation of clinical outputs, clinical informatics (the field that joins clinical work and information systems), AI safety and governance in a hospital.

Legal

Where AI helps

• Contract review: find important clauses, such as termination, liability and payment, in many contracts.
• Legal research: find relevant laws and earlier court decisions.
• Document review: sort many documents in a court case into relevant and not relevant.
• Client intake: collect the facts of a new case before a lawyer reads it.

What the AI team needs to know

• Every source must be checked. In 2023, a judge in New York fined two lawyers who filed a document that cited court cases that did not exist. A chatbot had made the cases up. The lawyers had not checked them. Lawyers are responsible for every citation, so a legal AI must show its sources, and the workflow must include the check.
• The place decides the law. The same question has different answers in different countries, and sometimes in different states of one country. This is called jurisdiction.
• Defined terms change meaning. In a contract, “Affiliate” or “Business Day” means exactly what the contract defines, which may be different from the normal meaning.
• A missing clause matters. A contract without a limit on liability is a risk. A system that only finds clauses that exist does not see it.
• Confidentiality and privilege. Communication between lawyer and client is protected, and some data cannot be sent to outside services.

A mistake without it

A team builds a tool that summarises leases for a property company. The summaries read well. A paralegal tests ten of them. One summary reports “notice period: 3 months”, but an addendum signed two years later changed it to 6 months. The tool read the main contract but not the later changes. For lawyers, the most recent signed version is the one that counts.

Roles that a domain expert can lead: legal subject-matter expert for a legal-technology product, legal engineer or legal operations lead, product manager for legal AI, AI governance and policy work, where lawyers often lead.

Retail

Where AI helps

• Demand forecasting: predict how much of each product each shop will sell, so the right amount is ordered.
• Product descriptions: write descriptions for thousands of products from their data.
• Customer service: answer common questions about orders, delivery and returns, as at Larkfield.
• Pricing and promotions: suggest which products to discount, and when.

What the AI team needs to know

• Zero sales does not mean zero demand. If a shelf was empty, nobody could buy the product. A model that learns “no sales on Tuesday” learns the wrong lesson. Store staff know which days the shelves were empty.
• Promotions and seasons change everything. A sales jump may come from a promotion, a holiday or the weather. The model must know about them, or it predicts the jump again next week.
• One product steals from another. A discount on one brand of coffee lowers the sales of the other brands. Total sales may not change at all.
• Some product claims are illegal. A product description must not say that a cream “cures” a skin condition, or that food is “organic” when it is not. A language model writes such claims easily if nobody stops it.
• Returns tell a story. A high return rate for one size can mean that the size chart is wrong, not that customers are difficult.

A mistake without it

A team forecasts demand for umbrellas from last year’s sales. The model predicts low sales for October. A store manager sees why: last October the umbrellas arrived three weeks late because of a supplier problem, so there were few sales. Demand was high. The shelf was empty. The model was about to repeat the problem.

Roles that a domain expert can lead: product manager for forecasting or merchandising tools, AI adoption lead in stores or e-commerce, evaluation of product content, customer service AI lead.

Four more fields, in short

Field	Where AI helps	One thing the AI team needs to know	Roles to lead
Human resources and recruiting	Sort job applications, answer employees’ policy questions, plan training.	Past hiring data contains past bias. In 2018, Reuters reported that Amazon stopped an experimental recruiting tool that had learned to score down CVs with the word “women’s”. Some places now have laws for AI in hiring. For example, New York City requires a bias audit for some automated hiring tools.	HR AI product lead, fair-hiring and AI governance, HR adoption lead
Manufacturing and maintenance	Predict machine failures, check product quality with cameras.	Technicians know which sensor readings are normal: a machine may be loud every cold morning. A fault is often recorded as “other” because the list of fault codes is long.	Maintenance AI lead, quality engineering with AI, operations product manager
Education	Tutoring, feedback on writing, help for teachers with planning.	A wrong answer shows a specific misunderstanding. A good tutor finds it. A tool that only gives the correct answer can stop the learning.	Learning design for AI tutors, education product manager, evaluation of feedback quality
Logistics and delivery	Predict delivery times, plan routes, sort parcels.	Drivers know the local problems: a road that floods, a building with no entrance on the street, a depot that closes early on holidays.	Operations AI lead, routing product manager, evaluation with drivers

Guided practice: write your domain card

1. In notes.md, add a heading: ## Domain card.
2. Copy the template below.
3. Choose one task in your field where AI could help. Write it as one use case, not a whole department.
4. Fill in the four parts. Use the cards above as examples.
5. For “What the AI team needs to know”, write at least three items. Use the parts of domain knowledge from the last lesson as a checklist: words, exceptions, real workflow, data meaning, expensive mistakes, trust and laws.

Domain:
Use case (one task):

Where AI helps:

What the AI team needs to know:
1.
2.
3.

A mistake without it (2–4 sentences):

Roles that I could contribute to, lead or build:

Check: read your card and ask, “Could a good engineer who never worked in my field write this?” If the answer is yes, your items are too general. Replace them with details that only experience gives you.

Independent variation: choose a field from this lesson that is not yours. Add one more item to its “What the AI team needs to know” list. If you do not know the field, ask a person who works in it. This is the same conversation that AI teams have with subject-matter experts every week.

Common mistake

A domain card full of general words

Symptom: your card says “AI can improve efficiency in my field”, “the team needs to understand the business”, and “mistakes can be costly”.

Cause: these sentences are true for every field, so they show no domain knowledge. It often happens when people write about their field from the outside, or try to sound like a report.

Fix:

1. Replace each general sentence with a specific example that happened in your work.
2. Use numbers, names of documents, and real steps: “Month-end close takes 5 days, and 2 of them are spent on matching invoices.”
3. Write the mistake as a short story with a cause, like the cards in this lesson.

Check your understanding

In the retail card, why is 'zero sales does not mean zero demand' a domain problem and not a programming problem?

Show the answer

The data has a zero, and the zero is recorded correctly. Nothing in the code is wrong. The problem is the meaning of the zero: the shelf was empty, so customers could not buy. Only somebody who knows how shops work, and which days the shelves were empty, can see this and give the team a way to correct it.

Check your understanding

Look at the 'mistake without it' stories. What do almost all of them have in common?

Show the answer

In almost every story, the technical work was correct: the model was accurate, the tests passed or the summaries read well. The failure came from a fact about the field that the team did not know, such as late fraud labels, an addendum or a late delivery. In most stories, a domain expert found it quickly by looking at real cases.

Recap

• Every field has AI use cases, and every use case needs domain knowledge that technical people do not have.
• The most common failure is not bad code. It is a correct model built on a wrong understanding of the field.
• Domain experts can lead AI products, evaluation, adoption and governance in their field.
• Your domain card shows your knowledge with specific examples. You use it in your career brief.

Next, you test the levels in your skills inventory with small tasks and real evidence.

Next: Evidence over self-ratings