AI in everyday products

What you will be able to do

See an AI product as five parts, not as “a model”. Each part is work that somebody does. When you know the parts, you can see where the jobs are, and where your own skills fit.

Before you startWhat to know and to have readyAbout 25 minutes

What you need to know

• Nothing. No earlier course is needed. You do not write code in this course.

What you need

• A browser and a text editor. Any editor that saves plain text works: Notepad on Windows, TextEdit on macOS (in plain text mode), or VS Code.
• In Set up your course notes, you make the folder and the notes file that you use for the whole course.

Set up your course notes

You write short answers in every lesson. At the end of the course, you use them to write your final assignment. Keep them in one file.

1. Make a folder with the name ai-career in your Documents folder.
2. Open your text editor, and make a new file.
3. On the first line, type # My AI career notes.
4. Save the file in the ai-career folder, with the name notes.md.
5. Close the editor. Open notes.md again.

Expected result: the file opens and shows the line # My AI career notes. If the file is not there, look in the editor’s “Recent files” list to see where it was saved. Then save it again in the ai-career folder.

If you use TextEdit on macOS, select Format › Make Plain Text before you save. Otherwise, TextEdit saves a rich-text file, and the .md name can change to .rtf.

The .md ending means Markdown: plain text, where a line that starts with # is a heading. You do not need to know more Markdown than that.

Words for this course

AI (artificial intelligence) means computer systems that do tasks that usually need human judgement, such as understanding text, recognising images or making predictions.

ML (machine learning) is the most common way to build AI today. Instead of writing every rule by hand, people show a computer many examples. The computer finds the patterns itself. The result is a model: a file of learned patterns that turns an input into an output.

For example, a spam filter is a model. It learned from millions of emails that people marked as spam or not spam. For each new email, it gives a score: how likely this email is spam.

Inference means using a trained model on a new input to get an output. Each time the spam filter checks one new email, that is one inference.

The example: a document assistant at Larkfield

In this course, you follow one company. Larkfield is an online shop that sells lamps, furniture and kitchen goods. Larkfield is fictional, but its problem is common.

Larkfield’s support team answers about 1,200 customer messages each week. Many answers depend on company policies: returns, delivery, warranty and payment. There are 40 policy documents, and some of them change several times a year.

Today, an agent searches the documents by hand. That takes about four minutes for each question. In a quality check, the support lead found that about 1 answer in 12 quoted an old rule.

Larkfield wants a document assistant. The agent types a question. The assistant finds the correct policy section, and writes a short answer with a link to that section. The agent checks the answer and then replies to the customer. The assistant never writes to the customer itself.

Five parts of an AI product

The five parts of Larkfield's document assistant. Operation feeds new problems back into the next round of work.

Text description of the diagram

1. User need: support agents need the correct policy rule, fast.
2. Data: the 40 policy documents, current versions only.
3. Model: finds the relevant sections, and writes a short answer.
4. Application: a panel in the help desk tool, with a link to the rule.
5. Operation: policy updates, feedback from agents, and checks every week.

An arrow goes from Operation back to User need: operation finds new problems and new policies, and these start the next round of work.

Part	The question it answers	At Larkfield
1. User need	Who has a problem, and what does a better result look like?	Agents waste time and sometimes quote old rules. Better: a correct answer with its source, in under a minute.
2. Data	What information does the AI use, and can we trust it?	The 40 policy documents. Only the current version of each one.
3. Model	What turns the input into the output?	A language model that reads the found sections and writes a short answer.
4. Application	How does a person use it, and what happens around it?	A panel inside the help desk tool. It shows the answer, the link to the policy section and the date of the policy.
5. Operation	How do we keep it working after the first day?	New policies are added automatically. Agents mark bad answers. Someone reviews the marked answers every week.

Follow one question through the parts

An agent gets this message from a customer: “I bought a lamp 40 days ago. It works, but I do not like the colour. Can I return it?”

Step	What happens	Result
1. The agent asks	The agent types: “Change-of-mind return after 40 days?”	A question in the assistant’s panel
2. Find the data	The assistant searches the current policy documents.	It finds Returns policy, March 2026, section 2.1: “Change-of-mind returns: within 30 days of delivery. Faulty items: within 90 days.”
3. Inference	The model reads the question and section 2.1, and writes an answer.	“No. Change-of-mind returns are accepted for 30 days after delivery. (Returns policy, section 2.1)”
4. Application	The panel shows the answer, the link and the policy date. The agent opens the link and checks it.	The agent replies to the customer, and offers help if the lamp becomes faulty.
5. Operation	The question, the answer and the agent’s rating are saved for the weekly check.	If the answer was wrong, the team sees it on the next review.

Look at how small step 3 is. The model is necessary, but most of the work is in the other four parts. Someone had to understand the agents’ problem. Someone had to collect the correct documents and keep them current. Someone had to put the answer inside the tool that the agents already use, and someone has to check it every week.

Each of those “someones” is a job. You meet them in the next lessons.

Guided practice: one everyday product

Here is a worked example with a product that you probably use: the arrival time in a maps app.

Part	In a maps app
1. User need	A driver wants to know when they will arrive, and which road is fastest.
2. Data	Road maps, speed limits, and the current speed of many phones on each road.
3. Model	Predicts the travel time on each road for the next hour.
4. Application	The app shows the route, the arrival time, and a new route if a road is blocked.
5. Operation	The team compares predicted and real arrival times, and updates the maps when roads change.

Now do one yourself:

1. In notes.md, add a heading: ## Five parts.
2. Choose one AI feature that you use. Examples: a spam filter, a phone photo search (“show photos of dogs”), or film suggestions in a streaming app.
3. Write one line for each of the five parts.
4. For part 5, write one thing that can go wrong after the first day.

Show an example answer for a spam filter

1. User need: people want important emails in the inbox, and no spam.
2. Data: emails that people marked as spam or not spam.
3. Model: gives each new email a spam score.
4. Application: emails over a threshold go to the spam folder. People can move an email back.
5. Operation: spammers change their messages, so the model must learn from new examples. Can go wrong: a new type of spam gets through, or an important email goes to the spam folder.

Check: each of your five lines must say something different. If parts 3 and 4 say the same thing, ask: “What does the person see and do?” That is the application.

Common mistake

Thinking that the model is the product

Symptom: a project plan says only “train a model” or “use a language model”. Nobody can say who uses the result, what happens when it is wrong, or who updates it.

Cause: the model is the most famous part of AI, so it gets all the attention. The other four parts look like “normal work” and are forgotten.

Fix:

1. Write the five parts for the project, one line each.
2. For each part, write the name of the person or team who does it.
3. If a part has no name, the plan has a gap. Fill it before you start.

Check your understanding

Larkfield changes its delivery policy, but nobody adds the new document. Which part failed, and what do the agents see?

Show the answer

Part 5, operation (keeping the data current). The model still works correctly: it writes good answers from the documents that it has. But the documents are old, so the agents see confident answers with an old rule. This is why the panel shows the date of the policy: it helps the agent see the problem.

Check your understanding

Is one question to the document assistant one inference, or many? Explain.

Show the answer

At least one. The model writes an answer for each question, and each run of the model on a new input is an inference. Many products use more than one model for one question. For example, one model may find the sections and another model writes the answer. Then one question is two inferences.

Recap

• AI systems do tasks that need human judgement. ML builds them from examples, and the result is a model. Using a model on a new input is inference.
• An AI product has five parts: user need, data, model, application and operation.
• The model is one part. Most of the work, and most of the jobs, are in the other four.
• You have a notes.md file for the whole course.

Next, you meet the main job families in AI, and what each one produces.

Next: The main job families