Understanding AI · Field guide

The AI Model Types Field Guide

“AI” covers many kinds of model, and each kind is built for a different kind of question. In our experience, choosing the wrong kind of model can cost more than choosing the wrong vendor, and many business problems are better served by an older, cheaper model than by a chatbot. Match three problems to a model first.

Guess first

Match the problem to the model

A bank wants to flag card transactions that are likely to be fraud. It has five years of transactions, each already labelled fraud or not fraud. Which kind of model fits?

A retailer wants to split its two million customers into segments for marketing. Nobody has labelled any customer with a segment yet. Which kind of model fits?

A contact centre wants to know how many agents to schedule for each hour of next month. Which kind of model fits?

The guide

What do you need the model to do?

The guide is showing all 10 model types.

Predict an answer

These models learn from past examples where the answer is known, then predict the answer for new cases.

Classification

Use it whenYou want to sort each case into one of a few buckets: fraud or not, likely to churn or not, which team a ticket belongs to.

How it failsThe model learns from past labels, so it repeats past decisions, including their biases. When one bucket is rare, a high accuracy figure can hide a model that misses most of the cases you care about.

Ask the vendorWhat are the precision and the recall on the cases we care about most, measured on our own data?

Practise it: Classification Challenge →

Regression

Use it whenYou want to predict a number for each case: a delivery time, a claim cost, the price a house will sell for.

How it failsThe model predicts well inside the range of its training data and poorly outside that range. A model that is right on average can still be far off on an individual case.

Ask the vendorHow wrong is a typical prediction, in dollars or days, and how wrong is a bad one?

Forecasting

Use it whenYou want to predict a value over future time: next month’s demand, call volumes by hour, cash flow by week.

How it failsA forecast assumes the future will look like the past. A price change, a new competitor or a pandemic can make the history the model learned from stop applying.

Ask the vendorHow did the model perform through the last disruption, and what range does it give rather than a single number?

Practise it: The Forecast Cone →

Find patterns

These models look for structure in data that nobody has labelled, so a person has to judge whether what they find is useful.

Clustering

Use it whenYou want to find natural groups in data that has no labels, such as customer segments or types of support request.

How it failsClustering will produce groups even in data that has none. The number of groups is often chosen by a person, and a different choice gives a different story.

Ask the vendorWould the same groups appear if we ran the model again on next quarter’s data?

Anomaly detection

Use it whenYou want to flag the unusual: a strange transaction, a machine running hot, a login from an odd place.

How it failsUnusual is not the same as bad. When real problems are rare, most alerts can turn out to be false alarms, and analysts learn to ignore them.

Ask the vendorOut of every 100 alerts, how many turn out to be real problems?

Practise it: The Anomaly Hunt →

Recommendation

Use it whenYou want to suggest the next product, article, course or action for each person, based on what similar people chose.

How it failsRecommenders favour what is already popular, and each recommendation shapes the next round of data, which can narrow choice over time. A new customer or a new product has no history to work from.

Ask the vendorHow does the system handle a new customer or a new product with no history?

Generate content

These models produce new text, images or audio. Their output is fluent by design, so checking it is part of the job.

Large language models

Use it whenYou want to work with language: draft, summarize, translate, extract fields from documents, or answer questions from a set of sources.

How it failsA language model writes fluent text whether the content is right or wrong. On its own, a language model is also unreliable at arithmetic and at forecasting, even when the answer sounds confident.

Ask the vendorHow will we check the output, and what happens when the output is wrong?

Practise it: Why AI Hallucinates →

Image and video generation

Use it whenYou want to create images, video or audio from a description: concept art, product mock-ups, illustrations.

How it failsDetails that must be exact, such as lettering, counts and logos, are the details most likely to slip. Rights to the output and the content of the training data are still being settled in court.

Ask the vendorWhat rights do we have to the output, and what was the model trained on?

Read the physical world

These models turn images and sound into data that other systems can use.

Computer vision

Use it whenYou want a system to read images or video: find defects on a production line, read scanned forms, count stock on a shelf.

How it failsPerformance drops when the lighting, the camera or the product differs from the conditions in the training images.

Ask the vendorHow was the accuracy measured, and were the test images taken in conditions like ours?

Speech recognition

Use it whenYou want audio turned into text: call transcripts, meeting notes, voice commands.

How it failsAccuracy drops with accents, background noise, people talking over each other and specialist vocabulary. The errors that matter most tend to be in names and numbers.

Ask the vendorWhat is the error rate on our own recordings, including our product names?

Use it this week

Name the model type in your next AI proposal

Take one AI proposal on your desk and write down which of these model types the proposal depends on. If nobody can tell you, ask the vendor question from that card before the next meeting.

Knowing what kind of tool you are buying is one part of thinking well in a digital workplace. The Digital Intelligence course builds the rest of that skill one day at a time.

See the 30-day course Decode the AI jargon