Skip to content
AI.info

AI literacy basics

What AI Systems Actually Do

Learn the major task families of AI and translate real problems into classifications, forecasts, rankings, retrievals, generations, optimizations, and actions.

By the end you can

Example

One support inbox can produce eight different AI tasks

“Use AI for support” is too vague to design against. It is also too vague to evaluate. The same inbox can produce several distinct outputs, each with its own labels, metrics, and risks.

  • Classification: assign each message to billing, technical support, cancellation, or another queue.
  • Priority scoring: estimate urgency or likely customer harm.
  • Ranking: order messages so agents see the most important cases first.
  • Retrieval: find the most relevant policy or previously approved solution.
  • Extraction: pull order numbers, dates, products, and requested actions into fields.
  • Forecasting: estimate expected resolution time or future ticket volume.
  • Generation: draft a summary or proposed reply for review.

A product brief should name the output before it names the model.

Case

The same inbox, measured in a real support centre

Which output you pick, and who receives it, is measurable. In “Generative AI at Work”, published in the Quarterly Journal of Economics in May 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered rollout of a generative AI conversational assistant to 5,172 customer-support agents. Issues resolved per hour rose 15% on average, but the effect was not evenly spread: less experienced and lower-skilled agents improved in both speed and quality, while the most experienced and highest-skilled saw small gains in speed and small declines in quality. One inbox, one output family — a suggested reply — and the result still depended on whose hands it landed in.

Visual

The output families behind most AI products

Many applications look unique at the interface level but rely on a small set of recurring computational tasks.

FigureHierarchy · 5 levels
  • Describe or detect

    Classification, extraction, recognition, segmentation, and anomaly detection describe what is present.

    • Estimate the future

      Regression, forecasting, risk scoring, and survival estimates predict quantities or events.

      • Order or retrieve

        Ranking, recommendation, search, and matching decide what should appear first or be considered relevant.

        • Create or transform

          Generation, translation, summarization, synthesis, and editing produce a new candidate artifact.

          • Choose or control

            Optimization, planning, and policy systems select actions under objectives and constraints.

Comparison

The same input can support very different output logics

A photo of a damaged package illustrates how task choice changes the question being answered.

The choice is also a bill for annotation. Microsoft COCO was presented by Tsung-Yi Lin and colleagues in 2014. It contains “2.5 million labeled instances in 328k images”. The paper reports that outlining them was “an extremely time consuming task requiring over 22 worker hours per 1,000 segmentations”. A classifier for those same photos needs one label per image. The detector needs a boundary drawn around every object in it.

FigureComparison · 4 columns

Classification

Choose a category such as “minor,” “major,” or “not damaged.”

  • Output: label or probability
  • Needs a category definition
  • Errors depend on class boundaries
  • Useful for routing

Detection

Locate the damaged region with a box or mask.

  • Output: coordinates or pixels
  • Needs spatial annotation
  • Errors include missed or false regions
  • Useful for inspection evidence

Generation

Produce a written description of the visible damage.

  • Output: open-ended text
  • Needs factual grounding
  • Errors can be fluent but unsupported
  • Useful for drafting a claim note

Decision policy

Recommend refund, replacement, or manual inspection.

  • Output: proposed action
  • Needs cost and policy constraints
  • Errors affect people and money
  • Useful only with a defined workflow

Retrieval, recommendation, and generation solve different selection problems

Retrieval searches a collection for items relevant to a query. Recommendation predicts which items may be valuable to a user or context. Generation composes a new artifact from patterns learned during training and the context supplied now.

A modern assistant may use all three: retrieve documents, rank passages, then generate a response. Evaluate only the final prose and three questions go unanswered — whether the evidence was found, whether it was ranked well, and whether it was used.

That combination has a founding paper and a name. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” was presented in 2020. Patrick Lewis and eleven co-authors wrote it, at Facebook AI Research, University College London and New York University. Its retrieval half was a fixed December 2018 Wikipedia dump. In it “each Wikipedia article is split into disjoint 100-word chunks, to make a total of 21M documents”. Its generation half wrote the answer. The reported score on the Natural Questions test set was 44.5 exact match. That is a single number. It says nothing about which of the two halves was responsible when an answer came out wrong.

Break compound products into task stages before assigning blame or credit.

Steps

Translate a user need into a measurable output

Good task formulation narrows the gap between an inspiring idea and an evaluable system.

FigureProcess · 5 steps
  1. 1

    State the user decision

    Describe the choice or work the user is trying to complete.

  2. 2

    Choose the output form

    Specify a label, number, ranked list, retrieved item, generated draft, or action proposal.

  3. 3

    Define the target event

    Clarify what counts as correct and when the truth becomes observable.

  4. 4

    Name the tolerance for error

    Separate reversible mistakes from errors with financial, safety, or rights consequences.

  5. 5

    Choose the handoff

    Decide whether the output informs, recommends, blocks, or acts automatically.

Case

A screening model meets the clinics that have to use it

What that gap costs is on record. In a field study presented at CHI 2020, Emma Beede and colleagues followed a deep-learning screening system for diabetic retinopathy through eleven clinics in Thailand. In the laboratory it read retinal photographs at more than 90% accuracy; in the clinics, where the lighting was poor, it rejected more than a fifth of the images as falling below its quality threshold, and those patients were told to come back another day. The output form was correct and the model was accurate on the images it accepted. What had not been specified was the target event for an unreadable photograph, the tolerance for that error, and the handoff — which turned out to be a second trip for the patient.

Key idea

A technically correct output can serve the wrong objective

A recommender can predict clicks accurately and still make the product less useful, more addictive, and less diverse. A support classifier can route every message correctly and still teach a team to avoid difficult customers. Accuracy is not the same as benefit.

The output is what the model produces. The objective is the behavior the organization rewards, and the outcome is what happens to people and systems afterward.

The people who built one of the largest recommenders in the world wrote this down. Paul Covington, Jay Adams and Emre Sargin presented “Deep Neural Networks for YouTube Recommendations” in 2016. They warn that “recommendation often involves solving a surrogate problem and transferring the result to a particular context”. They also name the surrogate that fails. “Ranking by click-through rate often promotes deceptive videos that the user does not complete (‘clickbait’) whereas watch time better captures engagement.” Their ranking objective is therefore expected watch time per impression, not click probability. The model did not become more accurate. The question it was scored on changed.

Never infer a good objective from an accurate prediction.

Analogy

A workshop for information

A workshop holds measuring tools, sorting bins, shelves, drafting tables, and machines that choose cutting paths. Each station performs a different operation on the same raw material.

AI task families resemble those stations: measure, sort, find, compose, choose. A cut plank stays where you put it. Information outputs change behavior and institutions, which is where the workshop stops helping.

Naming the operation clarifies what evidence and failure tests are needed.

Task-first thinking prevents model-first waste

Teams often begin with a model they want to use and then go looking for a problem; task-first design reverses the order, defining the decision, the output, the truth signal, the error costs, and the handoff before choosing a method.

The discipline also reveals when one product contains several tasks: a “copilot” may classify intent, retrieve context, generate a draft, check policy, and request approval, all in one interaction.

The unit of analysis is often a chain of tasks, not one magical model call.

Key takeaways