lesson
60 min

Start With the Problem, Not the Model

Turn a vague request like "we need a chatbot" into a problem statement you can design, test, and defend.

Objective

By the end of this lesson, you can turn a tool request into a written problem statement. It names the users, the outcome, how you will measure it, and what it costs when the system is wrong.

What an AI architect does

An AI architect decides how the parts of an AI system fit together. That means the model, the data, the prompts, the tools, the people who check the work, and the paths for when things fail.

In AI Engineering you built models. In this course you design systems around them. The hardest part is not picking a model. It is deciding what the system is for, and whether it should exist.

We don't just teach tools. We build judgment.

Requests are not problems

Most AI projects start with a solution: "We need a chatbot." "Can we use AI for intake?" A request like that tells you what someone saw in a demo. It does not tell you what is broken.

Your first job is to ask questions until the real problem shows up:

  • Who is struggling, and with what, today?
  • What happens now? Walk the current process step by step.
  • What does better look like? Name a measurable outcome.
  • What does wrong cost? Who is hurt, how badly, and can it be undone?
  • Who checks the work? Name the person or role.

A problem statement template

Write it down. A problem you cannot write in five lines is not ready to design.

PartExample: food bank phone line
UsersCallers seeking food help, many calling from prepaid phones, many preferring Spanish
Current painCallers wait on hold; most questions are about hours, locations, and ID rules
OutcomeMore callers get correct hours and location info without waiting
MeasureShare of common questions answered correctly, checked weekly against a test set; hold time
Cost of wrongA family goes to a closed site, or is told they need ID when they do not
Who checksVolunteer coordinator reviews flagged answers daily

Notice what the statement does not say. It never says "chatbot." The answer might be a better phone menu, a text line with fixed replies, or a printed flyer. AI is one option among several.

Measure before you design

An outcome you cannot measure is a wish. For each outcome, name:

  1. The metric. For example, correct answers on a test set of 100 real (de-identified) questions.
  2. The baseline. How the current process scores today.
  3. The bar. The score a new design must beat to be worth the change.
  4. The guardrail metric. Something that must not get worse, such as accuracy for Spanish speakers.

The guardrail metric matters most. Systems can improve the average while getting worse for one group. You will only see it if you measure each group.

Quiz: Framing check

  1. 01Which of these is a problem statement, not a tool request?

    Choose one.

  2. 02Why add a guardrail metric, such as accuracy for Spanish speakers?

    Choose one.

A national example

Practice: frame a real request

Frame a request

Private. Saved only in this browser. No one else can see it.

Your reflection stays in your browser. It is not sent anywhere.

Go deeper: AI, Minus the Hype: A Practical Guide to Solving Real Business Problems

Ask. Answer. Build together.