Choose a useful AI problem

Start with a decision, not a model

A useful AI project begins with an activity someone already needs to complete. Imagine a community repair workshop receiving questions about opening hours, required tools, and whether an item can be repaired. The aim is not to add a chatbot because chatbots are fashionable. The aim might be to help a volunteer find the correct published instruction in under one minute. This definition gives us a person, an input, a desired output, and a measurable benefit. It also leaves room for a simpler solution: a well-organized frequently asked questions page. Write down what happens without AI before proposing what should happen with it. If a search box or a short checklist solves the problem reliably, that comparison is valuable evidence. Model complexity is a cost to justify, not a sign that the project is more serious.

Three ways software can help

A rule-based program follows instructions written directly by a person. A workshop rule might reject any appointment outside opening hours. A predictive model learns patterns from examples and produces an estimate or category. It could estimate how long an intake conversation will take, provided appropriate historical examples exist. A generative model produces content, such as a draft explanation of a repair policy. These approaches can coexist in one product. A generated explanation may be followed by a strict rule that prevents booking outside opening hours. The interface can look equally conversational in all three cases, so appearance does not tell you how the answer was produced. Ask which parts are learned, which parts are retrieved, and which parts are enforced by ordinary software. This distinction helps locate mistakes and choose a proportionate fix.

Capability does not remove uncertainty

A language model can produce a persuasive sentence without having verified its claim. Its answer may combine useful patterns with an unsupported detail, such as inventing a service the workshop never offered. Friendly tone and confident wording are not measurements of correctness. Even a system that performs well on familiar questions may fail on unusual wording, outdated information, or a request outside its intended scope. Think in terms of conditions rather than universal labels such as smart or accurate. Does it work with this source collection, these languages, this audience, and this review process? A wrong opening time causes inconvenience; an invented instruction for handling dangerous equipment can cause greater harm. The acceptable amount of automation should depend on the consequence of error as well as the average quality of answers.

Write a bounded success statement

A strong first scope says that the assistant may answer published administrative questions and point to the relevant source. It must ask for clarification when a question is ambiguous and refer repair-safety judgments to a volunteer. Success might mean that eight of ten representative questions lead a reviewer to the correct source faster than the existing page does, with no invented policy details. That is a proposed learning target, not a claim about an untested product. Include the cost of review in the comparison: an answer that takes twenty seconds to generate and two minutes to check may save nothing. Finally, choose a stopping condition. If users mostly ask questions requiring personal judgment, improve the intake process rather than quietly expanding the assistant's authority. A small useful system with clear boundaries is a sounder starting point than an impressive demonstration without an owner.

Separate assistance from authority

This is a conceptual scenario. It splits the request into verifiable information, missing context, and a consequential judgment. It does not assume that a generated answer can authorize acceptance.

Request: Can I bring my damaged kettle tomorrow?
Allowed help: Find tomorrow's opening hours and the published intake policy.
Missing facts: Which date is tomorrow? Does the policy cover this item?
Human decision: Whether the damaged appliance is safe to accept.
Baseline: A volunteer searches the same approved policy page.

Ready to try it yourself?

ChatGPT sign-in takes you to OpenAI and back to CodeTrail. It keeps your learning account separate from other learners.