lesson
60 min

When Not to Use AI

Apply hard blockers, compare against a non-AI baseline, and practice saying no with evidence.

Objective

By the end of this lesson, you can apply hard blockers that stop an AI project, compare an AI design to a simpler baseline, and explain a "no" in terms a director will accept.

Saying no is an architecture skill

The most valuable thing an architect can say is sometimes "not this, not yet." A system that should not exist cannot be fixed by a better prompt.

AI is a good fit when the task is fuzzy, the volume is high, small errors are tolerable or caught, and you can measure quality. It is a poor fit when a rule already works, when errors cause serious harm with nobody checking, or when you have no way to know if it is right.

Hard blockers

A hard blocker is a condition that stops the project no matter how good the demo looks. Treat these as vetoes, not as points on a scale:

  1. High harm, no human review. Denying benefits, flagging fraud, or giving medical or legal direction with no person checking.
  2. No way to measure quality. If you cannot build a test set, you cannot tell good from bad.
  3. A simple rule already works. Routing by zip code does not need a model.
  4. No consent for the data. If people did not agree to this use, the data stays out (Module 6).
  5. The people affected were never asked. Design with them, not for them.

Then, only if no blocker fires, weigh softer factors like volume and error tolerance.

Run it. Then add a fourth case of your own, such as drafting thank-you notes to donors or translating a flyer. Decide its values honestly. Which blocker, if any, fires? Note that blockers are checked before any score is computed.
Python 3
Ctrl+Enter to run | Tab to indent | Python runs in your browser via Pyodide

The first run downloads Python into your browser (several MB). On phone data, you can read the logic here instead:

if case["harm_if_wrong"] == "high" and not case["human_review"]:
    blockers.append("high harm with no human review")
# ...more blockers...
if blockers:
    return "NO-GO: " + "; ".join(blockers)
score = case["volume"] + case["error_tolerance"] - case["sensitivity"]

Always build a baseline

A baseline is the simplest non-AI design that could work: a rule, a search box, a better form, a phone menu, a printed sheet. Build it, or at least estimate it, first.

The baseline does two jobs. It gives you a fair comparison. And it often solves the problem outright, with less to maintain and less that can go wrong.

QuestionBaseline (FAQ page + phone menu)AI assistant
Can a volunteer audit every answer?Yes, every answer is written downOnly by sampling
Works without internet or a smartphone?Phone menu doesDepends on design
What happens when it is wrong?One page to fixHard to predict
Handles unusual questions?NoSometimes, sometimes wrongly

Neither column wins by default. The point is to make the trade-off visible.

A national example

Practice: decide go or no-go

Go / no-go gate

0 of 8 done

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

Ask. Answer. Build together.