AI Literacy: When not to use AI

When Not to Use AI

What This Is

This guide helps professionals recognize situations where AI adds little value—or creates risk. Knowing when not to use AI prevents wasted time, compliance breaches, and poor outcomes.
Example: A hospital tried using a chatbot to triage emergency room patients, but the model misclassified a heart attack as indigestion, delaying critical care. Human judgment was irreplaceable.

Key Facts & Principles

• High-stakes decisions with irreversible consequences
  AI lacks accountability and nuanced judgment.
  Example: Don’t use AI to approve loans for marginal applicants without human review—regulators may penalize discriminatory patterns.

• Tasks requiring deep domain expertise or tacit knowledge
  AI models generalize from data but can’t replicate years of specialized experience.
  Example: A senior engineer’s gut feeling about a structural flaw in a bridge design is more reliable than an AI’s analysis.

• Low-volume, high-variability work
  AI thrives on patterns; one-off tasks with unique constraints waste time.
  Example: Drafting a custom legal clause for a niche contract is faster done by a lawyer than fine-tuning a prompt.

• Sensitive or regulated data
  Many AI tools process inputs as training data, risking leaks.
  Example: Don’t paste patient records into a public LLM—HIPAA violations can cost millions.

• Real-time, mission-critical systems
  Latency, hallucinations, or API failures can break workflows.
  Example: Using an LLM to monitor a power grid’s stability could miss a critical anomaly during a blackout.

• Creative work where originality is the goal
  AI remixes existing content; it doesn’t invent.
  Example: A marketing team using AI to generate a "unique" brand slogan may accidentally plagiarize competitors.

• Bias-sensitive contexts
  AI amplifies historical biases in training data.
  Example: Using AI to screen job applicants can perpetuate gender or racial imbalances if the model was trained on biased hiring data.

• Explainability requirements
  Some industries (e.g., finance, healthcare) require auditable decisions.
  Example: A bank can’t use a black-box AI model to deny a mortgage without violating "right to explanation" laws.

• Cost-prohibitive use cases
  AI isn’t always cheaper.
  Example: Training a custom model to sort 100 emails/day costs more than hiring an intern.

• Ethical or reputational risks
  Even if legal, some AI uses can damage trust.
  Example: A retailer using AI to dynamically price essential goods during a crisis may face backlash.

Step-by-Step Application

1. Map the task to AI’s strengths
   Ask: Does this require pattern recognition, speed, or scale? If not, AI may not help.
   Example: Automating expense reports (repetitive, rule-based) = good; writing a CEO’s apology email (nuanced, high-stakes) = bad.

2. Assess risk tolerance
   Score the task on:

3. Consequence of error (1–5, where 5 = life-threatening)
4. Regulatory exposure (1–5, where 5 = heavy fines)

5. Reversibility (1–5, where 5 = irreversible)
   If the total > 10, avoid AI.

6. Check data quality and volume

7. Is the data clean, labeled, and representative? If not, AI will fail.

8. Is there enough data? Rule of thumb: <1,000 examples = likely not worth it.

9. Evaluate alternatives
   Compare AI to:

10. Human effort (e.g., a junior analyst)
11. Rules-based automation (e.g., Excel macros)

12. Hybrid approaches (e.g., AI drafts, human edits)

13. Pilot and measure
    Run a small test (e.g., 10% of cases) and track:

14. Accuracy (vs. human baseline)
15. Time saved (or lost)

16. User satisfaction (e.g., surveys)

17. Document the decision
    Write a 1-pager justifying why AI was (or wasn’t) used. Include:

18. Risk assessment
19. Cost/benefit analysis
20. Fallback plan

Common Mistakes

• Mistake: Assuming AI is always faster.
  Correction: AI often adds latency (e.g., API calls, prompt engineering). Test speed vs. manual methods.

• Mistake: Ignoring "last-mile" costs.
  Correction: AI outputs usually need human review, formatting, or integration. Budget for this.

• Mistake: Overestimating AI’s creativity.
  Correction: AI generates combinations of existing ideas, not breakthroughs. Use it for ideation, not innovation.

• Mistake: Using AI for tasks with no clear success metric.
  Correction: Define KPIs (e.g., "reduce email response time by 30%") before deploying AI.

• Mistake: Assuming "AI" = "magic." Correction: AI is a tool, not a strategy. Align it with business goals, not hype.

Practical Tips

• Create an "AI No-Fly List"
  Maintain a shared doc listing tasks where AI is banned (e.g., "performance reviews," "emergency alerts"). Review quarterly.

• Use the "Grandma Test"
  If you wouldn’t trust your grandma to do the task unsupervised, don’t trust AI either.

• Default to "human-in-the-loop"
  Even for low-risk tasks, keep a human reviewer for edge cases (e.g., AI flags 90% of spam, but a person checks the 10% it misses).

• Monitor for "AI creep"
  Regularly audit AI use to ensure it hasn’t expanded into unintended areas (e.g., a chatbot meant for FAQs starts giving medical advice).

Quick Practice Scenario

Scenario: Your team wants to use AI to generate personalized training plans for employees. The data includes performance reviews, skills assessments, and career goals. Question: What’s the biggest risk, and how would you mitigate it?

Answer: Risk: Bias amplification (e.g., favoring certain roles or demographics). Mitigation: Audit the training data for imbalances, use a hybrid model (AI drafts + HR review), and exclude protected attributes (e.g., age, gender) from inputs.

Last-Minute Cram Sheet

1. AI fails at: High-stakes, irreversible, or bias-sensitive tasks.
2. AI thrives at: Repetitive, pattern-based, high-volume work.
3. ⚠️ Trap: "AI is free"—hidden costs include integration, training, and oversight.
4. Rule of thumb: If a task takes <10 minutes manually, AI may not save time.
5. Always ask: What’s the worst that could happen if the AI is wrong?
6. ⚠️ Trap: "AI is objective"—it inherits biases from training data.
7. Hybrid > pure AI: Use AI to augment, not replace, human judgment.
8. Regulated data? Assume AI is off-limits unless explicitly approved.
9. Low data volume? Skip AI—it needs examples to learn.
10. ⚠️ Trap: "AI will scale"—but poor outputs scale too (e.g., 1,000 bad recommendations).