The Reality

Everyone asks "can we use AI for this?" Not every problem needs AI. Sometimes a database query is better. Sometimes a rule beats a model. Sometimes a human is faster.

Good Use Cases

✓ Text Classification

Classify emails, receipts, tickets into categories.

Why: Fuzzy input. Typos, variations. Subjective boundaries.

Example: "Spent £45 at Tesco on groceries" → Groceries.

✓ Summarization

Extract key info from long documents, emails, transcripts.

Why: Subjective. What's "key" depends on context.

Example: Extract date, merchant, amount from receipt.

✓ Information Extraction

Pull structured data from unstructured sources.

Why: Handles formatting variation. Learns from examples.

Example: Extract {date, merchant, amount, category} from receipt photo.

✓ Semantic Search

Find similar documents, past conversations, relevant examples.

Why: Understands meaning, not just keywords.

Example: "I lost my password" matches "forgot credentials".

✓ Content Generation (With Guardrails)

Generate personalized copy, emails, explanations.

Why: Scale. Can't hire 100 writers. AI generates 10K variations.

Risk: Needs human review. Hallucinations happen.

Bad Use Cases

✗ Exact Fact Lookup

Don't use AI if you need the exact answer.

Example: "What's Q2 revenue?"

Why: Model guesses based on patterns. Query the database instead.

✗ Deterministic Logic

Don't use AI for "if A then B" decisions.

Example: "Apply 10% discount if order > £100".

Why: AI is probabilistic. Use a rule engine. Faster, cheaper, reliable.

✗ High-Stakes Instant Decisions

Don't use AI for instant, high-stakes calls.

Example: "Approve this loan application?"

Why: Needs human review. Defeats the speed advantage.

✗ Audio/Video Understanding (Today)

Vision models improve, but hallucinate on uncommon objects.

Why: Real-world data is messier than training data. Reliability is low.

✗ Tasks With No Training Data

AI learns from patterns. No examples = can't learn.

Why: Will hallucinate. Use rules or humans instead.

Decision Framework

Compare AI against alternatives before committing:

AI
  • Handles fuzzy input
  • Scales to millions
  • Learns from examples
  • Cost: £0.01–1/request
  • Latency: 2–10s
Rule Engine
  • Millisecond latency
  • No hallucinations
  • Cheap (once built)
  • Breaks on edge cases
  • Hard to scale
Database Query
  • Exact answers
  • < 100ms latency
  • Cheap
  • Structured data only
  • No flexibility
Human Review
  • Perfect accuracy
  • Understands context
  • Expensive (£10–100)
  • Slow (hours)
  • Doesn't scale

Quick Checklist

Reality check: 78% of AI pilots fail because they solved the wrong problem, not because the model was bad. The problem didn't need AI in the first place.

The Hybrid Approach

Best practice: Use AI where it excels. Use rules/databases everywhere else.

Example: Expense Categorization

Result: AI handles 70%. Rules handle 20%. Humans handle 10%. Cheaper and more reliable than AI-only.

The best AI projects aren't the ones with fancy models. They're the ones that solved the right problem with the right tool. Sometimes that tool isn't AI.