AI Use Case Selection Framework

How do you decide when you should use LLMs? LLMs hallucinate, and the challenge is not hallucination, but knowing when it hallucinates. The flowchart provides a framework for identifying whether a project is suitable for AI.

flowchart TD
    A["Can the task be solved deterministically?"]
    B["Do Not Use AI
Use software, rules engines"]
    C["Is the business value high?"]
    D["Too Much Effort to Verify
Avoid AI"]
    E["Can outputs be verified cheaply and reliably?"]
    F["Fully Automatable Tasks
Automate Aggressively"]
    G["Useful but Needs Review
AI drafts, humans decide"]
    H["Idea Generation & Exploration"]

    A -- Yes --> B
    A -- No --> C
    C -- No --> D
    C -- Yes --> E
    E -- Yes --> F
    E -- No --> G
    H -.-> E

Verification Economics Matrix

Hallucination Risk	Verification Cost	Value	Example Use Cases	Strategic Action
Low	Low	High	Email classification, chatbot FAQs, GitHub Copilot, SQL query writing, code debugging, receipt processing	Automate aggressively. Design for scale with minimal human intervention.
Low	Low	Low	Meeting notes, email summaries, reply suggestions, social media captions, message templates	Automate opportunistically. Deploy selectively for incremental gains.
Low	High	High	Legal contract review, medical record summarization, resume screening, compliance document analysis	Reduce verification cost before scaling. Use structured retrieval and strong guardrails.
Low	High	Low	Auto-generated status reports requiring full review, heavily-edited presentations and marketing content	Avoid. Verification overhead destroys productivity gains.
High	Low	High	Product brainstorming, market research, competitive analysis, strategic planning workshops	Use for ideation and discovery.
High	Low	Low	Creative writing experiments, marketing tagline testing, design mockups, brand concept exploration	Safe for exploration.
High	High	High	Legal brief writing, clinical decision support, compliance reporting, financial analysis, automated trading, database changes	Restrict to assistive roles only.
High	High	Low	Unsupervised customer commitments, auto-posting to social media, heavily-reviewed customer service emails	Avoid entirely.
Near-Zero Required	–	High	Payment processing, price calculations, user authentication, inventory tracking	Do not use LLMs. Prefer deterministic software, rules engines, or RPA.
Near-Zero Required	Low	Moderate	Database backups, scheduled reporting, batch email sends, file synchronization	Typed pipelines outperform probabilistic systems here. AI adds unnecessary risk.