A practical rule

Start with the task, not the model

When a new tool becomes popular, it can be tempting to look for reasons to use it. Start in the other direction: describe the result needed, list the constraints, and choose the simplest method that can produce a reliable result.

A spreadsheet can total numbers predictably. A script can rename hundreds of files in the same way each time. Search can find the official page where a policy changed. A human may be needed to weigh a sensitive decision. AI becomes useful when the task involves language, synthesis, drafting, exploring options, or translating an unclear problem into a plan—and when the result can be checked the result.

  • Define the deliverable and what a correct answer looks like.
  • Choose the lowest-complexity method that meets the need.
  • Estimate the cost of a wrong or incomplete answer.
  • Decide how the result will be verified before starting.

A simple comparison

Use the right tool for the shape of the work

Use ordinary code for deterministic tasks, databases for structured records, search for current source material, and AI for tasks where interpretation or language is the main difficulty. These categories overlap, but the distinction helps prevent unnecessary complexity.

  • Calculator or spreadsheet: repeated arithmetic, budgets, tables, and explicit formulas.
  • Script or application: repeatable transformations, scheduled jobs, validation, and operational workflows.
  • Search and primary sources: current prices, regulations, security advisories, release notes, and exact product specifications.
  • AI assistant: outline options, explain unfamiliar concepts, draft a first version, critique a plan, or turn notes into a structured checklist.
  • Qualified person or accountable owner: high-impact decisions that require context, professional judgment, consent, or formal responsibility.

AI can help navigate the other tools, but it should not replace a deterministic check when a deterministic check is available.

Workflow

A five-step test before trusting an AI answer

  1. Give it a bounded task. Explain the goal, context, constraints, audience, and required output.
  2. Ask it to expose assumptions. Ask what it is uncertain about and which claims need external verification.
  3. Verify important facts. Open the linked source, check publication dates, and compare the claim with the original document.
  4. Test the output. Run code in a safe environment, recompute calculations, or compare a recommendation against a checklist.
  5. Keep human ownership. Edit the result, record the decision, and remain responsible for what is published or executed.

For current events and product changes, answers remembered from older conversations are not current sources. Check official release notes or primary documentation when the information matters.

Proportionality

AI has costs beyond the subscription

AI systems use physical infrastructure: data centres, chips, networking, electricity, and cooling. The exact footprint depends on the model, hardware, workload, and energy system, so a broad claim such as “AI is always greener” or “AI is always wasteful” is too simplistic.

A practical rule for tool selection is: if a small script or clear checklist solves the task reliably, adding an AI step may add latency, uncertainty, privacy exposure, and operational cost without a useful benefit. If AI helps explain a complicated topic or compare many sources, the extra complexity may be justified.

For wider context, consult the International Energy Agency’s Energy and AI analysis. Treat it as a source to read, not as a single number that applies to every AI request.

Guardrails

Match review to the cost of failure

The higher the consequences, the more evidence and human review the workflow needs. A low-stakes brainstorm may only need an edit. A script that changes a production server requires a test environment, code review, backups, and a rollback plan. Financial, medical, legal, employment, or safety decisions need appropriately qualified human judgment.

  • Do not paste passwords, private keys, recovery codes, customer records, or confidential material into a service unless its approved data-handling arrangements explicitly permit it.
  • Ask for citations, then verify that the sources exist and support the specific claim.
  • Do not execute commands only because the assistant suggests them; understand their effect first.
  • Keep a record of decisions that materially affect a project or another person.
  • Use AI to challenge your thinking, not merely to confirm it.

Reusable prompt

Ask for a tool choice, not an AI sales pitch

Desired outcome: [specific outcome].
Context and constraints: [time, budget, tools, access].
Compare the simplest reliable options: ordinary software, a script, web research, an AI assistant, or a human expert.
For each option, explain cost, failure modes, privacy risks, and how the result can be verified.
Recommend the lowest-complexity option that meets the need. State assumptions and what you cannot verify.

The useful answer might be “do not use AI for this.” That is a good result when it reduces unnecessary work and risk.

Sources and further reading

Documents worth checking