Decide what AI should—and should not—do

Begin with work, not a tool

AI literacy begins before a prompt is written. It begins with being able to describe the work: what starts it, what information enters, what a good result looks like, who is affected, and who is responsible when the result is wrong.

This path uses one real workflow as the learning material. Choose something repeatable and reviewable, such as organizing intake notes, drafting a routine follow-up, classifying open tasks, or preparing a first-pass summary. Do not use live private data while learning. A redacted or invented example is enough.

The goal is not to automate the task. The goal is to make a defensible decision about where assistance belongs, what must remain human, and what evidence would justify a small pilot. Completing this path does not provide certification, professional authority, or a guarantee that AI is appropriate for your workflow.

Lesson 1 — Map the work as it happens now

Write the current process from trigger to finish. Name the input, each meaningful step, the output, the person who owns it, and the place where the work commonly stalls. Record what actually happens rather than the process you wish existed.

Then mark the evidence used to judge a correct result. A follow-up draft might need the requester’s actual question, an approved description of the offer, a clear next action, and a truthful statement about timing. If you cannot name the evidence, neither a person nor a model has a stable review standard.

• What event starts the task?
• Which inputs are required, and where do they come from?
• What finished output is produced?
• Who owns the final decision and any promise made?
• What error would cause real harm, confusion, cost, or lost trust?

Lesson 2 — Separate assistance from authority

Repeatable labor and accountable judgment are not the same thing. AI may help organize, compare, summarize, draft, format, or surface missing fields. Those are candidate assistance roles, not permission to decide or publish.

Keep authority with a person when the work makes a consequential promise, approves access, interprets unclear evidence, judges fairness, handles an exception, makes a moral or legal call, or decides that a claim is true. The more difficult it would be to repair a mistake, the stronger the human boundary should be.

• Candidate assistance: organize notes, extract explicit fields, prepare a draft, compare against a checklist, or format a reviewed decision.
• Human ownership: approve claims, choose an outcome, authorize access, handle sensitive context, accept an exception, or send the final communication.
• Keep manual: rare work with many exceptions, work without reliable evidence, or work where reviewing the output costs as much as doing it correctly.

Lesson 3 — Design the review before the draft

A human review step is meaningful only when the reviewer knows what to check. Write a short acceptance checklist before trying a tool. Use statements that can be verified: every date matches the source, no unsupported promise appears, required fields are present, and the named owner is correct.

Test with several safe examples, including an ordinary case, an incomplete case, and an exception. Record incorrect additions, missing facts, ambiguous wording, and the time needed to verify the result. A fluent answer is not evidence of a reliable process.

• What sources must the reviewer compare against?
• Which fields or claims must never be inferred?
• What condition requires the work to stop or return to manual handling?
• Who signs off before the output affects another person?

Lesson 4 — Set privacy, access, and failure boundaries

Use the minimum information needed for the task. Remove personal, confidential, financial, health, access, or client information from experiments unless there is a legitimate basis, an approved tool, and a reviewed handling process. Do not treat a convenient input box as permission to disclose data.

Keep early pilots reversible. Preserve the source material, retain a manual route, and state who can disable the assistance. Define what happens when the tool is unavailable, produces inconsistent output, or cannot explain an important result. A workflow that fails silently is not a dependable system.

• Minimum data: only the fields required for the stated task.
• Minimum access: no broad account or system permission for a narrow experiment.
• Visible failure: errors and uncertain outputs reach a person instead of disappearing.
• Manual fallback: the work can continue without pretending the tool is reliable.

Lesson 5 — Decide on one bounded pilot

Choose one of three honest decisions: assist one narrow step, keep the workflow manual, or stop until the evidence and controls improve. “Not yet” is a useful result when the task is unclear, the data is sensitive, the review burden is too high, or the failure cost is disproportionate.

If a pilot is justified, define its start and end, the safe examples it will use, the acceptance checks, the named reviewer, and the condition that stops the test. Track corrections and review time instead of assuming that generated speed equals saved labor.

At the end, compare the pilot with the original process. Keep it only if the work remains truthful, reviewable, maintainable, and meaningfully easier. The person responsible for the workflow owns that decision; AI does not approve its own role.

Practice lab — test the boundary, not the best-looking answer

Use four safe cases before deciding whether the assisted step is dependable enough to continue. Begin with an ordinary complete example, then remove one required fact, introduce an exception, and test the manual fallback while the tool is treated as unavailable. Use invented or properly redacted material. The exercise is about the review system, not the realism of private data.

For each case, keep the source evidence beside the output and mark every acceptance check. Record incorrect additions, missing facts, unclear language, and the time needed to review and correct the result. A case passes only when a person can trace the accepted output to evidence and the workflow handles uncertainty without inventing an answer.

The missing-information and exception cases matter because routine examples can hide a weak boundary. A model may produce fluent language even when the required evidence is absent. The reviewer must be allowed to stop, return the work to manual handling, or decline to produce an output. Completing a form is not more important than preserving a truthful decision.

• Ordinary case: all required evidence is present and the output can be checked directly.
• Missing-information case: at least one necessary fact is absent and must remain visibly unresolved.
• Exception case: the request falls outside the ordinary rule and reaches the named human owner.
• Fallback case: the work remains possible when the AI tool or integration is unavailable.

Review evidence before deciding to continue

Compare the pilot with the original manual step. Count review time, corrections, exceptions, and maintenance work as part of the cost. Generated speed alone is not saved labor when a person must reconstruct the evidence, repair invented details, or monitor a fragile connection.

Choose continued use only when the narrow assistance remains easier after review, the source remains visible, errors become visible, privacy and access are proportionate, and a named person can stop the workflow. Otherwise revise the test, keep the work manual, or stop until the evidence and controls improve.

Use the responsible AI pilot log to preserve this evidence and the final human decision. The log is a learning artifact, not a certification or approval from MethodCo. A real organization remains responsible for the professional, contractual, legal, privacy, security, employment, and other requirements that apply to its work.

Complete the review

Use the downloadable worksheet to keep the workflow map, human boundary, acceptance checks, privacy limits, fallback, and pilot decision together. The document is deliberately portable and can be completed without an account or an AI tool.

This learning path is educational material, not legal, privacy, security, employment, or religious advice. It does not certify a vendor, model, automation, or workflow. Consequential uses require the facts of the situation and appropriate human or professional review.