How Reliable Must the Result Be?
Work through the explanation, apply it to the example, and complete the quick check before continuing.
Not every mistake has the same cost
Compare these two tasks:
- Suggest five names for an internal team event.
- State the correct dose of a medicine for a patient.
A weak event name is easy to reject. A wrong medical instruction may cause serious harm. The second task needs qualified professional judgement and trusted clinical systems. A general AI response is not an acceptable authority.
Before choosing a tool, ask four questions.
1. What evidence is required?
Possible evidence levels include:
- No factual evidence: creative options that a person will choose from.
- Supplied evidence: a summary must match the given document.
- Current evidence: the result depends on information that may have changed.
- Authoritative evidence: an official rule, approved record, or specialist source is required.
2. What error is acceptable?
Acceptable error is the amount and type of error the task can safely tolerate.
For brainstorming, several weak options may be acceptable because a person will select one. For payroll, the correct amount is required. “Mostly right” is not enough.
3. What happens if the result is wrong?
Consider effects on:
- A person’s health, safety, rights, money, education, or employment.
- The organisation’s legal duties, security, reputation, or operations.
- Other people who may rely on the result.
The greater the possible harm, the stronger the evidence, control, and human authority must be.
4. Can the action be reversed?
A draft that has not been sent is easy to change. A payment, public message, deleted record, rejected application, or disclosed secret may be hard or impossible to reverse.
Reversibility does not remove the need for care. It helps you decide how much control is needed before action.
A simple reliability table
| Situation | Evidence need | Error consequence | Reversibility | Sensible approach |
|---|---|---|---|---|
| Private brainstorming | Low | Low | Easy | AI may suggest options; person chooses |
| Summary of approved notes | Supplied source | Medium | Easy before sharing | AI draft plus line-by-line check |
| Current visa requirement | Official current source | High | Decision may be costly | Check official authority; seek qualified help where needed |
| Invoice calculation | Exact values and rules | Medium or high | Sometimes difficult | Approved calculator/formula plus review |
| Hiring decision | Full authorised evidence | High | Hard for affected person | Human-led approved process; AI must not be final authority |
Stop conditions
A stop condition tells you when not to continue automatically.
Examples:
- A required source is missing.
- Two trusted sources disagree.
- The result is outside the reviewer’s expertise.
- Sensitive information appears unexpectedly.
- The proposed action cannot be reversed.
- The system gives a low-confidence or unsupported answer.
At a stop condition, pause and send the work to an authorised person or approved process.
Quick check
Which task can usually tolerate the most variation?
A. Calculating a final salary payment. B. Confirming an allergy from a medical record. C. Suggesting themes for an informal celebration. D. Approving access to a secure system.
Check the answer
Answer: C. The ideas can be reviewed and rejected before anyone acts. The other tasks require exact evidence and stronger authority.
Remember
- Match reliability to evidence, acceptable error, consequence, and reversibility.
- High-impact work needs stronger sources and human authority.
- Decide stop conditions before the tool begins work.
Next, we will match common needs to search, calculators, specialists, fixed software, and AI.