A Repeatable Way to Choose the Tool and Autonomy
Work through the explanation, apply it to the example, and complete the quick check before continuing.
Use the TASKS process
You do not need to guess whether AI is suitable. Use five steps:
- T - Task: Name and split the task.
- A - Accuracy and authority: Set the evidence, acceptable error, and decision owner.
- S - Sensitivity: Classify every proposed input.
- K - Kind of tool and control: Choose the tool, review, and level of autonomy.
- S - Stop conditions: Decide when the process must pause or escalate.
Step 1: Task
Write the task as an action and result.
Weak description: Work on the report.
Better description: Extract the five approved measures from the report and draft a 150-word summary for internal review.
If the task combines search, calculation, generation, and approval, split those parts.
Step 2: Accuracy and authority
State:
- Which evidence the result must use.
- What errors are unacceptable.
- What may happen if it is wrong.
- Who can review and approve it.
If nobody can check a high-impact output, AI is not a suitable shortcut.
Step 3: Sensitivity
Classify each input as public, internal, confidential, personal, or prohibited. Use the organisation’s stricter rule where it differs.
Do not hide a data problem by calling the task a test.
Step 4: Kind of tool and control
Choose the simplest method that meets the requirement.
One response
Suitable for a bounded draft or transformation where the source is supplied, the result is easy to review, and the system takes no outside action.
Example: rewrite an approved public paragraph in simpler English.
Controlled multi-step workflow
Suitable when the task needs separate search, extraction, calculation, drafting, and review stages. Each stage has a known tool and check.
Example: find an official rate, calculate an amount in a spreadsheet, then draft an explanation.
Tool-using or agentic process
May be suitable for repeated steps where the system can use tools or take actions. It requires stronger permissions, monitoring, limits, approval points, logs, and stop controls.
Do not choose an agent merely because the task has several paragraphs. Choose it only when delegated multi-step action adds real value and can be controlled.
Human-led or no automation
Choose this when authority, sensitivity, consequence, unclear evidence, or lack of oversight makes automation unsuitable.
Step 5: Stop conditions
Write them before work begins.
Example:
Stop if the official rate cannot be found, the documents disagree, personal information appears, or the calculated amount is outside the normal range. Send the case to Finance.
Worked recommendation
Task: prepare a public summary of an approved, non-confidential research note.
Task: transformation and extraction. Accuracy and authority: every claim must match the note; communications owner approves publication. Sensitivity: source is approved public information. Kind: one AI draft is reasonable, followed by claim-by-claim human review. Stop: pause if the draft adds evidence, numbers, or conclusions not in the source.
Recommendation:
Use AI for a first draft because the source is public and the output is reversible before publication. Require a human comparison with the approved note. Do not allow automatic publication.
Remember
- Start with the task, not the tool.
- Choose the simplest tool that meets the requirement.
- More autonomy needs more control.
- A recommendation is incomplete without evidence, data classification, authority, review, and stop conditions.
Next, you will apply the process to several situations.