Keep a work task manual when it crosses a confidentiality or organisational boundary, a wrong output could cause serious harm, nobody can independently verify the result, or the decision itself must be made by an accountable person. Use AI for a reversible draft only when the inputs are allowed and a person can check every material part. Between those cases, use AI only with a named human approver.
The useful question is not “Can AI do this?” It is “Can we detect and correct a bad result before it matters?”
First, define the task you are actually judging
Do not score a department or an entire workflow. Score one output and one decision at a time. “Prepare the monthly finance pack” is too broad. It might include:
- collecting files;
- copying approved figures into a template;
- explaining a variance;
- deciding whether to change a forecast; and
- approving the pack for distribution.
Each activity has a different risk profile. A rule-based copying step may be easy to check. A forecast decision may depend on context, ownership and approval rules. Scoring them together lets a low-risk step hide a high-risk one.
Write the task as one sentence:
Using [allowed inputs], produce [output] for [audience], before [deadline or decision].
This exposes the inputs, output and intended use. It also prevents an approval decision from being disguised as “just a draft”.
The SVCRA stop-or-proceed framework
This is a first-party decision aid, not a validated risk standard. It begins with a boundary gate, then scores five frictions: Sensitivity, Verifiability, Consequence, Reversibility and Approval. A boundary can stop the task before a convenience score rationalises it.
Step 1: Apply the boundary gate
Keep the task manual and escalate it if any of these is true:
- Your organisation, client or applicable professional rule does not allow the input or the proposed tool use.
- The task requires a person to make the judgement, but the workflow would delegate that judgement to AI.
- No qualified person can check the output before it is used.
- You cannot explain who owns the final decision.
A tool being available does not overrule a workplace boundary. If the rule is unclear, treat that as a stop and ask the appropriate owner.
Step 2: Score the five frictions
Score each dimension from 0 to 2. A higher score means a stronger reason to keep more of the task human.
| Dimension | 0 — Low friction | 1 — Mixed | 2 — High friction |
|---|
| Sensitivity | Inputs are public or explicitly cleared for the chosen tool. | Inputs are internal but cleared for this use, with defined handling rules. | Inputs are restricted, personal, client-confidential or not cleared. |
| Verifiability | A reviewer can check every material claim directly against approved inputs. | Some parts are checkable; others depend on interpretation or incomplete evidence. | There is no independent method, qualified reviewer or sufficient evidence to check the result. |
| Consequence | An error would be a minor internal inconvenience and would not alter rights, access, pay, health, safety or a material commitment. | An error could cause significant rework or a misleading message but should be caught before a binding action. | An error could alter rights, access, pay, health, safety, a binding commitment or another material outcome. |
| Reversibility | The output is an unsent draft or disposable note that can be replaced. | It can be corrected, but not without cost, confusion or a record of the first version. | It triggers an action that is hard to undo or persists in an important record. |
| Approval | No special sign-off is required before use. | A peer or manager must review it under a normal workflow. | A designated specialist, control owner or formal approver must sign off. |
Verifiability scores the difficulty of checking, not how fluent the output appears. A polished answer that nobody can check is a high-friction output.
Step 3: Map the total to an outcome
Add the five scores for a total between 0 and 10.
- 0–3: AI-assisted draft. AI may produce a working draft. A person checks it against approved inputs before use.
- 4–6: AI-assisted with approval. Name the reviewer before starting. The output does not move forward until that person approves it.
- 7–10: Keep manual. Keep the defining judgement, composition or action human. You may score a smaller, separate subtask rather than automating the original task.
Apply these overrides even if the total is lower:
- A 2 for consequence, reversibility or approval means the task cannot be treated as a simple draft. It needs at least the named approval path.
- A 2 for verifiability combined with a 2 for consequence means keep the task manual: the stakes are high and the result cannot be independently checked.
- A sensitivity score of 2 never overrides the boundary gate. Proceed only if an authorised path for that input and tool already exists.
If two people score an ambiguous item differently, use the higher score until the appropriate owner clarifies it.
Blank SVCRA worksheet
Copy this for each clearly defined task.
| Field | Your assessment |
|---|
| Task sentence | |
| Boundary gate: any stop? | |
| Sensitivity (0–2) | |
| Verifiability (0–2) | |
| Consequence (0–2) | |
| Reversibility (0–2) | |
| Approval (0–2) | |
| Total | |
| Override triggered? | |
| Outcome | Keep manual / AI-assisted draft / AI-assisted with approval |
| Named owner | |
| Evidence used to verify | |
The last two rows matter. “A human in the loop” is not a control until the human is named and knows what evidence to check.
Example 1: Draft an internal project update
Hypothetical task: Using a list of cleared project milestones, produce an unsent weekly update for the project team.
- Sensitivity: 0 — all inputs are explicitly cleared for the tool.
- Verifiability: 0 — the owner can compare each statement with the milestone list.
- Consequence: 1 — a wrong update could mislead colleagues and create rework, but it does not authorise an action.
- Reversibility: 0 — it remains an unsent draft.
- Approval: 0 — the task owner can review it in the normal workflow.
Total: 1. Outcome: AI-assisted draft.
The review is not a quick proofread. The owner checks every date, name, status and next step against the cleared milestone list. The human remains responsible for what is sent.
Example 2: Prepare a supplier schedule-change email
Hypothetical task: Using an approved purchase order and an internal note, prepare an email accepting a new delivery date.
- Sensitivity: 1 — the information is internal but cleared for this use in the hypothetical scenario.
- Verifiability: 0 — the purchase order and note provide direct checks.
- Consequence: 1 — a wrong date or term could cause significant rework.
- Reversibility: 1 — an email can be corrected, but the recipient may act on the first version.
- Approval: 1 — a procurement owner must review the change.
Total: 4. Outcome: AI-assisted with approval.
Name the procurement owner before drafting. That person checks the order number, date, quantity, terms and recipient. The email stays unsent until approved. If the change creates a binding commitment, reconsider the consequence and approval scores rather than reusing this result.
Example 3: Decide an employee’s compensation change
Hypothetical task: Using performance information and pay data, decide whether an employee’s compensation should change.
- Sensitivity: 2 — the task uses personal employment and pay information.
- Verifiability: 2 — the decision depends on judgement and cannot be checked as purely factual claims.
- Consequence: 2 — the decision changes a person’s pay.
- Reversibility: 2 — it can persist in pay and employment records.
- Approval: 2 — a designated people or control owner would need to approve it.
Total: 10. Outcome: keep manual.
The compensation judgement stays human. Do not treat an AI-generated recommendation as a neutral second opinion. A different, narrow task—such as formatting an already approved, non-sensitive template—would need its own boundary check and score. It does not change who owns the compensation decision.
A quick decision table
| If this is the dominant factor… | Default to… | Before proceeding… |
|---|
| The input or tool is not allowed | Keep manual | Stop and ask the policy or data owner |
| Nobody can independently verify the result | Keep manual | Define a verification method or use a qualified human process |
| The result is a reversible, unsent draft from cleared inputs | AI-assisted draft | Compare every material claim with approved inputs |
| An error can be found but would be costly to reverse | AI-assisted with approval | Name the approver and block use until sign-off |
| The output changes rights, access, pay, health or safety | Keep the decision manual; escalate | Follow the applicable expert and organisational approval path |
How to implement the decision in a real team
- Start with a short task inventory. List outputs, not job titles or vague “workflows”. Give each row an input, output, audience and owner.
- Remove unallowed candidates. Apply the boundary gate before discussing tools, prompts or possible convenience.
- Score independently. Have the task owner and proposed reviewer score the five frictions. A disagreement often reveals an unstated approval or verification gap.
- Design the check before the draft. Name the reference file, checklist or qualified judgement used to verify the output. If you cannot name the evidence, raise the verifiability score.
- Separate creation from action. Generating a draft must not automatically send an email, update a record, make a decision or publish a result.
- Record the control. For approval tasks, record the owner, approved version and evidence checked according to your organisation’s existing rules.
- Re-score when the task changes. A new input, audience, tool, approver, delivery method or automatic action can change the outcome.
A completed worksheet is a decision record, not proof that a workflow is safe. Regulated, safety-critical, legal, medical, financial or people-related work should follow the applicable expert and organisational approval path rather than relying on this general guide.
Limitations
- The scorecard is not validated. It is an editorial decision aid, not a prediction of error rates, safety or compliance.
- Numbers can create false precision. The boundary gate and overrides exist because a low total cannot repair a prohibited input or missing accountability.
- Scores depend on context. The same draft may be low-risk internally and high-risk once published, sent or used to trigger an action.
- Verification capacity varies. Score for the actual reviewer and evidence available, not an ideal reviewer.
- Manual does not mean error-free. People can also make mistakes. This framework addresses delegation and review design, not every source of risk.
- It does not replace expert advice or organisational rules. Follow any applicable rule, contract, professional duty or safety process.
FAQ
Which tasks should always stay manual?
There is no universal list based only on a task’s name. Keep a task manual if it fails the boundary gate, or if the stakes are high and the output cannot be independently verified. Score the actual inputs, output and action instead of relying on a generic category.
Is human review enough for any AI task?
No. Review only works when the reviewer is qualified, has reference evidence, has time to check the result and can block it before action. Otherwise, “human review” is only a label.
Should I use AI for a second opinion on a high-stakes decision?
Do not assume an AI opinion is independent, neutral or correct. If a decision affects rights, access, pay, health, safety or a binding commitment, keep accountable judgement with the appropriate person and follow the required approval process.
Can I split a high-risk task into a safer subtask?
Sometimes. Change the task, not its score. For example, separate formatting an approved template from deciding what belongs in it. The narrower task still needs its own boundary check, score and review path.
How often should a task be re-scored?
Re-score when an input, audience, tool, approver, delivery method or automatic action changes. A previous decision does not automatically apply to a new version of the task.
Where can I learn a practical AI workflow from here?
For a structured foundation in asking, choosing, reviewing and delivering AI-assisted work, browse the TakeAICourse course library. It includes the beginner course “AI Essentials: ChatGPT, Gemini, and Claude in 7 Lessons”. Your own policies and approvals still govern the task.