STEP 4 OF 7

Classify your Safety Integrity Level

For each hazard, answer how bad, how often and how correctable. AASIF derives its AI-SIL, and the highest one becomes your flow's Safety Integrity Level.

1Answer the fact questions for each hazard
2Confirm or change each proposal; every proposal shows its rule and reason
3Confirm each hazard with 'Confirm my rating · next hazard'

AI-SIL 3 Your flow: AI-SIL 3 · Safety Integrity Level 3Go to measures

14 of 14 hazards confirmed

Hazard

H01

Wrong decision or action from incorrect output (hallucination, misreasoning)

Could the agent act on a wrong fact or a wrong conclusion?

Expert view (S/E/C tiles)

FQ-S-ORG If this goes wrong, what is the worst credible harm to your organization?

Think of the credible worst case, not the typical case. Foreseeable harm counts even if it has not happened yet.

e.g. Worst case for the organization: in step 'Pays supplier invoice', the agent pays the wrong amount or the wrong party (typical €2k, maximum €250k). That could mean a direct financial loss.
e.g. If this goes unnoticed while step 'Pays supplier invoice' runs many times per day, the harm adds up before anyone stops it.

Reference cases (3)

P07 X0PA recruitment agents with human checkpoints (Pass D) · rated 2 · AI-SIL 2Public case · AASIF desk rating, not a finding by the organization
P13 GovTech phase-1 coding assistants (Pass D) · rated 2 · AI-SIL 2Public case · AASIF desk rating, not a finding by the organization
P02 Dayos IT ticketing agent (Pass N: controls stripped) · rated 2 · AI-SIL 1Public case · AASIF desk rating, not a finding by the organization

FQ-S-PER What is the worst credible harm to people outside the organization or to employees?

Include customers, applicants, citizens, patients, suppliers' staff and employees.

e.g. If in step 'Pays supplier invoice' the agent pays the wrong amount or the wrong party, the harm to suppliers could be money they are owed not arriving, or a wrong charge.
e.g. Ask whether any suppliers could lose money, access, privacy or a right through step 'Pays supplier invoice'.

Reference cases (3)

N09 McDonald's AI drive-thru ordering pilot (2021–2024) · rated 1 · AI-SIL 0Public case · AASIF desk rating, not a finding by the organization
P05 PwC Singapore report-drafting agents (Pass N) · rated 1 · AI-SIL 0Public case · AASIF desk rating, not a finding by the organization
N10 Anthropic Project Vend shop agent (2025) · rated 1 · AI-SIL 0Public case · AASIF desk rating, not a finding by the organization

at least S2RH12Step 'Pays supplier invoice' moves money and its maximum size is material (A08): at least S2 for the organization or for the affected persons (R-S-FLOOR).
suggested S2RH14The agent decides about, or informs decisions about, suppliers: consider at least S2 for them (73% of such hazards in the 60 cases were rated S2 or higher).

FQ-E How often does the situation occur in which this could go wrong?

Count each decision in a batch separately. For a detector, count how often a real threat passes by.

e.g. Step 'Pays supplier invoice' runs many times per day. Count how often the agent works from a wrong fact or draws a wrong conclusion there, not how often the step runs.
e.g. If the agent works from a wrong fact or draws a wrong conclusion in only one run out of a hundred, rate exposure from that rate, not from how often step 'Pays supplier invoice' runs.

Reference cases (3)

N09 McDonald's AI drive-thru ordering pilot (2021–2024) · rated 4 · AI-SIL 0Public case · AASIF desk rating, not a finding by the organization
P07 X0PA recruitment agents with human checkpoints (Pass D) · rated 4 · AI-SIL 2Public case · AASIF desk rating, not a finding by the organization
P13 GovTech phase-1 coding assistants (Pass D) · rated 4 · AI-SIL 2Public case · AASIF desk rating, not a finding by the organization

proposal E4RH21not confirmed yetStep 'Matches invoice to purchase order' runs many times per day; for 'Wrong output' the hazard situation usually occurs about as often as that (E4).

FQ-C If it happens, can the harm be corrected before it really hurts?

This is about the harm, not the action. An email can be 'undone' by a correction, but a wrong promise in it may already bind you.

e.g. If the agent works from a wrong fact or draws a wrong conclusion in step 'Pays supplier invoice' and harm follows, the money can only be recalled with effort, and recalls often fail.
e.g. For step 'Pays supplier invoice', a person checks only exceptions. The error may only show in the next reconciliation.

Reference cases (3)

P07 X0PA recruitment agents with human checkpoints (Pass D) · rated 2 · AI-SIL 2Public case · AASIF desk rating, not a finding by the organization
P13 GovTech phase-1 coding assistants (Pass D) · rated 2 · AI-SIL 2Public case · AASIF desk rating, not a finding by the organization
N10 Anthropic Project Vend shop agent (2025) · rated 2 · AI-SIL 0Public case · AASIF desk rating, not a finding by the organization

proposal C2RH05not confirmed yetStarting point: in the 60 reference cases the harm of 'Wrong output' could typically be corrected only with effort or delay (C2). Confirm it for your case.

AI-SIL for this hazardAI-SIL 2

S2 + E4 + C2 − 6 = AI-SIL 2

S: proposal S2 (RH14) → your value S1 · down

This hazard: AI-SIL 2 → 2 · Flow: 3 → 3 · 0 measures change grade

Back

Classification is deterministic — AI only adapts wording and suggests, you confirm.