TEST VERSION · no company data needed — an abstract description of your use case is enough
AASIF

SAFETY INTEGRITY LEVELS FOR AGENTIC AI

Not every AI agent needs the same safeguards.Find out which ones yours needs.

AASIF rates your use case on a Safety Integrity Level from 0 to 4 and recommends measures scaled to that level, each with its reasoning.

See a finished Safety Concept

Free test version · about 15 minutes + 2 minutes of your view · no company data needed

AASIF SAFETY CONCEPT

Supplier invoice payment agent

PDF · DOCX

AI-SIL 3 · Safety Integrity Level 3

1
Use case and scope3 agent actions
2
Risk assessment14 hazards · 1 critical even if rare
3
Measures50 selected · 46 required
4
Roadmap and decisions4 open points
5
Compliance legendEU AI Act · ISO/IEC 42001

Built on ISO 26262 · ISO 21448 SOTIF

Contributes to EU AI Act · ISO/IEC 42001 · NIST AI RMF · Singapore MGF

THE QUESTION EVERY AI OWNER FACES

Your agent can approve, pay and send on its own. How much safeguarding is enough?

Too little

One wrong action reaches customers, money or people's rights before anyone notices.

Too much

Every agent gets the same heavy controls, and useful projects stall in review.

Not defensible

Nobody can explain to the board or the auditor why this level of control is the right one.

WHAT YOU GET

Walk away with a Safety Concept you can defend.

One document per use case, in a common language your board, CISO, legal team, engineers and auditors can review together.

  • ✓Your AI-SIL and the reasoning behind it
  • ✓Likely and critical side effects, including the rare ones
  • ✓Measures scaled to the level, to prevent or to detect and recover
  • ✓A roadmap with the scope decisions your team still has to take
  • ✓What each measure contributes to the EU AI Act, ISO/IEC 42001, NIST AI RMF and Singapore MGF
Open the full example

AASIF SAFETY CONCEPT · 3

Measures

Authenticated and validated inter-agent messagingPrevent++
Output and egress control (no auto-rendered external links or images; egress allow-list)Prevent++
Kill switch: halt and lock autonomous actionDetect & recover++
Decision logging (inputs, actions, model version, rationale)Detect & recover++
Why? Agents must trust only verified messages.

Contributes to EU AI Act · ISO/IEC 42001 · NIST AI RMF · Singapore MGF

FOR YOUR ROLE

What AASIF does for your role

Persona · illustrative Tester · real and consented
1 / 8

HEAD OF AI

Deciding which agent pilots go into production

“I can say yes to the right agents and show why the others need more work first.”

Persona · illustrative

WHAT YOU GET

  • Lets the company say yes to more agents safely, instead of a blanket no or a blanket yes.
  • A short path for low-risk cases keeps the effort proportionate too, not just the safeguards.
  • Faster go-live approval brings value from AI use cases forward.
  • Lets a non-expert run a structured safety analysis without hiring a safety engineer.
  • The free tool gives an initial risk view that would otherwise take the first days of consulting.
  • Bring your own classification: in-house risk tiers map onto AI-SIL instead of being replaced.

HOW IT WORKS

Three steps, about 15 minutes.

01

Describe your use case

What the agent does, which decisions it takes and where its limits should be.

02

Classify the risk

Guided questions surface what could go wrong. AASIF rates the use case from AI-SIL 0 to 4.

03

Get your Safety Concept

Measures scaled to the level, a roadmap and a document to download as PDF or DOCX.

The classification follows fixed, versioned rules. AI only adapts wording and suggests examples; you confirm every value.

PROVEN PRACTICE, TRANSFERRED

Built on the method that made autonomous cars safe.

An emergency brake decides in 50 ms. A payment agent can commit €250k in seconds. Both need safety designed in, not checked afterwards.

01

Proven

Decades of automotive functional safety: ISO 26262 and ISO 21448 SOTIF.

02

Transferred

The same integrity-level logic, adapted to AI agents; effort scales with risk.

03

Explained

Deterministic classification; every rating and measure comes with its reason.

04

Connected

Each measure shows what it contributes to today's AI governance references.

Show me the method

WORKED EXAMPLE · FICTIONAL COMPANY

Walk through a finished analysis — no sign-in.

A supplier invoice payment agent, analyzed end to end through the real screens, down to the downloadable Safety Concept.

Open the worked example
1Use case and scope3 actions
2Grey areas11 decided · 4 open
3Risk assessment14 hazards · AI-SIL 3
4Measures50 selected
5Roadmap and Safety Conceptready

FAQ

Questions before you start

Find out what your AI use case needs.

Open the worked example