Assurance

Not every product that “does risk assessment with AI” is the same. Fluent answers are easy. An assurable operational risk assessment — grounded in IRAM 2.0, owned by the server, reviewed by experts, and verified in software — is not.

Not all AI assessments are equal

A general chatbot — or even a Custom GPT with instructions and knowledge files — can produce text that looks like an IRAM-style assessment: phases, threats, controls, likelihood language, and a polished summary. That only proves the model can write in the genre. It does not prove that the same method always ran, that scores were not invented, that evidence was used, or that a human gate existed before “done.”

The table below compares the main assurance attributes. ChatGPT and Custom GPT can help you draft; Riskonami is built to run and verify an operational risk assessment.

Assurance capability comparison between ChatGPT, Custom GPT, and Riskonami
Assurance attribute	ChatGPT	Custom GPT	Riskonami
Fluent IRAM-style / risk assessment prose	✓	✓	✓
Custom instructions and uploaded knowledge files	—	✓	✓
IRAM 2.0 assessment spine owned by the product (not the prompt)	—	—	✓
Server-enforced phases — model cannot skip or reorder the method	—	—	✓
Schema-validated structured outputs per phase	—	—	✓
Deterministic scoring / likelihood tables (server source of truth)	—	—	✓
AI limited to bounded enrichment (not the whole assessment)	—	~	✓
Evidence intake with noise / irrelevant media rejection	—	~	✓
Persistent assessment state and session memory	~	~	✓
Human-in-the-loop — product-enforced preview / accept before commit	—	—	✓
Human judgement remains the gate for risk acceptance	~	~	✓
Expert enhancements / expert validation of issued reports	—	—	✓
Attribution — distinguish AI enrichment, user edits, and expert input	—	—	✓
Provenance for assessment runs and report generation	—	—	✓
Version control of assessment / report artifacts (revision lineage)	—	—	✓
Report history — issued reports retained and listable per session	—	—	✓
User-controlled report issuance (no auto-publish of conclusions)	—	—	✓
Traceable findings linked to locked assessment artifacts	—	—	✓
Issued audit-oriented report (structured PDF / export path)	~	~	✓
EU data residency for assessment state (product-controlled)	—	—	✓
GDPR-aligned processing with limited server retention + export	—	—	✓
Optional EU-hosted LLM for enrichment (enterprise)	—	—	~
Automated unit + integration coverage of the assessment spine	—	—	✓
End-to-end scenarios that simulate real assessment journeys	—	—	✓
Fixture-based verification of AI enrichment quality	—	—	✓
Invariants: no invented scores/IDs; no silent method rewrite	—	—	✓

✓ = provided by the product · ~ = partial / depends on user discipline · — = not a product capability

Chat-as-assessment

The model is the assessment. Prompts and conversation history decide what happens next. Outputs are essays. Completeness is “it looks finished.” Comparability across runs is weak.

IRAM 2.0 product with AI enrichment

The server owns the method. AI assists inside bounded steps. Structured state is the source of truth. Experts preview and accept before a report is issued.

Riskonami is the second kind. AI removes repetitive structuring, drafting, and evidence organisation — it does not replace judgement, invent the methodology, or auto-publish conclusions.

What we encode: method before model

Assurance starts with a methodology you can defend, not with a prompt. Riskonami implements an IRAM 2.0-aligned assessment spine: organisation and system context, architecture, threat profiling, controls, likelihood, scenarios, treatment, and traceable reporting.

Structured assessment phases

Assessments follow a defined process with expected inputs and structured outputs — not open-ended prompting. Each phase has a purpose the platform enforces.

Evidence-aware assessment

Uploaded documents and system information inform assessment outputs instead of generic, disconnected analysis. Noise and unrelated material can be rejected rather than absorbed as “context.”

Model governance

AI is used inside a controlled workflow. The platform owns phase routing and state transitions. The model cannot skip phases, invent orchestration paths, or become the source of truth for scores and locked artifacts.

Assessment memory

Sessions preserve context so work can be resumed, refined, and reassessed as systems and controls change — without restarting from a blank chat.

See the product workflow

Accountability at the point of decision

Operational risk work is accountable work. The product is designed so experts remain responsible for judgement, validation, risk acceptance, and final report approval.

Human-in-the-loop review

Riskonami supports expert judgement; it does not replace it. Users review, validate, and accept outputs before finalisation.

Provenance, history, and attribution

Assessment runs and reports keep revision lineage and report history. Outputs can distinguish AI enrichment, user edits, and expert validation — so accountability is inspectable, not implied.

Final report control

Users decide when an assessment is complete and when a report is issued. Riskonami does not auto-publish conclusions.

Traceable report outputs

Findings connect back to assessment context, evidence, controls, and reasoning so reports support governance and audit — not an ungrounded narrative.

Data handling and residency

Assessment state and context are kept in Europe on Google Cloud (Netherlands / europe-west4), designed for GDPR-aligned processing. See the data and model section below for what stays in the EU versus what is sent transiently for AI enrichment.

Expert enhancements

Optional expert validation strengthens issued reports beyond self-serve AI drafting — a product capability ChatGPT and Custom GPTs do not provide.

GDPR, where data lives, and which models we use

Riskonami is built for GDPR-aligned operational use: we minimise what we keep, we store assessment records in the European Union, and we ask you to export anything you need for your own archive because server retention is limited.

Persisted in Europe (our control plane)

Accounts, sessions, worksheets, uploads, structured phase state, issued reports, and billing entitlements run on Google Cloud in the EU (Cloud Run, managed database, and private object storage in europe-west4). That is the durable product record.

Transient AI enrichment (default path)

When you affirm an AI enrichment step, bounded prompts and context are sent to the configured OpenAI model (product default via OPENAI_MODEL, typically a current GPT-class reasoning model) for that request only. Enrichment output is validated and then stored back in the EU assessment state. OpenAI is not the system of record for your assessment.

Optional EU-hosted LLM (enterprise)

For enterprise customers who need reasoning without leaving an EU processing boundary, we plan an optional EU-based LLM path (same bounded enrichment contracts, alternative provider/region). This is not the default free/self-serve path — contact us for enterprise terms.

Retention and cleanup

Forever-free and inactive sessions are intended for shorter retention (target 90 days after last activity). Paid and enterprise workspaces follow longer contractual windows. Always download reports you need to keep.

Billing, free plan, and retention Pricing

How we verify: software that can be tested

A serious AI risk product is still software. Method claims are empty if the implementation cannot be exercised. We treat assurance as an engineering obligation: cover the assessment spine with automated tests that fail when behaviour drifts.

Unit tests

Phase contracts, schemas, scoring rules, classification helpers, and report composition — the pieces that must stay correct in isolation.

Integration tests

Server-owned orchestration: phase transitions, locks, validation, enrichment boundaries, and persistence of structured assessment state.

End-to-end scenarios

Scripted and browser paths that simulate real assessment journeys — so the product path users take is the path we continuously exercise.

This is the difference between a demo that “worked once in chat” and a system you can regress-test when prompts, models, or phase logic change.

How we verify AI enrichment

We do not ask “is the model generally rational?” We ask a narrower, auditable question: for a known problem with known expected characteristics, did the AI stay inside its bounded role and produce output that is structurally valid, grounded, and materially aligned with what the method requires?

Shape and contract — schema-valid JSON, required fields, and phase-task completeness.
Invariants — no invented scores or IDs; no contradicting locked assessment state; no silent methodology rewrite.
Claims and concepts — required themes present; forbidden expansions caught (for example treating noise as architecture).
Semantic checks — similarity and judge layers against golden fixtures for enrichment quality, separate from orchestration tests.

Live enrichment quality is checked on fixtures and release cadence — not left as an untested hope that the next model call will “sound right.”

What we claim — and what we do not

We claim

An IRAM 2.0-aligned, server-orchestrated assessment product where AI enrichment is bounded, human review is required before issuance, outputs are traceable to structured state, and software plus enrichment verification support ongoing assurance.

We do not claim

That a language model alone is a complete risk methodology; that AI replaces expert acceptance; or that every free-form chat answer is an auditable assessment record.

Built for accountable risk work

See the shared assessment spine, or talk to us about how Riskonami fits your governance and assurance expectations.

See how it works Talk to us About Riskonami