Skip to main content

AI systems and impact assessments

An AI system is the governance unit above agents and teams. The risk registry classifies the objects that hold capabilities — an agent, a team, a function — but an auditor asks a different question: what AI systems do you operate, what are they for, and who owns them? An AI system record answers that. It maps to ISO/IEC 42001 controls on AI system documentation (A.6.2.9), defined use and misuse (A.6.2.10), intended use (A.9.3) and retirement (A.6.2.7).

Open Governance → AI Systems. Systems are listed in a tree grouped by lifecycle state. Retired systems aren't shown in the list.

Anyone with the Viewer role or above can view AI systems, their system cards and assessments. Changing anything needs the Editor role or above.

note

New organisations may already have AI systems in the list: FYAI publishes starter systems alongside its starter agents and teams, and each organisation gets its own copy when it's provisioned.

Register an AI system​

  1. Click the add (+) button at the top of the AI Systems tree.

  2. Fill in the form headed Register AI System:

    FieldRequiredNotes
    NameYes
    DescriptionNo
    Intended UseBefore going liveWhat this system is for.
    Declared MisuseNoWhat this system must not be used for.
    OwnerBefore a High or Critical system goes liveThe accountable person.

  3. Click Save. The system starts in Draft.

Attach agents and teams​

Select the system, then under Constituents pick agents in Agents and teams in Teams. Changes save as you make them.

The system's risk tier is the highest risk level among its agents and teams. It's recalculated whenever you attach or detach a constituent, and again before every lifecycle transition, so a system whose agent later gains a dangerous tool can't keep an out-of-date tier.

Evaluation​

The Evaluation section shows whether the system's agents and teams are tested, using their test cases (see Testing agents and teams):

  • A summary such as 2/3 tested, plus N failing and N stale chips where they apply.
  • One row per constituent with Tests (the number of test cases, None, or Unknown if they couldn't be read), Failing, and Last run (No tests, Never run, Today, Yesterday, or N days ago).
  • A constituent is Stale when its most recent test run is more than 90 days old.

Lifecycle​

Use Move to… to change the lifecycle state. Only allowed next states are offered.

FromAllowed next states
DraftIn Development, Retired
In DevelopmentDraft, Live, Retired
LiveDeprecated, Retired
DeprecatedLive, Retired
RetiredNone — retirement is permanent

Moving to Live is refused unless:

  • Intended Use is filled in;
  • for a High or Critical tier, Owner is filled in; and
  • for a High or Critical tier, the system has a current, approved impact assessment that isn't stale.

If a transition is refused, the error message says which condition wasn't met.

Retiring a system asks for confirmation: a retired system can't re-enter service or be edited. To bring the capability back, register a new system. Every change to an AI system, including each lifecycle transition, is written to its hash-chained change history.

System card​

Click System Card to see a card generated from the system's live configuration rather than a hand-written document. It contains the system's name, description, intended use, declared misuse, owner and lifecycle state; its declared and reconciled risk tiers and whether the declared tier is understated; each constituent agent (with its risk level, data classification, categories, owner, review date and tools) and team; the models its agents use; and a gaps block naming what's missing — no intended use, no declared misuse, no owner, no constituents, or constituents whose risk classification hasn't been reviewed.

Impact assessments​

An impact assessment records the effect of the system on people — a different question from security threat modelling, which assesses risk to the system. It maps to ISO/IEC 42001 A.5.3 and A.5.4 and follows the shape of ISO/IEC 42005.

Start an assessment​

  1. Select the system and, under Impact Assessment, click Start Assessment.

  2. Fill in:

    FieldRequired for approval
    Intended purposeYes
    Affected stakeholdersYes
    Potential harmsYes — choose from confabulation, harmful bias, data leakage, prompt injection, over-reliance, value-chain opacity, or type your own
    BenefitsNo
    MitigationsYes
    Residual riskYes
    Residual risk levelNo — critical, high, medium, low or informational

  3. Click Save Draft. The assessment is saved as a new version (v1, v2, …) in draft status, with you recorded as Assessed by.

A draft can't be edited from the app once saved; start a new version instead, or update the draft through the API (PATCH /ai-systems/{system_id}/impact-assessments/{assessment_id}).

Approve an assessment​

  1. In the assessment table, click Approve on the version you've reviewed.
  2. Confirm. The approved version becomes the system of record for going live.

Approval is refused if any required question is unanswered, or if you wrote the assessment — independent review is the point of the step. Approving a version marks any earlier approved version superseded. Approved and superseded assessments can't be edited.

When an assessment goes stale​

When an assessment is approved, the platform fingerprints the parts of the system that bear on its impact. If any of them changes afterwards, the assessment shows Stale and no longer satisfies the go-live check:

  • the system's intended use, declared misuse or risk tier;
  • an agent being attached or detached, or an agent's risk level or model changing;
  • an agent gaining or losing a tool, or a tool's risk level or approval override changing;
  • an MCP server changing the set of tools it advertises;
  • a team being attached or detached, or its risk level changing.

Renaming the system or editing its description doesn't make an assessment stale.

The table shows each version's Version, Status, Assessed by, Approved by and Stale flag. If the system can't go live because of its assessment, a warning above the table says why.