AI Trust Center

The AI Trust Center is the most important document Attestly generates. It's the document your enterprise prospects ask for in their security review and the document your EU regulator may eventually ask to see.

What it contains

Section	What's in it
Executive summary	A 2–3 sentence overview, generated from your stack.
AI systems	One row per AI system: provider, model, inputs, outputs, risk class, human-in-the-loop.
Annex IV documentation	General description, data and inputs, monitoring/oversight, post-market monitoring, change management — per system.
Human oversight	Concrete oversight measures, opt-out path, override mechanism.
Evaluation	Bias and fairness, robustness, accuracy metrics, red-teaming posture.
Upstream provider terms	Training opt-out, data residency, provider SLA.
Data flows	Source → AI provider → output, with category labels.
Training-data policy	A statement on whether customer data is used for training (defaults to "no").
Acceptable use	Prohibited inputs/outputs, automated abuse detection.
Incident response	Classification (P0–P3), detection, response SLA, customer-notification commitment.
End-user disclosure	What users are told about AI processing.
Child safety	Statement on minors' data.
Conformity	Whether Annex IV technical documentation applies.

Every row carries a source field — see Source citations.

Risk classification

Attestly assigns each AI system a default risk class using the following heuristics:

1. If the system processes special_category data (Article 9 / Annex III) → High.
2. If the system is in the credit, hiring, or law-enforcement domain (heuristic on file paths) → High.
3. If the system generates user-facing content with no human review → Limited.
4. Everything else → Minimal.

You can override the default per-system in the dashboard. The override is preserved across regenerations as long as the underlying detector key doesn't change.

Schema

The generator output is constrained to this Zod schema:

const aiTrustCenterSchema = z.object({
  intro: z.string(),
  aiSystems: z.array(
    z.object({
      id: z.string(),
      purpose: z.string(),
      modelProvider: z.string(),
      modelName: z.string(),
      inputs: z.array(z.string()),
      outputs: z.array(z.string()),
      riskClass: z.enum(["minimal", "limited", "high", "unacceptable"]),
      riskRationale: z.string(),
      humanInTheLoop: z.boolean(),
      source: z.object({ filePath: z.string(), line: z.number().nullable() }),
    }),
  ),
  dataFlows: z.array(
    z.object({
      from: z.string(),
      to: z.string(),
      categories: z.array(dataCategorySchema),
      purpose: z.string(),
    }),
  ),
  trainingDataPolicy: z.string(),
  conformity: z.object({ annexIVApplicable: z.boolean(), notes: z.string() }),
});

The model cannot introduce a field that isn't in the schema. This is the single most important safety primitive in Attestly — it's why the same scan produces the same document every time.