Reg. (EU) 2024/1689 - Articles 12, 19, 26, 50

Prove what your AI agents did. Every event, sealed.

AttestAI turns the agent traces you already emit into a tamper-evident audit ledger, then generates the evidence packs your enterprise buyers and EU regulators ask for. Point your OpenTelemetry exporter at one endpoint; keep shipping.

Request early accessRun the readiness checkBuilt for teams selling AI into banks, insurers, and healthcare.
AGENT TRACE / OTLP INLIVE
model.callgpt-5.5 / 1,204 tok
SEQ 0841sha256 9c41d2aa07f3b6e1PREV 62b0f5 - linked
tool.callpolicy_lookup
SEQ 0842sha256 5e88a1c94d02f7b3PREV 9c41d2 - linked
decisionpayout EUR 12,400
SEQ 0843sha256 b37f04e6a8c15d92PREV 5e88a1 - linked
human.approvalm.reeves approved
SEQ 0844sha256 e02c69fb31a4d875PREV b37f04 - linked
session.endoutcome: settled
SEQ 0845sha256 47da3b18c5f2e690PREV e02c69 - linked
LEDGER / APPEND-ONLY / RETENTION 183D FLOORCHAIN VERIFIEDHEAD 47da3b18

The problem

“How do you audit your AI agents?” is now a deal-blocking question.

If you sell AI products into regulated industries, the security questionnaire has new rows in it. Your observability stack was built for debugging, not for evidence: logs can be edited, retention is whatever the plan defaults to, and nobody can attest that the record is complete.

  • Enterprise procurement asks today. Vendor risk teams at banks and insurers already require documented agent audit trails, retention policies, and incident processes before they sign.
  • The EU AI Act makes it law. High-risk AI systems must technically allow automatic event logging over their lifetime (Article 12) and keep those logs at least six months (Articles 19 and 26(6)).
  • Debug logs are not evidence. A regulator or auditor needs to know the record was not altered after the fact. Mutable log stores cannot prove that; a hash-chained ledger can.

How it works

Ingest, seal, attest. Three steps from traces to evidence.

STEP 1

Ingest

Point your existing OpenTelemetry exporter at the AttestAI endpoint, or use the SDK. Spans that follow the GenAI semantic conventions (gen_ai.*) map to a typed event taxonomy: model calls, tool calls, decisions, handoffs, human overrides, disclosures.

STEP 2

Seal

Every event is appended to a per-project hash chain: sha256(prev, content). The database rejects updates and deletes at the trigger level, retention is floored at six months, and signed checkpoints anchor the chain over time. Tampering breaks the chain visibly.

STEP 3

Attest

Generate an evidence pack on demand: event coverage, human oversight activity, retention configuration, and a chain-integrity attestation, mapped to the EU AI Act articles they evidence. PDF for the auditor, JSON for the machine.

Try to tamper with it

One edited byte. Watch the ledger notice.

The claim your auditor cares about is not “we log things” - it is “nobody changed the record afterwards.” Mutable log stores cannot prove that. A hash chain can, and you can watch it happen.

Ledger excerpt / project claims-agentCHAIN VERIFIED - HEAD 47da3b18
0841model.call{ "intent": "settle claim CLM-2207" }9c41d2aa07f3b6e1SEALED
0842tool.call{ "tool": "policy_lookup", "policy": "P-88341" }5e88a1c94d02f7b3SEALED
0843decision{ "payout": 12400, "currency": "EUR" }b37f04e6a8c15d92SEALED
0844human.approval{ "approver": "m.reeves", "action": "approved" }e02c69fb31a4d875SEALED
0845session.end{ "outcome": "settled" }47da3b18c5f2e690SEALED

This is a live simulation of the verification pass. Edit the payout amount the way an attacker with database access would - then watch the chain answer.

Chain of custody

Five layers between your record and anyone who would rewrite it.

Integrity is not a policy document here; it is enforced in the database, in the hash math, and in a signature scheme - each layer independently checkable.

LAYER 01

Append-only at the trigger

The events table rejects UPDATE and DELETE inside the database itself. The single admitted exception is the retention purge: payload to NULL, everything else byte-identical. Even a compromised application server cannot rewrite history.

RAISE EXCEPTION 'ledger events are append-only'
LAYER 02

Canonical content hashing

Every event is serialized deterministically (versioned, sorted keys) and hashed. The content hash is stored beside the payload and survives it - purge the content, keep the proof.

content_hash = sha256(canonical(event))
LAYER 03

The hash chain

Each seal folds in the previous one, so the ledger is a single tamper-evident object, not a pile of rows. Appends are serialized per project with row locks - concurrent ingest cannot fork the chain.

event_hash = sha256(prev_hash || content_hash)
LAYER 04

Signed checkpoints

On a fixed cadence the chain head is countersigned with Ed25519. Checkpoints anchor the timeline: rewriting any earlier event would have to forge every signature that followed it.

sign(projectId : seq : head_hash)
LAYER 05

Retention with a statutory floor

The six-month minimum from Article 19 is a database check constraint, not a setting. Legal hold pauses purging; purged events stay verifiable through their hashes.

CHECK (retention_days >= 183)
scrypt password hashingkeys and sessions stored as SHA-256 onlyorg-scoped roles on every routeadministrative audit trailEU hosting by default

Evidence packs

Answers for the questionnaire, exhibits for the audit.

Each pack states what was recorded, how long it is kept, who intervened, and whether the ledger verifies - with the article references your reviewer wants to see. Regenerate any period at any time; the underlying record cannot be quietly rewritten.

AttestAI evidences your logging, retention, and oversight controls. It is not legal advice, and no tool alone makes an AI system compliant - we are precise about what we prove.

Evidence pack - excerpt

EU AI Act - Record-keeping and logs

Automatic event recording (Art 12(1))ACTIVE
Events recorded, period184,209
Distinct sessions traced12,441
Human oversight events (Art 14)326
Retention configured (Art 19(1))365 days
Six-month statutory floorMET
Ledger integrityVERIFIED
Chain
verified

The regulatory clock

Honest dates, after the 2026 Digital Omnibus.

The June 2026 omnibus deferred the high-risk obligations - it did not remove them, and it left transparency duties on the original date. Compliance projects that finish by these dates start now.

2 Aug 2025
In force: GPAI obligations and penalties.

Penalty framework active - up to EUR 35M or 7 percent of global turnover for the most serious violations.

2 Aug 2026
Article 50 transparency - NOT deferred by the omnibus.

Users must be told they are interacting with AI; synthetic content must be machine-readably marked. AttestAI records disclosure events so you can show it happened.

2 Dec 2027
High-risk obligations, Annex III systems.

Article 12 logging, Article 19 retention, Article 26 deployer duties, technical documentation, registration. Roughly 17 months of runway - about one enterprise procurement cycle.

2 Aug 2028
High-risk obligations, Annex I embedded systems.

AI embedded in regulated products (machinery, medical devices, vehicles) follows eight months later.

Sources: Regulation (EU) 2024/1689; Digital Omnibus on AI, final compromise text ST 9247/26, Council approval 29 June 2026.

Pricing

Priced like a tool, not like a GRC platform.

Enterprise AI-governance suites start at six figures. AttestAI is self-serve, transparent, and an order of magnitude cheaper - because passing a questionnaire should not require a procurement project of its own.

Pilot

$0 / 30 days

  • 1 project, 100k events
  • Full ledger and verification
  • 1 evidence pack
  • Community support
Start pilot

Scale

$999 / month

  • 10 projects, 10M events / month
  • 3-year retention and legal hold
  • Questionnaire answer library
  • SSO, API access
  • Priority support
Request access

Enterprise

Custom / annual

  • Self-hosted or EU-dedicated
  • Custom frameworks and DPA
  • Onboarding and audit support
  • Security review pack
Talk to us

Questions

Asked by every serious buyer.

Do I need to re-instrument my agents to use AttestAI?

No. If you already emit OpenTelemetry traces with GenAI semantic conventions - as LangChain, LangGraph, the OpenAI Agents SDK, and the Anthropic SDK integrations do - you point your OTLP exporter at the AttestAI endpoint and events start flowing. The SDK adds typed events (human overrides, disclosures, incidents) where you want richer evidence.

How is this different from LangSmith or Langfuse?

Observability tools are built for debugging: their stores are mutable, retention follows the plan you bought, and they make no integrity guarantees. AttestAI is built for evidence: an append-only, hash-chained ledger with enforced retention floors, signed checkpoints, and audit-ready exports mapped to regulatory articles. Many teams run both, and some feed AttestAI from the same OTel pipeline.

The omnibus delayed the high-risk deadline. Why start now?

Three reasons. Article 50 transparency duties still apply from 2 August 2026. Enterprise buyers already ask for agent audit trails in security reviews regardless of statutory dates. And an audit trail only covers the period after you switch it on - evidence for December 2027 starts accruing today, not in November 2027.

Can AttestAI make my AI system compliant?

No tool can, and we will not claim otherwise. Compliance involves risk management, data governance, documentation, and organizational measures. AttestAI gives you provable logging, retention, oversight, and incident-support evidence - the parts a ledger can honestly carry - and maps each export to the article it addresses.

Where does the data live?

EU-hosted by default, with self-hosted deployment available on the Enterprise plan. Payloads can be purged after your retention window while the chain remains verifiable - hashes stay, content goes.

Early access

Be audit-ready before your buyer asks twice.

We are onboarding a small group of design partners - AI product teams selling into regulated industries. Fifty percent off the first year, direct line to the founder, your questionnaire becomes our roadmap.

Request early access