# See how regulated AI actually ships

> Static walkthroughs of artifacts Impetora delivers to clients. Not live SaaS demos. Not marketing renders.

Source: https://impetora.com/demo
Last updated: 2026-04-28

These are static walkthroughs of artifacts Impetora delivers to clients. A contract clause that gets parsed into obligations. An evaluation harness output. A support conversation with the audit trail intact. For a live conversation about your specific problem, book a discovery call: https://impetora.com/#discovery-call

## Showcase 01 - Document obligation extraction

A contract clause goes in. Obligations, risks, and quantitative thresholds come out, each with a citation back to the source line.

**Sample input clause (supplier agreement, section 9.3):**

> The Supplier shall deliver the Services in accordance with Schedule 2 and shall meet each Service Level set out in Annex A. Failure to meet a Critical Service Level for two consecutive months gives the Customer the right to terminate without penalty on 30 days' written notice. The Supplier shall maintain professional liability insurance of not less than EUR 5,000,000 per claim.

**Extracted:**

- Obligation: Deliver services per Schedule 2 and Annex A SLAs
- Obligation: Maintain professional liability insurance (>= EUR 5M per claim)
- Risk: Termination right after 2 consecutive Critical SLA breaches
- Notice period: 30 days written notice for termination
- Quantitative: EUR 5,000,000 minimum coverage

Production deployments cite back to clause line numbers in the source file.

## Showcase 02 - TRACE evaluation harness

Anonymised snapshot from an evaluation harness output. Each test case ties a regulatory question to an expected answer, the model response, a pass or fail badge, and the citation chain. This is the artifact we ship before going to production.

**Eval suite: regulatory-qa-v3 (5 of 6 passing)**

1. EVAL-001 - Pass. Under MiFID II Article 25(2), when must a firm assess suitability for a retail client? Citation: MiFID II Dir 2014/65/EU, Art 25(2).
2. EVAL-002 - Pass. What is the maximum administrative fine under GDPR Article 83(5)? Citation: GDPR Reg 2016/679, Art 83(5).
3. EVAL-003 - Pass. Does the EU AI Act classify a CV-screening tool as high-risk? Citation: EU AI Act Reg 2024/1689, Annex III(4).
4. EVAL-004 - Fail. What is the cooling-off period for distance contracts under EU Directive 2011/83? Source not cited; expected Dir 2011/83/EU Art 9(1).
5. EVAL-005 - Pass. Under DORA, who is in scope for ICT third-party risk obligations? Citation: DORA Reg 2022/2554, Art 2.
6. EVAL-006 - Pass. How long must KYC records be retained under AMLD5? Citation: AMLD5 Dir 2018/843, Art 40.

Real evaluation suites are sized to the client problem and re-run on every model or prompt change before promotion.

## Showcase 03 - Customer support triage

A scripted exchange showing how an AI agent triages a refund request, looks up the order, applies policy, and resolves without human escalation.

- Customer: Hi, I bought a coat last week and the lining is coming apart at the seam. Can I return it?
- Agent: Sorry to hear that. Items with a manufacturing defect are eligible for a full refund or replacement within 60 days of purchase. Could you share your order number so I can pull up the case?
- Customer: Order is #A-44218
- Agent: I see it: coat ordered 6 days ago, EUR 189.00. I am opening a defect return. You will receive a prepaid label by email in two minutes. Refund is processed when the warehouse confirms the defect, typically within 3 working days of arrival.
- Customer: No that's perfect, thanks.
- Agent: Thank you. Have a good day.

Eval-set metrics: 74% resolved, 42s median time, 4.6 / 5 CSAT. Production agents are tuned to client policy and tone.

## Pipelines

### Document extraction pipeline

Upload (source clause) -> Extract (bounded model) -> Review (human gate) -> Output (cited result). Every stage writes to the audit log.

### Evaluation harness pipeline

Test set (frozen cases) -> Run (candidate system) -> Score (pass / fail / cite) -> Gate (promote or block). Each run is logged with model version, prompt hash, and per-case scores.

### Support triage pipeline

Intent (classify) -> Lookup (order data) -> Policy (apply rules) -> Resolve (or escalate). Resolution and escalation paths both write to the audit log.

## TRACE methodology

Every showcase on this page was built using TRACE. Trust, Readiness, Architecture, Citations and Evidence are visible in every artifact, not bolted on. Read the methodology at https://impetora.com/methodology.

## FAQ

**Are these real Impetora systems?** Each showcase is a static walkthrough of an artifact we ship to clients. Sample data, not client data. Production runs on customer documents under their data residency.

**Why static walkthroughs and not a live demo?** Live demos hide the audit trail. We prefer static snapshots with input, output, citation chain, and pipeline visible in one frame.

**Can I get a custom walkthrough on my data?** Yes. Book a discovery call. We typically prepare a recorded reference on a sample of your data before the call.

**Do these systems work in languages other than English?** Document extraction and support triage have been deployed in English and Lithuanian. Other languages on a per-engagement basis.

## Contact

Email: info@ainora.lt
Discovery call: https://impetora.com/#discovery-call