AI risk
AI builds the pipeline. Who checks the result?
SQACE is the independent verification layer, neutral to the producing system. It verifies the business correctness of AI-generated data pipelines and validates the outputs of embedded AI components, with documented evidence for supervisors and internal audit.
30 minutes, on your own use case.
The problem
AI changes the testing conditions.
When an AI agent writes a pipeline, the code is deterministic, but no one knows the rules the agent implicitly applied. When an AI component runs inside the pipeline, for example for classification or extraction, the result itself is non-deterministic.
Classical regression tests fall short in both cases: they check the behavior of the pipeline, while the business correctness of the results stays open. For high-risk AI, the EU AI Act, BaFin guidance and comparable frameworks require documented tests against clearly formulated expected requirements.
The SQACE solution
One independent verification layer, neutral to the producing system.
Verification against expected results. For deterministic sections, SQACE recomputes the business rules and compares with the output. This works even when the pipeline was produced by an AI agent: the agent does not have to explain what it did. The result must match the business requirement.
Validation of AI-produced outputs. For non-deterministic AI outputs, SQACE checks against business plausibility and consistency rules: sums must stay constant across buckets, extracted fields must match value ranges, distributions may shift only within defined limits. Violations are flagged for human review.
Scenarios
Where teams assure AI processing with SQACE.
Verify AI-generated pipelines against expected results
An agent builds a pipeline from a business specification. SQACE checks whether the pipeline actually implements the specification.
Validate AI outputs against plausibility rules
A component extracts fields from documents or incoming mail. SQACE checks value ranges, mandatory fields and consistency with the records, and sends violations into human review.
Build EU AI Act evidence for high-risk systems
For credit scoring, valuation or claims handling, SQACE documents the tests, the inputs, the rulebooks and the anomalies.
In concrete terms
- Business verification remains possible, even when AI built the pipeline.
- Validation delivers assurance where an exact expected-result calculation is fundamentally impossible.
- One platform for both modes, one continuous audit trail.
At a glance
| Pipeline type | SQACE mode | Result |
|---|---|---|
| Classical, deterministic | Verification | Full business assurance |
| AI-generated, deterministic | Verification | Output correct, independent of agent behavior |
| AI component, non-deterministic | Validation | Statistical assurance plus individual-case review |
Compliance
Compliance evidence as a by-product.
SQACE reports are directly usable for the EU AI Act (Art. 9 risk management, Art. 17 quality management), BaFin guidance on the use of AI, MaRisk AT 4.3.4 as well as DORA testing of ICT systems. For classification and extraction models, this additionally builds a data basis for drift monitoring.
How SQACE isolates data spaces and stores evidence is described in Security & architecture.
FAQ
Frequently asked questions
Is SQACE an AI tool?
No. SQACE is the independent verification layer that checks the result of a data processing against a business rulebook. It stays neutral to whether a human or an AI produced the processing.
Does SQACE replace model evaluation or MLOps tools?
No. Model evaluation checks model quality on test sets, MLOps monitors the deployment. SQACE checks whether the business logic of the entire data processing, in which the model is embedded, is correct.
Which models and providers are supported?
SQACE is model-independent. What gets checked is the result, regardless of which model or stack produced it.
How is non-determinism evaluated?
Via tolerance bands, distribution comparisons and defined consistency rules. What counts as a deviation is defined from a business perspective and versioned.
How are hallucinations handled?
Plausibility hallucinations typically show up as violations of value ranges or cross-references to master data. These hits go into human review.
How does this fit ISO 42001?
ISO/IEC 42001 requires an AI management system with documented controls. SQACE provides the control for the data-processing part, with an audit trail.