HomeAI/ML Assurance & Certification Readiness

AI/ML assurance & certification readiness

Understand the AI.
Build the evidence.

Safety and certification planning for AI-enabled aviation systems and AI-assisted engineering.

Define what the AI is allowed to do, how its errors can affect the aircraft or the engineering work, and what evidence is needed to support its use.

Explore the FAA & EASA context

Start with the application

Two uses of AI.
Different assurance questions.

Select the context closest to your work. The starting point changes with the role of the AI and the consequences of relying on it.

A practical place to start

Assurance of AI-assisted engineering.

Requirements analysis, document review, test preparation and safety-assessment support can all use AI. The question is what error the tool could introduce or fail to detect, and how that error is controlled before it enters the product or compliance evidence.

We help define permitted use, review responsibilities, acceptance checks, source traceability and change controls. Tool qualification considerations depend on the tool's actual role, the credit taken and the applicable development process.

Example question

An assistant proposes a requirements-to-test mapping. Who checks omitted requirements, incorrect links and unsupported coverage claims before the mapping is used?

Work through an example →

Practical programme support

Turn the use case into
a reviewable assurance plan.

Aerospace Consulting brings systems, safety and certification judgement to the work. Specialist model evaluation or testing can be scoped with the client and relevant technical specialists.

01

Define the application and approval questions

Clarify function, users, system boundary, operational limits, model-update behaviour and the approval sought. Identify which aspects need early agreement with the authority.

Typical output Use-case and certification-planning brief, assumptions register and authority questions.

02

Connect AI behaviour to system safety

Examine erroneous, missing, late and inconsistent outputs. Review dependencies, human oversight, monitors, fallback and the safety requirements needed to support each claimed protection.

Typical output Safety integration assessment, architecture findings and allocated evidence actions.

03

Review the data, model and evaluation approach

Assess whether the proposed evidence addresses the intended conditions, data provenance, partitioning, performance limits and deployment configuration. Identify where specialist analyses or additional evaluation are needed.

Typical output Data and model assurance review, evidence gaps and a scoped evaluation plan.

04

Prepare the evidence and manage change

Link claims to requirements and results. Establish the baseline, open limitations and change triggers. Prepare a coherent technical position for programme reviews and authority discussions.

Typical output Evidence map, configuration and change-control plan, and assurance-readiness review.

These are proposed engagement outputs, tailored to the application. They are not a universal regulator-mandated document set. Certification decisions and delegated findings remain subject to the applicable authority, project arrangements and authorisations.

Evidence that supports the engineering claim

Look beyond a single
performance number.

Our review connects the model to its data, implementation, aircraft interfaces and operating assumptions. The required depth follows the application and safety consequences.

Function and operating limits

Define what the function may do, when it may be used and what conditions are excluded. Make those limits meaningful to the system and its users.

Data and evaluation

Identify provenance, labels, coverage gaps and separation of training, validation and test data. Assess relevant conditions and rare scenarios, not only an aggregate score.

Implementation and integration

Connect the trained model to preprocessing, sensors, software, compute platform, timing and downstream functions. Verify the configuration that will actually be used.

Protection and recovery

Examine detection coverage, monitor failures, intervention time, transition behaviour and recovery capability. Substantiate any independence credited between paths.

Human responsibilities

Establish what the pilot, operator or engineer sees, understands and must do. Review workload, timing and whether the person can detect the error being credited to them.

Controlled evolution

Track changes to data, model, thresholds, software and operating conditions. Define the evidence that must be revisited before a changed version is released.

A useful test of the argument

For every claim, ask: under which conditions, for which configuration, supported by what evidence?

Read our guide to connected safety evidence →

FAA & EASA context

Know the status
of the source you are using.

Sources checked . A roadmap, a concept paper, a special condition and a research report serve different purposes. Establish applicability and the proposed means of compliance for the project.

FAA

Start with the function
and its safety implications.

The Version I AI Safety Assurance Roadmap distinguishes a fixed, learned model from AI that changes during operation. It sets a strategic direction for incremental assurance work within aviation's safety framework. Roadmap source

The 2025 automation framework also focuses on the particular task and the roles of pilots and operators. AI and autonomy are not interchangeable descriptions of an application. Framework source

EASA

Separate applicable material
from emerging proposals.

SC-AI-01 Issue 1 is a final special condition with a stated CS-23 scope and references to Concept Paper Issue 01. Its application needs to be established for the project. Special condition source

Concept Paper Issue 2 is published guidance. Proposed Issue 3 and the DS.AI proposal broaden the developing framework, but remain proposals in the sources reviewed. Compare the source editions

Questions worth resolving early

The technical detail
changes the answer.

Does strong model accuracy establish aircraft safety?

Accuracy is one result for a defined evaluation. The safety argument also needs to address the kinds of errors, their operating context, duration, timing and system consequences. A benchmark percentage is not automatically a failure probability per flight hour. Related FAA research →

Can a monitor or a fallback make the AI acceptable?

It may support the architecture, but the credit depends on what it detects, how quickly it responds and whether the recovery remains effective. The monitor, switching and fallback functions have their own failure modes and assurance needs. Runtime-assurance research →

Are two different models necessarily independent?

No. Different architectures or suppliers may still share training data, requirements, preprocessing, sensors or operating weaknesses. Diversity is a design choice; the independence claimed in the safety argument needs substantiation.

What changes when a model is retrained?

The revised model needs an identified baseline and an impact assessment. Revisit the evidence affected by changed data, behaviour, limits and interfaces. Offline retraining followed by controlled release is different from adaptation during operation. FAA learned/learning distinction →

Does every AI engineering tool need qualification?

The answer depends on intended use, the errors it could introduce or miss, and the development or verification credit being taken. Establish the applicable process and how outputs are checked. Neither an AI label nor a generic promise of human review settles the question. Software and tool context →

How does this relate to ARP4761A and ARP4754B?

Safety assessment and systems development provide the structure for functions, effects, requirements, architecture and evidence. AI-specific data, learning and evaluation concerns need additional treatment. Establish the applicable editions and authority approach for the project. Read the safety assessment guide →

Research perspective

Follow the research.
Test its relevance.

LawZero's Scientist AI research explores approaches that limit goal-directed behaviour while supporting prediction and reasoning. It is relevant to questions about agency and constrained use, while an aviation application still needs evidence for its own function and safety consequences.

Explore LawZero research

Research reference only. No affiliation, endorsement or aviation approval is implied.

Begin with one defined application

A focused review.
A practical next step.

Bring the use case, available architecture, current evidence and the next decision you need to make. We can agree a defined review or ongoing programme support.

AI use-case & assurance review

A scoped engagement to identify the safety or process risks, review the proposed controls and establish the evidence gaps.

  • An agreed application boundary and assumptions
  • A prioritised findings and evidence register
  • A recommended work plan and authority questions