An assurance case is rarely created in one go. The argument matures with the system: it is developed, commented on, supplemented, reviewed and corrected, and each of these steps requires the time and effort of experienced engineers. Below, we show how this cycle can be accelerated by using AI tools connected to PREMIS via MCP. Importantly, users can work with their own AI tools that they are already familiar with, and these tools may already have knowledge of the system the argument is about.
AI support for work on assurance cases
The Model Context Protocol (MCP) is an open standard that allows language models to use functions exposed by an application. Through the MCP interface, PREMIS provides AI tools with the same operations that an engineer uses: reading the assurance case, creating elements with links, editing and moving elements, modifying relationships, and adding comments with a priority. The AI is not limited to producing content in the chat window; it can work directly on the argument the team is developing. The user still decides whether the AI only proposes changes or actually applies them to the argument in PREMIS.
The user controls the AI’s involvement in the work on the assurance case in two ways. The first is technical: permissions in PREMIS. If the access token grants read-only permissions, the AI cannot modify the argument. The second is the instructions the user gives to their own AI tool, which determine what the AI does in the chat and what actions it performs on the argument in PREMIS.
Using an engineer’s own AI tool also gives them the ability to manage access to documentation or system models that are used as evidence in the argument. As a result, the AI tool may already have knowledge of the system the argument is about when it starts working on an assurance case and connects to PREMIS for the first time. This allows the AI to develop and review the argument more effectively.
AI-supported activities
As a simple example, consider a functional safety argument containing a claim about hazard identification and analysis. Let’s assume that this claim is supported by only one piece of evidence, the HARA report. This example is shown in the video below. Although the HARA report is the main evidence and usually refers to the full hazard analysis documentation, relying on this report alone is a major simplification and leaves out any argument about the hazard analysis process itself. For example, the auditor is given no justification that all hazards have been identified and has to infer this from the report. As a result, the auditor has to analyse the HARA report themselves and can only accept or reject it as a whole. Such oversimplified arguments, which shift analytical work onto auditors, should be identified during reviews.
The first activity in which AI can support an engineer is reviewing the argument. The AI can carry out the review, while the engineer defines its scope and criteria. The results of the review can be presented in the chat or saved in the argument as prioritised comments. For the claim mentioned above, the review comment should point out the need to develop the argument instead of referring to only one piece of evidence.
The second activity is proposing and discussing changes. The AI presents a specific change proposal that can be discussed. The engineer can evaluate the proposed change, indicate the elements to be corrected, and finally accept or reject it. It is important for the expert to confirm whether the AI correctly understands a given part of the argument. On the one hand, the AI can point out weaknesses in the argument that a human has missed, and on the other hand, every comment and suggested change must be analysed and verified by an expert. A change can be discussed for as long as necessary before it is accepted.
Making changes is the third activity. Changes can be scattered and affect many places in the argument. Having the AI make such changes often saves considerable time compared with doing them manually. After making a change, the AI tool reports the scope of the change, for example by listing the added or modified elements. How the change is presented depends on its type: for modified elements, the before and after states are shown.
The next important activity is tracking how the review findings are addressed and verifying the changes made. The AI tool summarises the changes made, how many review comments have been addressed and how many are still outstanding. If the review findings exist only in the AI chat, progress can only be tracked there. If they are recorded as comments in PREMIS, progress can be tracked independently of the AI, including by other team members. Systematically handling review comments is as important as the review itself. To close the review, we need to be sure that each comment has been checked and either addressed or rejected with justification
Process efficiency
The main goals are, of course, to make the process more efficient and effective. Work on the assurance case requires the involvement of experts, so any reduction in their workload is a measurable saving. All work must be carried out under expert supervision, but some tasks can be successfully delegated to AI tools. In practice, a model in which AI advises on changes to the argument or checks them works well. To make the best use of experts’ time, it is also worth using AI to identify minor inconsistencies, reference or versioning errors, typos, and inaccurate wording. While these issues are of little substantive importance, they can be time-consuming, especially if detected late. Overall, the involvement of AI in assurance case work enables experts to focus more on the conceptual aspects of the argument and assessing its strength, rather than on mechanical editing tasks. The MCP interface in PREMIS makes this possible. PREMIS thus becomes a platform for discussion and collaboration between the engineering team and AI, ensuring high process efficiency.
However, these efficiency gains come with risks. Insufficient or absent expert supervision can lead to serious problems with the quality of the argument. If the certification body or regulator finds problems, the result may be rejection due to errors and misuse of AI, as illustrated by FDA Warning Letter 320-26-58 issued in April 2026. The benefits of using AI can only be achieved if it remains a tool for the expert, who is responsible for the quality of the assurance case at all times. AI can support the expert, but it is up to them to decide which tasks can be delegated to the AI. The role of AI is to be an assistant to engineers, and this approach is supported by PREMIS.