AI Documentation helps companies reduce manual document processing time, improve efficiency, and increase customer service. It also frees up resources to focus on innovation and customer needs.
Interviewees from banks, consultancies, and software developers point to challenges with describing AI solutions in such a way that they can be audited. These challenges arise because AI is different to traditional software.
Requirements for AI Documentation
Documenting AI is a complex task that requires careful consideration of several important requirements. These include the technical viability of AI projects, data security, and explainability of AI solutions.
For example, if you’re working in a regulated industry (like healthcare or finance), you should also account for data protection laws and ensure that your AI is compliant with them.
In addition, you should document how the AI works, so that new employees can easily learn it. This will help your organization to maintain its information systems in the long run and keep AI development agile.
In this study, we explored the requirements for AI documentation and interviewed 16 different interviewees. The results were grouped into four groups, and an interview guideline was developed to elicit the most essential and challenging requirements for AI documentation.
Interviewees’ Approaches to AI Documentation
In this study, we conducted semi-structured interviews with AI professionals from commercial banks. These professionals were able to provide valuable insights into the challenges that commercial bankers face when it comes to AI documentation.
We grouped interviewees into three groups: “prepared adopters,” who are risk-averse and apply AI in use cases with a high level of risk; “relaxed adopters,” who adopt AI without much fear; and “risk-averse non-adopters,” who do not adopt AI at all.
Prepared adopters apply AI in use cases with a high risk and make compliance requirements a top priority. They argue that AI is an excellent way to achieve this goal and present interesting workarounds to the challenge of applying AI in practice.
Recommendations for AI Documentation
Artificial intelligence (AI) is a promising technology that can revolutionize document processing and help businesses of all sizes automate time-consuming data entry, analysis, and reporting. However, AI is not without its limitations.
One challenge for AI implementation is its black-box character that makes it difficult to audit and validate AI applications. This is particularly true for regulated use cases, where software needs to be certified or validated before it can be deployed.
As a result, it is crucial to establish and implement documentation standards that ensure fairness, accountability, and transparency in the context of AI. But how do you design such standards?
The use of AI has never been more widespread, and as the number of applications increases, so does the risk of vulnerabilities. The biggest challenge is ensuring fairness, accountability, and transparency in AI-based systems.
Luckily, one way to ensure that these requirements are met is by documenting the AI. Aside from providing a solid understanding of the AI, documentation helps in verifying the quality and accuracy of the results generated by AI models.
For example, the source code of a machine learning (ML) algorithm can be interpreted as a storyboard that explains how the model has been trained and what data it used to do so. This is the most efficient and meaningful way to provide the required information in the context of ML algorithms.
However, a proper AI Documentation can be challenging to create. The current state of the art is based on software documentation guidelines, which may not be applicable to AI-based systems due to their unique features and characteristics.