The advantages of curative AI in digital compliance
“No Creation without Validation!”
In the current era of digital content proliferation, ensuring content compliance against brand, business and regulatory guidelines represents a significant challenge for companies across various industries. This is particularly the case when the objective is to achieve compliance at speed and at scale, ensuring content is up to date with the evolving regulatory review and branding requirements across markets, business units and products.
This process can now be supported with AI, particularly curative AI, as it enables the implementation of a rules-based review system curated by human professionals to ensure safety and compliance standards are met.
What is curative AI?
It is a subfield of artificial intelligence that can extract information from assets and describe it in ways that allows further operations such as validation, alteration, transformation, and correlation with numerical data.
This process is entirely curated by humans from end to end, ensuring strict control over the data used to train the AI models and process execution, which is channelled by a set of pre-defined rules that reflect the regulatory and branding requirements. These pre-defined rules work as guardrails within which the data processing and execution is done. This allows us to create specialised AI models that are highly effective in executing asset validations in line with the required regulatory and brand standards.
Following six years of developing curative AI, we are proud to have begun co-developing it with Microsoft since 2019. During this time, we have gained considerable experience of working with clients who have multi-brand, multimarket and multilingual requirements, and we are excited to share our conclusions with you.
The advantages of curative AI in comparison to other AI genres
From a macro perspective, other forms of artificial intelligence (AI) are being trained on a broad spectrum of generic information, and not on specialised regulatory guidelines of a market, their interpretations, and/or branding guidelines of a specific company, market or other requirements. In effect, these AI models fail to have the required training to validate content against a set of unique requirements. Likewise for execution, the processing is not channelled to stay in line with the guidelines/regulatory requirements while being flexible and agile enough to adapt to the regulatory and market changes. This introduces a potential risk: we cannot tune the parameters of the process to such degree that we can ensure accuracy, sensitivity, and specificity of the output, which may have dire consequences. While such missteps may not have significant financial/reputational implications in the context of generative AI, they could have far-reaching legal and compliance consequences while reviewing content from a legal, medical, or branding angle.
This is why curative AI is the dominant force in this field. Firstly, these AI models are trained only on small sets of specialised data and are fine-tuned only on variations of these datasets. This approach creates a highly specialised model capable of exceptionally high accuracy, sensitivity and specificity levels in its outputs within a specific type of content.
As a rules-based system, curative AI is designed to be constrained by a set of predefined rules, which human experts in the relevant field curate prior to deployment. This approach ensures that the system does not act autonomously in any way but only in accordance with these predefined rules. The parameters of the rules can be adjusted whenever required to reflect the ever changing regulatory requirements. The capacity for machine learning is also subjected to rigorous regulation and is only allowed to operate under human supervision’s guidance and oversight. In this way, the AI system does not learn autonomously from the datasets it processes after receiving them from the users.
In practice, the process consists of two stages, both of which comprise of a collaborative effort between QA professionals, SMEs, and curative AI experts.
Initially, the foundations are laid by the experts from the brand or regulatory authorities by defining the requirements. Next, these requirements are translated and scoped into clearly defined and unequivocal rules. These rules are then matched with specific datasets and used to train the AI models. In this manner, the AI models are trained within a rigorously regulated setting curated by human experts.
The second stage comprises the fine-tuning of AI models using user inputs and experiences, as well as content pieces that have been subjected to the quality assurance process during usage. In this phase, the AI models are not permitted to learn autonomously like other forms of AI do. The user experience is collated and refined through the appropriate channels, approved by experts, and the model is then fine-tuned only in accordance with the approved inputs.
This human curating of the rules which govern the checks and the specialised datasets for training sets curative AI apart from other genres, and provides the level of security and comfort in terms of accuracy, sensitivity and specificity of the AI models.
Experiences in streamlining multimarket/multi-brand requirements
VML Enterprise Solution’s experience deploying curative AI for almost five years has demonstrated that AI is an effective tool for ensuring brand and regulatory compliance across various industries, markets, brands, products, and in up to 62 languages. The key to success is to deploy it with a holistic approach in mind.
One of the main drivers in successful Curative AI deployment is gaining a comprehensive understanding of the quality assurance (QA) processes employed across the entire business. This understanding facilitates the establishment of a set of fundamental, overarching QA principles – that is, rules – in collaboration with the relevant professionals in the field. These core rules can then be adapted to align with the specific regional, local, regulatory, branding, and language requirements when and where required. This approach allows for minimal effort that can be re-used, while ensuring quick deployment of the AI systems across the business. It also ensures that there is no duplicating of work throughout the organisation.
Consequently, instead of conducting a manual review of tens of thousands of asset items and over 100-page-long documents quality assurance professionals can instead assess the visual reports, in which the essential information has already been extracted and pre-checked by the highly efficient curative AI.
The automation of the QA review process is an extremely resource- and time-effective method that has reduced content review time by up to 92% and content review costs by up to 83%. This reduces the time to market for products, instilling confidence in the AI system’s capabilities.
The biggest gains have so far been seen in automating QC processes in regulatory requirements, where Pharma, MedTech, Lifesciences and highly regulated consumer packaged-goods have been the greatest beneficiaries.
Our almost five-year experience deploying curative AI for regulatory and brand review has given us the understanding and experience needed to ensure that with enough human expertise is provided in the early stages of training the AI models, curative AI can become a huge risk mitigator in content review automation unmatched by human reviews or other forms of AI.
