“AI Guardrails” are often used as a method to not only get AI programs on track, but also as a way to accelerate AI investments. Projects and programs that fall within the guardrails should be easy to approve, govern, and manage – whereas those outside of the guardrails require further review by a governance team or approval body. The concept of guardrails is familiar to many tech businesses and are often applied in areas such as cybersecurity, digital initiatives, data analytics, governance, and management.
While guidance on implementing guardrails is common, organisations often leave the task of defining their specifics, including their components and functionalities, to their AI and data teams. To assist with this, Ecosystm has surveyed some leading AI users among our customers to get their insights on the guardrails that can provide added value.
Data Security, Governance, and Bias
- Data Assurance. Has the organisation implemented robust data collection and processing procedures to ensure data accuracy, completeness, and relevance for the purpose of the AI model? This includes addressing issues like missing values, inconsistencies, and outliers.
- Bias Analysis. Does the organisation analyse training data for potential biases – demographic, cultural and so on – that could lead to unfair or discriminatory outputs?
- Bias Mitigation. Is the organisation implementing techniques like debiasing algorithms and diverse data augmentation to mitigate bias in model training?
- Data Security. Does the organisation use strong data security measures to protect sensitive information used in training and running AI models?
- Privacy Compliance. Is the AI opportunity compliant with relevant data privacy regulations (country and industry-specific as well as international standards) when collecting, storing, and utilising data?
Model Development and Explainability
- Explainable AI. Does the model use explainable AI (XAI) techniques to understand and explain how AI models reach their decisions, fostering trust and transparency?
- Fair Algorithms. Are algorithms and models designed with fairness in mind, considering factors like equal opportunity and non-discrimination?
- Rigorous Testing. Does the organisation conduct thorough testing and validation of AI models before deployment, ensuring they perform as intended, are robust to unexpected inputs, and avoid generating harmful outputs?
AI Deployment and Monitoring
- Oversight Accountability. Has the organisation established clear roles and responsibilities for human oversight throughout the AI lifecycle, ensuring human control over critical decisions and mitigation of potential harm?
- Continuous Monitoring. Are there mechanisms to continuously monitor AI systems for performance, bias drift, and unintended consequences, addressing any issues promptly?
- Robust Safety. Can the organisation ensure AI systems are robust and safe, able to handle errors or unexpected situations without causing harm? This includes thorough testing and validation of AI models under diverse conditions before deployment.
- Transparency Disclosure. Is the organisation transparent with stakeholders about AI use, including its limitations, potential risks, and how decisions made by the system are reached?
Other AI Considerations
- Ethical Guidelines. Has the organisation developed and adhered to ethical principles for AI development and use, considering areas like privacy, fairness, accountability, and transparency?
- Legal Compliance. Has the organisation created mechanisms to stay updated on and compliant with relevant legal and regulatory frameworks governing AI development and deployment?
- Public Engagement. What mechanisms are there in place to encourage open discussion and engage with the public regarding the use of AI, addressing concerns and building trust?
- Social Responsibility. Has the organisation considered the environmental and social impact of AI systems, including energy consumption, ecological footprint, and potential societal consequences?
Implementing these guardrails requires a comprehensive approach that includes policy formulation, technical measures, and ongoing oversight. It might take a little longer to set up this capability, but in the mid to longer term, it will allow organisations to accelerate AI implementations and drive a culture of responsible AI use and deployment.
The challenge of AI is that it is hard to build a business case when the outcomes are inherently uncertain. Unlike a traditional process improvement procedure, there are few guarantees that AI will solve the problem it is meant to solve. Organisations that have been experimenting with AI for some time are aware of this, and have begun to formalise their Proof of Concept (PoC) process to make it easily repeatable by anyone in the organisation who has a use case for AI. PoCs can validate assumptions, demonstrate the feasibility of an idea, and rally stakeholders behind the project.
PoCs are particularly useful at a time when AI is experiencing both heightened visibility and increased scrutiny. Boards, senior management, risk, legal and cybersecurity professionals are all scrutinising AI initiatives more closely to ensure they do not put the organisation at risk of breaking laws and regulations or damaging customer or supplier relationships.
13 Steps to Building an AI PoC
Despite seeming to be lightweight and easy to implement, a good PoC is actually methodologically sound and consistent in its approach. To implement a PoC for AI initiatives, organisations need to:
- Clearly define the problem. Businesses need to understand and clearly articulate the problem they want AI to solve. Is it about improving customer service, automating manual processes, enhancing product recommendations, or predicting machinery failure?
- Set clear objectives. What will success look like for the PoC? Is it about demonstrating technical feasibility, showing business value, or both? Set tangible metrics to evaluate the success of the PoC.
- Limit the scope. PoCs should be time-bound and narrow in scope. Instead of trying to tackle a broad problem, focus on a specific use case or a subset of data.
- Choose the right data. AI is heavily dependent on data. For a PoC, select a representative dataset that’s large enough to provide meaningful results but manageable within the constraints of the PoC.
- Build a multidisciplinary team. Involve team members from IT, data science, business units, and other relevant stakeholders. Their combined perspectives will ensure both technical and business feasibility.
- Prioritise speed over perfection. Use available tools and platforms to expedite the development process. It’s more important to quickly test assumptions than to build a highly polished solution.
- Document assumptions and limitations. Clearly state any assumptions made during the PoC, as well as known limitations. This helps set expectations and can guide future work.
- Present results clearly. Once the PoC is complete, create a clear and concise report or presentation that showcases the results, methodologies, and potential implications for the business.
- Get feedback. Allow stakeholders to provide feedback on the PoC. This includes end-users, technical teams, and business leaders. Their insights will help refine the approach and guide future iterations.
- Plan for the next steps. What actions need to follow a successful PoC demonstration? This might involve a pilot project with a larger scope, integrating the AI solution into existing systems, or scaling the solution across the organisation.
- Assess costs and ROI. Evaluate the costs associated with scaling the solution and compare it with the anticipated ROI. This will be crucial for securing budget and support for further expansion.
- Continually learn and iterate. AI is an evolving field. Use the PoC as a learning experience and be prepared to continually iterate on your solutions as technologies and business needs evolve.
- Consider ethical and social implications. Ensure that the AI initiative respects privacy, reduces bias, and upholds the ethical standards of the organisation. This is critical for building trust and ensuring long-term success.
Customising AI for Your Business
The primary purpose of a PoC is to validate an idea quickly and with minimal risk. It should provide a clear path for decision-makers to either proceed with a more comprehensive implementation or to pivot and explore alternative solutions. It is important for the legal, risk and cybersecurity teams to be aware of the outcomes and support further implementation.
AI initiatives will inevitably drive significant productivity and customer experience improvements – but not every solution will be right for the business. At Ecosystm, we have come across organisations that have employed conversational AI in their contact centres to achieve entirely distinct results – so the AI experience of peers and competitors may not be relevant. A consistent PoC process that trains business and technology teams across the organisation and encourages experimentation at every possible opportunity, would be far more useful.
When non-organic (man-made) fabric was introduced into fashion, there were a number of harsh warnings about using polyester and man-made synthetic fibres, including their flammability.
In creating non-organic data sets, should we also be creating warnings on their use and flammability? Let’s look at why synthetic data is used in industries such as Financial Services, Automotive as well as for new product development in Manufacturing.
Synthetic Data Defined
Synthetic data can be defined as data that is artificially developed rather than being generated by actual interactions. It is often created with the help of algorithms and is used for a wide range of activities, including as test data for new products and tools, for model validation, and in AI model training. Synthetic data is a type of data augmentation which involves creating new and representative data.
Why is it used?
The main reasons why synthetic data is used instead of real data are cost, privacy, and testing. Let’s look at more specifics on this:
- Data privacy. When privacy requirements limit data availability or how it can be used. For example, in Financial Services where restrictions around data usage and customer privacy are particularly limiting, companies are starting to use synthetic data to help them identify and eliminate bias in how they treat customers – without contravening data privacy regulations.
- Data availability. When the data needed for testing a product does not exist or is not available to the testers. This is often the case for new releases.
- Data for testing. When training data is needed for machine learning algorithms. However, in many instances, such as in the case of autonomous vehicles, the data is expensive to generate in real life.
- Training across third parties using cloud. When moving private data to cloud infrastructures involves security and compliance risks. Moving synthetic versions of sensitive data to the cloud can enable organisations to share data sets with third parties for training across cloud infrastructures.
- Data cost. Producing synthetic data through a generative model is significantly more cost-effective and efficient than collecting real-world data. With synthetic data, it becomes cheaper and faster to produce new data once the generative model is set up.
Why should it cause concern?
If real dataset contains biases, data augmented from it will contain biases, too. So, identification of optimal data augmentation strategy is important.
If the synthetic set doesn’t truly represent the original customer data set, it might contain the wrong buying signals regarding what customers are interested in or are inclined to buy.
Synthetic data also requires some form of output/quality control and internal regulation, specifically in highly regulated industries such as the Financial Services.
Creating incorrect synthetic data also can get a company in hot water with external regulators. For example, if a company created a product that harmed someone or didn’t work as advertised, it could lead to substantial financial penalties and, possibly, closer scrutiny in the future.
Conclusion
Synthetic data allows us to continue developing new and innovative products and solutions when the data necessary to do so wouldn’t otherwise be present or available due to volume, data sensitivity or user privacy challenges. Generating synthetic data comes with the flexibility to adjust its nature and environment as and when required in order to improve the performance of the model to create opportunities to check for outliers and extreme conditions.
The process of developing advertising campaigns is evolving with the increasing use of artificial intelligence (AI). Advertisers want to optimise the amount of data at their disposal to craft better campaigns and drive more impact. Since early 2020, there has been a real push to integrate AI to help measure the effectiveness of campaigns and where to allocate ad spend. This now goes beyond media targeting and includes planning, analytics and creative. AI can assist in pattern matching, tailoring messages through AI-enabled hyper-personalisation, and analysing traffic to communicate through pattern identification of best times and means of communication. AI is being used to create ad copy; and social media and online advertising platforms are starting to roll out tools that help advertisers create better ads.
Ecosystm research shows that Media companies report optimisation, targeting and administrative functions such as billing are aided by AI use (Figure 1). However, the trend of Media companies leveraging AI for content design and media analysis is growing.
WPP Strengthening Tech Capabilities
This week, WPP announced the acquisition of Satalia, a UK-based company, who will consult with all WPP agencies globally to promote AI capabilities across the company and help shape the company’s AI strategy, including research and development, AI ethics, partnerships, talent and products.
It was announced that Satalia, whose clients include BT, DFS, DS Smith, PwC, Gigaclear, Tesco and Unilever, will join Wunderman Thompson Commerce to work on the technology division of their global eCommerce consultancy. Prior to the acquisition, Satalia had launched tools such as Satalia Workforce to automate work assignments; and Satalia Delivery, for automated delivery routes and schedules. The tools have been adopted by companies including PwC, DFS, Selecta and Australian supermarket chain Woolworths.
Like other global advertising organisations, WPP has been focused on expanding the experience, commerce and technology parts of the business, most recently acquiring Brazilian software engineering company DTI Digital in February. WPP also launched their own global data consultancy, Choreograph, in April. Choreograph is WPP’s newly formed global data products and technology company focused on helping brands activate new customer experiences by turning data into intelligence. This article from last year from the WPP CTO is an interesting read on their technology strategy, especially their move to cloud to enable their strategy.
Ethics & AI – The Right Focus
The acquisition of Satalia will give WPP and opportunity to evaluate important areas such as AI ethics, partnerships and talent which will be significantly important in the medium term. AI ethics in advertising is also a longer-term discussion. With AI and machine learning, the system learns patterns that help steer targeting towards audiences that are more likely to convert and identify the best places to get your message in front of these buyers. If done responsibly it should provide consumers with the ability to learn about and purchase relevant products and services. However, as we have recently discussed, AI has two main forms of bias – underrepresented data and developer bias – that also needs to be looked into.
Summary
The role of AI in the orchestration of the advertising process is developing rapidly. Media firms are adopting cloud platforms, making IP investments, and developing partnerships to build the support they can offer with their advertising services. The use of AI in advertising will help mature and season the process to be even more tailored to customer preferences.
