From AI pilot to production: 7 key factors for achieving ROI

Capgemini

2020-04-29

The AI bandwagon is gathering speed. Gartner estimates that AI global business value will reach $2.9 trillion in 2021.

But although most organisations accept that AI is key to growth and success, many struggle to reap actual benefits and achieve ROI.

Achieving trusted and sustainable AI

Introducing AI is a journey towards end-user adoption. It’s sensible to start with proof of concept and pilot projects. But unless these initiatives fit into a wider framework, they can stall. Stakeholders become disillusioned, putting the AI agenda in jeopardy.

And data science alone isn’t enough. To make sure your AI journey is a success, you need to understand these 7 critical factors before you start.

1. Understanding what’s different about AI

Understanding the key characteristics of AI and how these differ from other technologies underpins all the other factors required for success.

Data-centric thinking – Many IT systems focus on particular business functions, with data for each function kept in its own silo. AI solutions need data from a variety of sources. You may need to change to an enterprise-wide view of data and solutions.

Bias and unintended consequences – Machine learning models are probabilistic and only as good as their training data. Bias in that data means the AI will make biased decisions.

Changing relationship between people and technology – Customers and staff who interface with the AI need to understand and trust the AI’s decisions and rationale.

2. Developing a roadmap for sustainable AI

You need to develop a roadmap for your AI investments. There are widespread examples of use cases to improve customer experience, operations and the supply chain. But as well as identifying use cases and prioritising them, you need to prepare for the future.

Scaling AI to an enterprise scale needs a transformation mindset. How do you need to reorganise your architecture and data to scale and sustain your AI solutions? How should you address any concerns your people might have and prepare them for the change?

3. Developing trusted AI

AI and people work in different ways and it is that difference that makes the combination more powerful than either individually. But the combination will only work if people trust the AI.

AI can have a significant impact both on the people working in your organisation and your customers. The relationship between people and technology will progressively change as AI makes, or recommends, more business decisions. People need to understand how it reaches its decisions.

To be trusted, it must be ethical, understood and transparent. Using explainable (white box) AI can help by providing the reasoning behind decisions.

Your workforce can be your best asset – their domain knowledge can identify opportunities and test whether developed solutions are working successfully. But staff must be supported as they adapt to changes in their roles.

Because AI is different to previous technologies, organisational change is more important than ever.

4. Designing your architecture

AI solutions range from modules integrated into third-party products to bespoke AI with models designed to meet specific requirements.

You need to ensure that your architecture is data-centric to make the right data and AI components available as they are needed. You can’t get value from data held in inaccessible silos.

And the architecture must connect all the required components: processes, data engineering, model development and run-time AI.

As AI tools become readily available in cloud services, how does AI fit with your cloud strategy? To what extent do you want to use cloud-native capabilities?

Since the technology and related tools are changing fast, you should embed agile architecture skills and capabilities into your approach.

5. Analysing data requirements

AI solutions require vast quantities of data. But what data is needed and under what conditions is it available? There may be a range of privacy regulations or other access restrictions to consider.

6. Ensuring good governance and business assurance

You need good governance to champion AI initiatives, manage the impact on the business and define ethics to guide AI development.

The risks and challenges of AI, such as unintended bias, cybersecurity attacks and privacy breaches, can damage reputations and brands if not adequately addressed.

AI can impact the amount you automate business processes. Any changes to the AI model may have wide-spread operational impacts that you need to understand.

You can’t rely on testing alone. Business assurance is critical. You need to understand the impact on the business of implementing or changing AI models. And it isn’t a static activity—you need to understand how to provide feedback to models in order to improve them.

7. Updating your delivery model and testing

You need to adapt your delivery model for AI to incorporate elements like model development and data engineering.
Your AI testing regime needs to recognise that:

• The end-to-end process includes everything from the acquisition of data onwards
• The models used are probabilistic and their impact will only emerge from a statistically appropriate set of test cases
• Models can deliver unforeseen outcomes (like bias) if the underlying data is not appropriate

This degree of change from current practice has implications for delivery processes and governance as well as skills.

Conclusion

• Realising the full potential of AI is a journey.
• The 7 key elements discussed in this article can help you develop AI that is trusted and sustainable, with successful user adoption.

Author

Paul Walton is a digital transformation consultant and architect with 20 years experience at Capgemini. He leads an Automation and AI community in a team that combines specialists in business and technology architecture, organisational change, organisational culture and training.