Artificial intelligence (AI) is the transformational technology of our time. Faculty was founded on the belief that AI must be made real. To achieve its promise, it must be taken out of the lab and reach production systems in companies. Many companies are talking about the subject. Some are moving forward with small proof of concepts. While others are building exciting things but failing to deliver changes promised. Few have succeeded in creating business value.
To do data science properly you need to combine business strategy with deep, technical data science delivery expertise. This avoids data first strategies failing to address problems that matter. And top down traditional strategy analysis failing to get beyond powerpoint slides.
In the last two years we have designed AI strategies for a wide array of organisations, ranging from venture-backed start-ups building their business on AI, to large FTSE 100 enterprises in sectors ranging from fashion retail, to water utilities and financial trading. In the public sector, we are proud to have recently worked with the UK’s Office for Artificial Intelligence and Government Digital Service to identify the most significant opportunities to introduce AI across government with the aim of increasing productivity and improving the quality of public services.
Over the course of successfully completing over 300 data science projects, we’ve begun to codify what separates those that are able to systematically achieve business value with AI. This can be summarised by developing AI strategies that:
- focus on value
- prioritise speed to impact
- build capability, not dependency
These can be broken down into the following 9 lessons.
Focus on value:
1. Stop innovating: AI strategy is too often confined to the remit of isolated innovation teams, or the ambitious ‘horizon three’ of digital roadmaps. In our experience, a focus on innovation absent of clearly defined value levers rarely leads to successful return on investment. The only way AI will achieve its promise is when it is tied to core business strategy and has accountability to the executive team.
2. Find the intersection between your data and the problems that matter: Good strategies find the meeting place between your data, your important challenges and what’s possible with data science today. Start by getting a comprehensive view on your data assets and the extent to which they support the application of machine learning to solve important business problems. For some problems you may need to start collecting new data, or integrate with external data sets.
3. Remember it is data SCIENCE: Good data with bad science will result in failure – models that overfit, misjudged conclusions and results that don’t generalise. Place your bets on scientific rigour and real world testing over the promise of a large clean dataset alone.
Prioritise speed to impact:
4. Build working AI software not proofs of concept (POCs): POCs deliver little practical value, and can often deceive by disguising the gap between a vision and implementation. Instead focus on finding the subset of problems that can be immediately solved by building and deploying AI software. This proves value to the organisation, builds traction and develops the institutional reflexes that only come from working through an AI project end to end.
5. Don’t wait for your data transformation projects to finish: The requirements of AI models and data scientists are nuanced and specific to the problems at hand. Aggregating and simplifying data within structured databases without considering the value of the signals in the raw data can often limit the ability of data science to deliver results. What may seem like a logical warehouse design decision (to, for example, drop this feed, or aggregate these values) may prevent data scientists from building the best models. Build your data infrastructure in parallel with getting started with modelling – this way the data scientists can iteratively feedback requirements to your engineers and your infrastructure to support production AI from the outset.
6. Breakthroughs take time, so create a portfolio of investments: AI can provide significant competitive advantage for a business, even create breakthroughs that change the market entirely. But all breakthroughs that matter take time. Therefore, focus on building a portfolio of investments. Shorter term investments should seek to optimise core business, unlocking cash for pursuing more risky, game changing research.
Build capability, not dependency:
7. Build bespoke and be wary of out of the box solutions: Every business has unique customers, challenges and data. Out of the box solutions will not be tailored to the unique nuances of your situation, risking averaging down performance to the lowest common denominator. For anything core to the way you do business consider remember advantage isn’t gained with all your competitors using variants of the same model.
8. Invest in the best tools for your data science team: At each stage of the workflow, data science teams face a number of challenges ranging from building and deploying machine learning models to running them in production and ensuring ongoing prediction accuracy. These difficulties are exacerbated in large organisations, with the need for stable, scalable and secure processes often conflicting with the desire for rapid iteration and feedback. The result is a technical stack that is complex, and different to construct. Data science platforms like Faculty Platform can dramatically streamline the work of data scientists, freeing them up to focus on delivering value to the business.
9. Demystify data science for the business: Bring the business with you on your data science journey. Start by educating the executive team on what AI is, and what it can (and cannot do) for the business. This helps ensure everyone starts from a common understanding and expectations are aligned from the outset. As you build momentum, consider training your existing analytical teams in the basics of data science and help everyone speak with a common language and approach to solving problems with data.
If you would be interested in finding out more about how we might be able to help you on your AI journey get in touch.