How to do AI if you are not Google

This is a guest blogpost by Matt Jones, lead analytics strategist at Tessella, in which he argues companies with physical products and infrastructure cannot simply cut and paste the tech giant’s AI strategy

Much written about AI seems to assume everyone wants to emulate Google, Facebook, or other companies built around data.

But many organisations look nothing like these tech giants. Companies in manufacturing, energy, and engineering – long standing, multi-billion-pound industries – derive revenues from physical products and infrastructure, not from targeting adverts at groups or individuals. Their data is usually collected from industrial machines and R&D processes, not people and internet spending habits. Their data collection is often bolted onto decades-old long lived internal processes, not built-in by design.

This type of data will deliver insights such as whether a factory can operate safely or predict the active properties of a new drug like molecule; not whether clicks turn into sales. This is very different from the insights that companies like Google are generating and looking at, and these pre-digital companies must take a very different approach to deriving benefit from AI.

CIOs at these companies can learn from the tech giants but trying to cut and paste their approach is a route to AI failure. Based on our work with companies built in the pre-digital age, we at Tessella recently produced a white paper outlining 11 steps that these pre-digital companies must take if they are to drive growth and stay competitive with AI. Broadly, these steps fall into three categories: building trust into AI, finding the right skills, and building momentum for AI programme delivery.

Trust is important

A key difference between the digital native companies and pre-digital enterprises is that the latter are often looking for very specific insights. Digital companies can afford to experiment and accommodate imprecision; a badly targeted advert will do a little harm. But an AI designed to spot when a plane engine or off-shore oil rig subsurface structure might fail demands absolute certainty.

Pre-digital companies cannot simply let an AI loose on all their data and see what patterns emerge, such unsupervised training experiments may provide estimations or suggestions, but they cannot be depended upon to inform an empirical solution. In these high-risk cases, there is a higher need to find the right data in order to effectively train AIs in a supervised learning regime.

Too many companies start by trying to pool all their data, perhaps looking admiringly at what Facebook and Amazon can do. For most, this is costly and unnecessary, at least in the short term. Companies should start by defining the problems AI can solve, identify the data needed to solve that problem, put people, technology and processes in place to collect and tag that data, then turn it into AI training data.

As AI is developed, there is also a need to maintain oversight to ensure the AI is delivering trustworthy results. Basic AI governance in high risk situations must include random sampling of AI outcomes and checking them against human experts for accuracy.

Finally, AI interaction, the user experience, must be intuitive, or it will not be taken up. AI decision support must take advantage of data visualisation and search technologies to ensure results are presented in meaningful ways. We can learn from digital native companies here, who are experts at making things easy for users: Google Photos runs neural networks, image analysis, and natural language understanding, but all the user needs to master is a search bar.

People not platforms

The temptation can be to completely hand over the problem to so-called data experts, or to buy in expensive technology platforms. But this misses an important point: that AI isn’t about spotting patterns, it’s about understanding what those patterns mean.

AI needs people who understand that data represents something in the real world – material strain, temperature readout, chemical reactions, maintenance schedules – and who can put together effective training regimes. AI should therefore be designed by people who understand the underlying data and what it represents within this business context. The best teams include representatives from IT, operations and business teams, domain experts partnered with embedded AI and data analytics experts who not only possess technical expertise but can also translate between these different roles.

We can again learn from the digital native companies. It is notable that these companies spend their budgets hiring the best people to design AIs which are right for them, not on buying in off the shelf technologies. Whilst the pre-digital companies will need different skill sets and more specific industry understanding in their AI teams, the focus must remain upon finding these right skills. This is the key to AI success, regardless of industry.

Build momentum

The digital native companies started from scratch and created the digital world, which they went on to lead. Longer established companies do not have this luxury – they come with decades of development in a pre-digital world, which has now been upturned and potentially disrupted. Many of their staff and processes are not ready for this new data driven world. They cannot just switch overnight; however ambitious their CIOs might be.

Such companies should set long term goals of digitalising processes and identifying where they see AI automating and advising. But they must work towards this goal determinedly and transparently keeping their people informed and engaged with the digital transformation; gradually shifting the business model and bringing existing staff with them on the journey. Starting too big without a carefully planned digital roadmap often undermines effectiveness and impact.

Pre-digital companies should initially focus on well-understood opportunities that can be executed quickly, with clear measurable milestones to demonstrate success built into their roadmap. This should be accelerated by running multiple agile AI projects in parallel, ensuring the best ideas are progressed rapidly. This will build a critical momentum for AI change programmes.

As they go, they should monitor their many AI projects, checking relative performance of each, immediately abandoning the bad ideas, and using successes (and failures) to improve training regimes. This agility is how digital companies deliver innovation but is lacking in many pre-digital organisations.

To summarise: physical enterprises undergoing digital transformation can and must harness the disruptive potential of AI. If they don’t, they will quickly be outpaced by competitors, startups or even tech giants with an eye on expansion. They start from very different positions to digital native companies. If they want AI to deliver business impact, they must mindfully find their own approach to people, processes, technology and management and form close, strategic partnerships with those that will build momentum behind an AI enabled digital transformation.