Seeing the wood for the trees: 5 tips for executives looking to get real business value from data and AI

Commercial use of what has now been labelled as ‘artificial intelligence’, or AI, has become the hot topic of the past few years. But if it is to be more than a fad, AI will have to deliver on the hype many of its supporters have ladled upon it.

First, we should define our terms. AI is not just large language models (LLMs) such as ChatGPT, or even deep neural networks in general, which can also be used for image generation by the like of DALL-E or Stable Diffusion. In my opinion, what differentiates ‘AI’ from plain old machine learning (which some of us remember as the hot topic of the 2010s) is that it is wrapped in a user interface (i.e. a digital product) that enables the user to solve a problem. It is this useability-focused definition, usually achieved by chaining several types of machine learning models or approaches together, that could make modern AI such a powerful tool for business.

However, this potential is balanced by the fact that deployment of AI is at least as difficult as any other major digital innovation, for a variety of technical and cultural reasons. Early data suggests that AI implementation programmes have failure rates far above the average for digital transformation initiatives. I have seen suggestions that as many as seven out of every eight AI projects fail.

So the challenges are substantial. From talking with clients and potential clients, I believe that many companies are currently in a process of investigation and exploration with AI tools. The question is: when organisations move from discovery to delivery, how can they overcome the odds of failure and ensure that they are able to deliver the kind of transformational change that is promised?

Below are some reflections, drawn from my experience as a consultant, on how executives should think about the opportunity to use AI in their business, to strategically build long-term, sustainable value.

1)      Think holistically about the potential use cases for digital technology and innovation in your organisation – not just about AI

In order to think strategically about use of AI, don’t start by thinking explicitly about ‘AI’, instead think about how you could use data and digital technology (including AI) to better serve, and create value for, your clients, customers, or service users. The reason for this is that AI uses data as its raw material and sits within a digital ecosystem, and it cannot be disentangled from this context. This means you will need to take account of the digital maturity of your business in a wider context. AI can’t be used as a ‘cure-all’ for legacy technical debt, and if built on inadequate foundations, it will be likely to disappoint.

It is especially important not to have our view on AI dictated by whichever new piece of technology is flavour-of-the-month, seemingly taking the world by storm. The rapid advances in large language models (LLMs) in recent years has led to embarrassment for more than one large firm when, after their digital transformation centred around deploying LLM-driven chatbots, they have been shown to be overhyped and not ready for primetime. 

Instead, think (without jargon and in detail) about how data, AI and digital technology could change how your business runs – the actual specific processes and operations. Then derive a strategy by building up from there.

If this is a chatbot, great! But it probably isn’t just a chatbot. So apart from LLMs, the analytical techniques that anchor your highest value data use cases could also involve:

• A single customer view – this involves standardising customer data, building data-driven typical customer profiles (for example, by using clustering techniques). It often involves enriching your data with third-party sources (see point 3 below). It is a core sales and marketing asset for retailers and other companies selling to large numbers of consumers.

• Digital twins – these are models that encode every detail of your product in software. This enables use cases such as predictive maintenance, scenario modelling or stress testing. Particularly relevant to manufacturing businesses, especially those with complex products.

• Planning and optimisation algorithms – These are useful in any use case that involves achieving an objective with a fixed set of resources and known constraints. They are especially relevant to Manufacturing and logistics companies. It a used in a range of scenarios, including supply chain optimisation, financial planning, and resource planning.

This line of thinking often leads to the conclusion that often you don’t need ‘AI’ per se. What you really need is better RPA/automation, supply chain analytics, customer journey models, process control, or security. These could involve lots of technologies, from image recognition, to NLP, to web design for a better UI/UX.

2)      Getting to grips with your messy, unstructured data is often where the really big wins come from

Since I started working in digital transformation, I’ve seen many CTOs grimace when I ask about their unstructured data. In a traditional technology function, unstructured data is the unloved stepchild – shoved into storage, shut out of the data warehouse, rarely thought of, and even less appreciated.

But, in AI, unstructured data is the gold mine. It was commonly under-utilised because it was messy, and difficult to extract value from. But, as ChatGPT has shown, clearly this is no longer true of text data, but other kinds of unstructured or semi-structured data, such as images, video, audio, are also easier than ever to manipulate. In fact, libraries and tools to manipulate and bring order to this data may be the defining accomplishment of the past decade in AI. It is now almost trivially easy to effectively summarise a large text corpus, or build an object detection algorithm for your images.

I have undertaken numerous projects involving never-before used unstructured data that have delivered extraordinary value. In one case, I was able to use a text mining algorithm, to analyse tens of thousands of responses to a survey on a pending reorganisation, for a large government department. By identifying specific key complaints I was able to provide senior leaders with the knowledge to defuse many of these complaints with minor changes to their plan that did not affect the impact of the reorg, but which reduced the cost of staff turnover in the next 24 months by several tens of millions of dollars. So my strong recommendation is: don’t be put off by messy data. There’s gold in them there hills!

3)      Investigate what other sources are out there to enrich your data

Open data, and proprietary third-party data, can be invaluable for enriching your internal proprietary datasets. Think about how you can use both open and internal/proprietary data and integrate them both. The answer to this question will probably require getting your data experts and your business/ops experts in a room together.

Open data often comes from governments and international organisations. This could include environmental or transport data for a country which can be invaluable for any organisation operating there.

Proprietary datasets can often be purchased across a wide variety of industries – it is often worth doing some research to understand what is out there in yours – you might be surprised. Commercial companies will often be able to provide rich datasets of potential clients in your industry.

Corporations such as Dun & Bradstreet have datasets that might allow you to better understand your upstream supply chain risk exposures and therefore avoid excessive dependence on a single supplier, even if indirectly. Alternatively, private surveys of procurement managers in your industry could be a way to gauge the confidence of your customers and may predict changes in their demand.

In addition, high quality datasets on financial markets are available from providers such as NASDAQ and Bloomberg. This can be invaluable for companies with significant exposure to interest rates, or raw commodity prices. Financial data published by public companies can also be of interest for reasons of competitor intelligence.

4)      Remember the unglamourous stuff too

Although it is usually the focus on a company’s digital transformation and AI effort, front of house (production, operations, sales & marketing, commercial) is not the only game in town. Much of your business’ costs will be hidden in non-core, non-revenue producing functions. In my experience, legal, HR, finance, commercial, facilities management, and other administrative and support functions often make up 20-30% of headcount (and cost base), or more.

Since senior executives of a business tend to be expert in, and think most naturally of, the revenue-generating functions, these also tend to receive the most attention when the firm thinks about technology investment. It is not unusual, in my experience, to find a manufacturing companies running the latest CNC precision machine tools, while their HR and finance offices struggle with obsolete processes, manually manipulating spreadsheets for updating quarterly forecasts or administrative records.

The lesson is do not neglect the ‘back office’. Taking cost out of your business adds value because it allows you to expand production, and/or cut the unit cost to consumers. The advantage of looking at your support functions is that, since almost every business has them, many of these areas will have mature, AI-driven software that make it easy to take off-the-shelf solutions that could solve many of your business needs.

5)      Proofs of Concept (POCs) save a lot of money, time and wasted effort

Finally, whatever you decide, you probably want to set up a test or a trial first. Don’t assume your ideas will work first time. Develop a POC or introduce a trial (sometimes called a pre-release, Alpha or Beta product depending on its maturity), prove value, then expand the user base and develop the product.

Don’t try to boil the ocean first time. Try to prove the concept, using 1-10% of your likely budget. Rather than authorise a £3mm, 24mo programme of work, first prove the concept with a 3-month pathfinder project for £30-300k. This derisks the project. Over the medium term, this will save a lot of money, blame games and headaches.

A well-built proof of concept will enable your team to understand the technical problems to be solved (including the unanticipated ones), identify skills gaps in the team that could impede delivery, while also helping the user to ‘try before they buy’ – by using a minimal version of the product, potential issues and impediments can be identified faster and more reliably.

It is very important that the POC be carefully thought out. It should include, literally, just the components necessary to ‘prove the concept’ – i.e. to solve the key challenges and demonstrate the solution in principle. In the data space, this can be difficult to scope, since much of the work is in data engineering and manipulation, model building, data cleaning and quality validation.

Therefore, compromises must be made – using only a subset of the data, for example, or using a static dataset initially. The exact functionality required will depend on the ‘concept’ being proven. If your deliverable is ‘an app that provides critical real-time information on a customers’ utility usage for field staff’, showing the ability to deliver the information in real-time seems critically important. But in other instances – such as ‘developing a mobile app for senior managers to convey key information about the business’, an overnight data refresh may be sufficient, and more traditional elements of the user experience are key – is the app easy to navigate? Are graphs and charts clearly labelled and easy to understand and manipulate. Can the charts be attached to, or linked in, an email?

Conclusion

I’ve laid out some of the key issues I’ve observed in organisations attempting to use AI to achieve real transformation in their business. These points are far from comprehensive, and there are numerous additional factors to consider – the critical importance of recruiting the right team, with the right skills and the right leadership being a critical one. But hopefully they provide the reader with something to reflect on as they think through the potential role of AI in their business.

While hype cycles come and go, I believe that, if properly handled, AI has the potential to have truly transformational positive impact on our economies and societies.