Building Strong Foundations for your AI Services: How digital infrastructure can enable your data strategy

Data, particularly the critical information that encapsulates your organization’s unique intellectual Property (IP), is the lifeblood of Generative AI (GenAI). So, the digital infrastructure over which that data travels must be properly engineered to support the GenAI use cases on which your future depends. Your ability to achieve this will determine whether your infrastructure is an enabler of your data strategy – or a constraint.

Hiding amid the clamour about the possible impact of GenAI is an uncomfortable truth: the large language models to which most large companies have signed up are generic platforms.

By way of a metaphor, if two Olympic divers each perform the same complicated 4 1⁄2 somersault in pike position, the only difference in the results they achieve will be in how well they execute the manoeuvre. Similarly, suppose two companies in the same industry each take up a Co-Pilot subscription: in that case, the only difference in the results will be down to their ability to leverage an identical set of capabilities – the technology will not on its own deliver the breakthrough results that either is seeking.

The power of AI is only truly unleashed when it becomes part of the bloodstream of your company; when it is joined up with your Intellectual Property and used to drive the raison d'être of your company. That’s what we mean when we say, ‘the tools you have and the algorithms you’re deploying are no longer differentiators, but your data can be.’ But that’s a big ask.

At conferences, I poll my audience to see if they are comfortable with exposing the core of their company’s IP via an AI service. Not many hands go up. I don’t blame them. It would take either a brave person or a fool to take that step. The difference between the two is trust.

A big part of the solution or a bigger part of the problem

If you are betting the business on a high-bandwidth AI service, can you be sure that your network can cope with the traffic without reducing other applications to a crawl? And that your infrastructure can grow with your AI ambitions? Can you trust your security systems to protect vital company information as it traverses the network even though this presents a larger attack surface to potential bad actors? Can you be sure that the data your LMMs is trained on is of sufficient quality to prevent hallucinations that will tarnish your company’s reputation? If opening up access to every document in your company is a key use case, can you be sure that you have the robust privilege processes in place to ensure the integrity of sensitive information?

If you can answer ‘yes’ to all of these questions, then you should congratulate yourself because you are wise indeed: you have successfully operationalised a fledgling GenAI concept into a production-grade service. But very few companies are this advanced on their GenAI journey. If you answer’ no’ to at least one of the above, then you’re not alone, but you do need to proceed with caution. The point here is that if you can trust your infrastructure, then it becomes an enabler of your data strategy. If you can’t, then it becomes a constraint.

Keep calm and carry on

If you can’t do all of this today, then don’t panic. The recent furore around DeepSeek shows that the sands of AI can shift dramatically – and in unexpected ways. (Who could imagine that an unheralded Chinese AI start-up could wipe out over a trillion dollars in market value from companies like Nvidia, Microsoft, and Alphabet in a single day?)

Like a good idea that doesn’t work out in practice, don’t bin a projected AI service because it can’t be achieved today. Instead, put it in the ‘parking lot’ for later review. The AI market moves so swiftly that what is impractical or too expensive to deliver today could be perfectly viable in six months. For example, until recently, voice interaction with Chat GPT was considered too complex to be implemented – now it is available off-the-shelf from the Azure AI foundry.

Nuggets to go

Another good example of this is unstructured data – the documents, presentations, and emails that may hold golden nuggets of information. GenAI is capable of processing rich and varied types of information, ranging from instructions and notes to images and schemas. Before it can be used, of course, it will have to be digitized and stored (again, a trusted infrastructure will be an enabler of this process) and then absorbed within your overall data strategy. This is a non-trivial exercise, but it will be the basis of some potentially game-changing GenAI services.

The most obvious use case for this is maintenance: one can easily foresee field engineers arriving at a remote site and their voice-enabled GenAI-driven co-pilot identifying the malfunctioning equipment, asking relevant questions to the engineer to help diagnose the problem and then providing technical manuals and training videos that will ensure it is fixed far more rapidly than was previously possible.

Marketing provides a more surprising use case. CMOs are already imagining the creation of digital twins of customers to drive hyper-personalisation. Not only would this provide individual offers and incentives, but – because this is a generative form of AI – it could create text and images that were customised to that specific customer. Lots of data, both structured (e.g., product pricing and availability) and unstructured (e.g., customer social media feeds) would have to be loaded (and secured) into your AI model, but the impacts on marketing performance could be huge.

Neither of these use cases can be delivered overnight, not least because of the need for a robust change management process to ensure appropriate levels of adoption. Also, how do you assess the quality of this unstructured data? How do you sort the emails that have the nuggets from the ones that add no value? (We are working with our customers to help them find answers to these questions, but they are far from obvious.)

Nonetheless, if this is your vision, you can work towards it in stages – whatever challenges you face. Do what you can today and wait for the technology to catch up with your dreams – it will probably be sooner than you think.

Growing Pains

The lack of maturity in the industry can be evidenced by the preponderance of architects crafting today’s AI solutions. Generally, engineers significantly outnumber architects because software design principles are well understood, and most of the resource is applied to developing the application. Today, companies are still trying to understand how best to design AI services, so the ratio of architects to engineers is almost 1:1.

We are very conscious of these challenges and are doing our utmost to help customers overcome them. This is, for example, guiding the selection of partners for our AI ecosystem. As always, Orange Business is vendor-agnostic, and we will always work with best-of-breed vendors – but we look very favourably on those vendors that allow for a high degree of interoperability because we know our customers are struggling to integrate the components of their solution into a coherent structure.

For the agile organisation, change is an opportunity, not a threat

Thirty years ago, the first corporate websites were being put up on the internet. At that point, it would have been impossible to imagine the services – e-commerce, social media, streaming platforms, etc – that are now part of our daily lives. We are at that same point with GenAI today. All we can say for certain is that the future of GenAI will be different from what we imagine.

Tomorrow’s GenAI methodologies and tools will be very different from those we use today. We will witness breakthroughs we cannot predict today. And smaller language models will also emerge that require less compute power (and consequently lower costs and a reduced carbon footprint).  In that context, agility is important. We know from a worldwide survey conducted for us by GlobalData that lock-in with the foundational GenAI model of choice was a key concern (lagging only behind data management and legacy integration) – so GenAI leaders are clearly conscious of the need to keep their options open.  
The best advice I can give to anyone trying to operationalise a GenAI service is to be brave and build an infrastructure you can trust to deliver the data strategy you envisage. And, while it is important to keep your eyes on the prize that tomorrow will deliver, remember that (in the words of Mahatma Gandhi), the future depends on what you do today.