As the UK government continues its push towards AI deployment and development, investments in infrastructure and AI skills have been paramount to achieve its AI ambitions. 

Some of the numbers being thrown around are quite staggering: It was revealed in January that the UK AI sector was attracting £200 million pounds in investment… every single day.

Then, just this year, the first ever UK-US Tech Prosperity Deal revealed £31bn in AI infrastructure investment, with Nvidia CEO Jensen Huang predicting that it will make the UK an “AI superpower.”

This is all good and well – of course we need the foundation and infrastructure to enable innovation but there’s a huge elephant in the room.

Data underpins AI. It’s not unrealistic to even say it almost completely dictates the usefulness (or potential for misuse) that the technology represents. It’s a vital bedrock that is still often misunderstood, overlooked, and underappreciated in the rush to slap on an ‘AI-enabled’ sticker on any given product or service.

The fact that data has such an overwhelming bearing on this technology so poised to be transformational surely means that data governance is front and centre of the discourse? 

Unfortunately, not so. Data governance is falling behind compared to AI implementation, leaving a critical gap in AI policy, sabotaging return on investment, eroding trust, and preventing AI from reaching its full potential. 

To delve into this critical gap, DIGIT spoke with Stuart Harvey, CEO of Datactics, a data management and governance firm, to discuss data governance best practice and the importance of strong data foundations when it comes to AI. 

The Plumbing 

AI is sexy – fresh, new, attracting a lot of attention from investors, driving the stock market in a tizzy, and governments can’t get enough of it. Everyone is talking about it, for better or worse. 

Data governance? Less so. Stuart Harvey likens it to plumbing – foundational, integral, but ultimately less exciting than flashy large language models or copilots. Still, without plumbing, you’re going to end up with a lot of, well, problems. 

Proper data management ranges from unifying and consolidating fragmented data systems and data silos, to managing data bias and limitations. 

It involves a master data management system, a data lineage system, and a quality matching system which Harvey believes is essential. 

“This extracts the critical data element from the data catalogue,” Harvey explaines. “This means you know how important the data is, you know who owns it. You know what it is in terms of the information type, it shows you the critical pieces in the data journey. 

“And so if we can measure data, it’s timeliness, it’s accuracy, it’s criticality at each step through that journey.”

All of this essential data ‘plumbing’ means that any AI system relying on this data is auditable, accountable, and explainable. These three key elements are essential to making an Ai model responsible and trustworthy. 

But these data plumbing foundations are often being overlooked or not neglected. 

Data is proving to be something of a headache to organisations trying to realise the full potential of their AI investments. 

Fragmented data from multiple legacy systems and departmental silos can cloud the ability of AI to process all the data it needs to make important decisions and truly revolutionise industries. 

Without this, Harvey says AI is simply not achieving its full potential. 

“I think there’s two ways of looking at AI: one is as machine learning – it’s like a new tool, a new feature. It’s the equivalent of say, spell check in Word 95,” Harvey started. 

“Or, it’s a paradigm shift where it completely replaces a business function or the people associated with it.”

What Harvey has seen so far with clients – which range from large financial institutions to the Home Office and other UK government departments – is the former, rather than the latter. 

“It’s an extremely useful tool for productivity and software development, maybe to process and summarise documents, but that’s a feature extension.” 

And this could be what is stalling ROI when it comes to AI – people are spending millions on AI because it’s shiny and new, but it is not necessarily transformational. 

“My concern is that people are going full tilt on adopting GenAI and LLMs as a solution to problems, but these tools can be rather opaque and lack the ability to join up the breadcrumbs. 

Datactics works largely with organisations at the forefront of governance and regulation – large financial institutions that must have strong security and handle mass amounts of data, and government organisations that are perhaps bearing the brunt of the ethical dilemmas that come with AI’s hasty deployment. 

What it comes down to, according to Harvey, is how AI is being implemented. 

“You have deterministic reasoning, and probabilistic reasoning.”

Harvey distinguished the two: deterministic follows a strict set of rules and ensures explainability and accountability. It can be explained that why x may be equal to y but greater than z because it uses rules and follows a strict set of code. 

Probabilistic is more problematic, and can be seen in generative AI models that lack explainability, can issue different answers to the same question, and hallucinate inaccuracies.

This is not to say that probabilistic – and by extension, genAI – does not have value or the potential for extreme transformation, but mitigating its faults, making it accountable, needs solid data plumbing.  

Data Governance Best Practice 

When it comes to actually implementing good data ‘plumbing’ and good data governance, firms of all sizes face challenges. 

Smaller firms typically struggle to get the funds for the skilled staff they need to clean up their data and implement it safely and securely into any AI. Pairing this with the growing AI skills gap and the risk of shadow AI, many SMEs may simply avoid implementing AI in any truly impactful manner, or risk data protection and security issues. 

Larger companies typically have the funds to hire outside help or hire top talent, so their challenges delve into the nitty gritty of data. 

Legacy data held on multiple inoperable platforms, duplicate data, augmented data can all create a quagmire that larger institutions can sink into. Data security, especially when it comes to breaches, is a top concern for larger firms embarking on their data journey, Harvey finds. 

Recommended reading

• Report: The Data Crisis Behind Business AI Failures
• Poor Data Governance Holding Back AI Adoption
• 95% of UK and US Firms Struggle With Data Issues, New Report Finds

But let’s venture a step further: what does good data governance mean when it comes to AI? 

“You have to prioritise accurate data. You have to have consistency in your model, and ensure that you have good guidance and standards around data relevance to the problem you are trying to solve,” Harvey explained some of the paramount features of proper data governance. 

Continuous monitoring, no matter the size of the organisation, is also vital when it comes to implementing AI. “Data quality isn’t a one time thing.”

None of these concepts are exactly new – there are organisations providing guidance and tools to organisations on data best practice, like the Enterprise Data Management Association and the Data Management Association. 

“At the moment it sort of feels like people are trying to solve a particular niche problem, rather than dealing with foundational issues,” Harvey said. 

And these foundational issues are integral for a reason. 

“Without those being in place, you don’t have any transparency, you don’t have the ability to say if your data comes from a trusted source,” Harvey said. 

Issues like inconsistent, erroneous data, and targeted attacks leading to data poisoning can all corrupt models, Harvey says. The stakes are high. 

“Without tools like data quality matching tools, cataloguing tools, lineage tools that sit on top of data, you don’t have the means for transparency.”

This can be devastating for building the trust essential in rolling out AI across essential services that impact people’s everyday lives. 

Working with government departments, Harvey is seeing a conservative approach to AI integration, with officials questioning processes to ensure transparency and explainability across AI outputs and decision-making. 

At Datactics, the firm uses an internal policy group to review any AI tools and how they might be deployed within the organisation. 

Other large organisations are held to high standards for data governance and data protection, such as large banks. 

But with the UK government’s push to become an AI superpower, it remains to be seen if all companies are taking this careful approach to AI integration. 

It brings questions and concerns to AI rollouts in areas like immigration, biometrics, as well as the prominent use of shadow AI seen across enterprises. 

To mitigate these issues, Harvey and Datactics urge for a deterministic model approach to AI adoption with robust data foundations to protect people from the downfalls of hastily created AI models. This may also usher in a more transformative era of AI, allowing not only the financial return on investment the market has been waiting for, but a return on innovation as well.