Data quality, integration and trust: the key to AI adoption

Challenges preventing organisations from realising true value from AI.

HotTopics C-Suite exchange in partnership with Salesforce

Existing organisational data challenges are creating roadblocks to the successful adoption of artificial intelligence (AI) in the enterprise.

A recent HotTopics C-Suite Exchange debate revealed the data challenges leaders face, whether they are working in financial services, public sector, or non-profit.

This debate found that organisations were investing in AI through proof of concept (POC) projects, but not yet seeing a clear return on investment (RoI) or the delivery of tangible business value. A number of attendees said this was often caused by the organisation not having a clear vision of how AI aligned with the business strategy, so the POC was not matching the desired business outcomes.

Overview: 

• AI is stuck in pilot phase
• Poor data quality, process and culture limit AI success
• Data skills shortage pushes organisations to AI COPs
• Why trust matters for AI adoption
• The Agentic Enterprise

AI is stuck in pilot phase

A nuance that the group identified early in the conversation was that large enterprise and commercial organisations are extending the timelines for AI to deliver business value, while the third sector is restricted by both its funding model and limited resources.

One attendee from the sector said a ten percent improvement in productivity is helpful, but anything larger required a wholesale change to the business processes of the charity - processes that the staff were either unwilling or unable to make.

This poses one of many challenges for AI to be successful. It requires a sustained investment in data, platforms and culture, but this takes time and money - both of which remain in relatively short supply given ongoing macroeconomic instability.

Poor data quality, process and culture limit AI success

AI may be a new tool and way of working, but it only works if the organisation’s underlying foundations are in good order.

The most striking, and common, problem is data quality. Poor data quality, as one attendee noted, can result in data bias which, in turn, can skew the effectiveness of these models, and ultimately lead to questioning the veracity and value that AI brings.

In regulated sectors, such as financial services, there are significant issues around AI governance and regulation. Much of the data a bank or insurance company holds cannot be utilised because it is sensitive intellectual property or customer data, and as such comes with strict security controls.

“We have to make sure that that is contained, or secured, before we can allow agentic AI process automation,” said one attendee from the sector.

In the same sector, but revealing a common problem across all industries, another attendee said their organisation didn’t have reliable data integration between systems such as the customer relationship management (CRM) and other back-end systems. Adding AI to such an environment would open the potential for incorrect outcomes and brand reputational damage.

On the existing data challenges hampering AI, one attendee said disparate and siloed data is going to cause large language models (LLMs) problems. But they also took a positive note that the demand for AI is another opportunity for data leaders to make sure the LLM understands the meaning of data and its context. In other words, by ensuring the AI models truly understand the context of the organisation and its place in market, it was far more likely that the output would be correct, and there would be far less likelihood of the old data science analogy; ‘garbage in, garbage out’.

While data quality, context and integration remain challenging for data leaders, there is also the issue of greenfield AI experimentation, and changes to traditional enterprise hierarchies, resulting in newfound bottlenecks around teams and decision-making.

“Applied to an existing organisation model, AI just made the bottlenecks appear faster,” said one executive.  

“Having AI can help one team deal with their backlog faster – but then you find Team B has stored their data in a different way, uses different terminology and different systems, which means it just creates another backlog.”

In response to this, another attendee attributed much of this bottleneck to the culture of the organisation.

“When you've got an organisation that's siloed, the data practices become siloed and the systems stop. You can speed any one of those up, but as soon as you get handoffs, you just make bottlenecks. That's been our key learning. It doesn't paper over those gaps; it actively shapes them.”

Data skills shortage pushes organisations to AI COPs

Like their cousins in IT, data teams have struggled with skills shortages in recent years, while demand from the business has exponentially increased.

The advent of AI requires new skills, and not just in data teams but also in the senior leadership of the organisation. A number of attendees here said they were looking to develop a community of practice (COP) around AI.

“Get those pockets of people that are doing things really well to share what they're doing for the benefit of the rest of the organisation,” said one attendee.

COPs are often a way organisations can tackle cultural issues towards a new way of working and have been used to great success in driving the adoption of cloud-based collaboration tools.

With AI, a COP may address one of the greatest concerns members of the organisation have - can AI be trusted? To create trust, one organisation changed the vernacular to that of AI being a way to increase the number of capabilities each of the team members had.

Why trust matters for AI adoption

Trust is not a new problem for data leaders to address, as one attendee noted: “We've always had this in the data analytics space. In the past, if I look at a dashboard and the numbers aren't quite matching up, or I know something's not to be true there, I don't necessarily trust it.

“If we don't trust what we're seeing, we're not going to go and use it, and actually this becomes more compounded with AI, because AI is generative. AI can look at our data and make its own understanding of what it is, and it can provide us answers that on the surface might look very credible – but, actually, might not necessarily be true.”

A single source of the truth has, for many organisations and their data leaders, been the holy grail.

Trust often fails because the data presented to the organisation lacks context, and Matthew Littlefield, Field CTO for analytics at Salesforce, says this knowledge layer is even more important to ground the LLM in what the business is, and what outcomes it is trying to achieve.

Littlefield explained that a knowledge graph helps the LLM understand data and the semantics around that data, as semantics provides the definition of what the data is.

“Consider it your map that you use to get around the place, and the knowledge graph is actually the journey on the map,” he said.

This will make AI trustworthy for organisations, he added.

“When an LLM doesn't have that context or understanding, that's when the generative part takes over, and it will guess what it thinks it means.”

The discussion highlighted a reality many organisations are now discovering: AI success is ultimately a challenge of data and contextualisation. Organisations that invest in data quality, integration, AI governance, trust and semantic context are far more likely to realise value from AI than those treating AI as a standalone technology initiative.

”The old ways in which we deploy new technology, or make any kind of change management within our organisations, all stand true. Yes, the technology is bigger, better, faster, stronger, but actually having the groundwork there is where you really start to benefit from AI, with a strong community of practice and strong skill sets,” summarised Littlefield.

The Agentic Enterprise

In partnership with Salesforce, HotTopics brings together forward-thinking technology leaders to explore what it truly means to build an agentic enterprise—one that balances innovation with control, speed with security, and automation with human ingenuity.