Dun & Bradstreet - accurate data must be the basis for any serious enterprise use of generative AI

AI - especially generative AI, such as Large Language Model-based predictive text engines such as ChatGPT - is something organizations believe has huge potential in the enterprise. However, concerns about hallucinations, as highlighted by Salesforce CEO Marc Benioff last week, and shallow knowledge bases, means a lot of CIOs are taking a ‘wait and see’ approach until things shake out.

A company right in the middle of this debate is Dun & Bradstreet, whose main business is providing accurate corporate information for American businesses.

After all, its DUNS (Data Universal Numbering System), which assigns a unique numeric identifier to a single business entity, is almost universally used in credit search.

And like many organizations, Dun & Bradstreet initially wasn’t quite sure about the right move forward when it came to Gen AI. 

In the words of its Chief Data and Analytics Officer, Gary Kotovets:

Data is at the core of what we do - we sell data and we sell insights to drive efficiencies, because data is the fuel e-commerce depends on. But more and more, we’re starting to focus not so much on data as it is, but on how we can improve our clients' existing workflows and uses of their data.

Prior to the AI hype, that meant a Dun & Bradstreet-wide emphasis on accuracy of business data, he points out.

He adds:

“Dun & Bradstreet data is always clean, validated, structured, and as we like to say ‘DUNS-ified.’

“In the context of AI, that means that if you develop a model and deploy it, if it points to a dat set that's not validated, it will give you the wrong answer - and if it's pointing to a data set that's not uniquely identified, it could potentially give you the answer about the wrong entity.

“And I think we can all agree that the AI debate is pretty consistent - that your AI is only ever going to be as good as your data.

But taking baby steps with Gen AI, while understandable, risks missing opportunity and immediate cost-savings by using text AI to speed up some business processes.

The organization decided to take positive action, do the work needed to turn this around and start to see what practical benefits it could get from AI in the short-term.

However, any such work had to avoid issues such as allowing users to train models without enough guidance and introduce errors, or allow personal or sensitive information, like PII (personal identifiable information) to leak.

Limited access 

Investigation of how to do generative AI the right way has been led by Kotovets and his small (three-person), but highly focused, data and analytics team.

He says:

I joined the organization about three and a half years ago, and my responsibility is to work on improving all of our foundational data, as well as creating new insights and analytics services we could commercialize.

The first step was to establish a compliance framework, he says, to minimize such risk as far as possible. 

Next up was to build a technological risk framework, and then put a chosen internal Dun & Bradstreet generative AI team through what he says was “pretty robust training” based on the two frameworks.

As a result, only staff with such training processes were granted access to generative AI tools - those he and his team believed would conform to the legal and technological parameters that had been set.

He says:

After we qualified and validated the output of the models, we had a pretty robust quality assurance process that was making sure there were no hallucinations in the output.

After an intensive testing and validation process, Kotovets and his team decided to apply these new Gen AI Dun & Bradstreet builder tools.

Positive results, he says, soon started to be derived in two business workflows: procurement and sales and marketing.

Procurement is a heavy user of data, he points out.

A lot of time must be spent on researching and identifying risks within the supply chain.

Procurement professionals - the CPO, the Chief Procurement Officer - therefore spend 70%-80% of their time in managing existing suppliers, understanding costs, assessing supply chain disruptions, and checking out for supplier failure, he states. 

He adds:

All of these tasks need to be manually researched, identified and automated - but smart use of generative AI can help surface various risks, potential disruption, and identify alternative suppliers to help the CPO make decisions faster and more dynamically.

Another aspect of the CPO’s job is the 20%-30% time procurement practitioners spend answering questions from internal stakeholders - e.g., on the status of an order, or information about a supplier.

Here, Kotovets says, a self-service Procurement support generative AI-fueled chatbot can deal with many of these first-level questions, freeing up the Procurement team to focus on more critical problems.

Dun & Bradstreet then looked at helping potential sales and marketing applications of a refined business data and AI combination.

Here, he adds, creating a target list is a regular activity for a sales or marketing professional, which can sometimes end with a database full of errors and which, he claims, can take up to 10% of a professional's time. 

Against the right dataset, a Gen AI model can help create such lists quickly. 

He says:

Just by asking simple questions, it can produce a useable list within seconds.

Connecting AI back to your data

Now, the firm is looking to share its learnings and best practice in a more structured way than these initial proofs of concept.

This is in the form of a new Dun & Bradstreet commercial service - D&B.AI Labs.

This, he claims, will be a transformative hub for organizations to pursue co-development of AI-powered solutions.

Customers will be able to do that by leveraging the power of both Dun & Bradstreet’s proprietary data and analytics and all his team’s work on Gen AI. 

Kotovets says:

We’ve got data scientists, data engineers and solution specialists with extensive innovation experience as well as deep expertise in not just AI but also advanced business analytics. 

The team will work side-by-side with customers to formulate solutions in real-time, build prototypes, and rapidly deploy solutions to help clients realize the power of our data and analytics.

Which brings us back to the role of data - as Kotovets stresses the importance of good data in any attempt to make generative AI a dependable enterprise tool.

He concludes:

You need to be sure that you can always connect back to your data to understand where any answer came from.

In other words, you not only need to see the answer - you are also going to want to know the underlying document or publicly available source or any other private source that produced the answer your AI’s giving you.