Setting Up the Right Data Foundation to Implement AI Models in Your Organization

By Dan Kronstal, Principal Architect, Solutions, Insight Enterprises

It shouldn’t come as a shock, but Insight is now officially leveraging OpenAI’s ChatGPT — well… an internal version of ChatGPT called Insight GPT to meet our own specific day-to-day needs as a Solutions Integrator.

Now, it’s a fairly big step, but it isn’t a shock, because Insight is far from alone. With The Harris Poll, Insight asked leaders at businesses with over 1,000 employees about generative Artificial Intelligence (AI), which is what ChatGPT leverages. The responses were compiled in our report, “Beyond Hypotheticals: Understanding the Real Possibilities of Generative AI.” As an illustration, 82% of leaders say their company has established or is in the process of developing internal generative AI policies.

While ChatGPT is gaining traction all on its own, it’s one of several AI models that companies from various sectors can use and are using as a foundation to build their own. They’re each presumably leveraging their own data to serve their own purposes, though. Based on our experiences at Insight, that would be the right way to go about it, anyway.

AI models are like librarians?

Our friends at Microsoft at a recent Executive Briefing Centre (EBC) event shared another analogy to AI models that I think fits perfectly. Think of large language models (so, ChatGPT) as librarians. Like librarians, they know about books in the library and can answer questions about what’s in those books. This is where it gets tricky, though.

Like a librarian, AI can have access to excellent references in the same vein as encyclopaedias, biographies, etc. Also like a librarian, AI can have access to great works of fiction too, though. Ask the AI a general-knowledge question and it won’t necessarily know the difference between the two, generating an answer.

Now, that doesn’t mean the info is faulty or that the librarian is at fault themselves, just that the librarian lacks the right context. The hypothetical mistake falls on the shoulders of those putting data into the model (or at least on the shoulders of those unreasonably expecting more out of the model than the data allows). If only non-fiction works are included, there won’t be that same inaccuracy issue. Incorrect information just means that the info made accessible to the model is wrong. The model can still be right.

Road-deficiency detection as a case study

Take a real-life case study as another example. A county in Ontario worked with Insight to leverage AI to automatically identify and log road deficiencies, like potholes. The county mounted cameras on their patroller trucks. Meanwhile, an in-vehicle AI used an object-detection model developed using Microsoft Azure Cognitive Services to process what was getting filmed as patrollers went about their normal everyday routines driving from location to location.

Insight built the model using public image sources, but we’re consistently improving it by using the client’s own footage taken in the county itself. Otherwise, the model would only have roads from other parts of the world as a comparison. They look different in Canada, a fact that could conceivably lead to misidentifications… were the AI not getting fed more and more images of the roads it’s now regularly encountering.

The next frontier in business productivity

It all speaks to the challenges companies now face regularly, as AI — and generative AI specifically — becomes more and more integrated in best business practices and processes. According to our data (again compiled by The Harris Poll), the vast majority of respondents see the technology boosting productivity (72%) and customer engagement (66%), while 90% say it’s poised to enhance a wide range of roles, generally speaking. So, it’s at least seen as a challenge well worth overcoming.

The question becomes “how?” The field is admittedly fairly new. So, speaking to the aforementioned employees who would be inputting data, companies must determine which roles to adopt to properly classify and curate the data, among other tasks.

However, before companies get that far, they first have to understand the data they have. Very few organizations have the required level of visibility into their existing data to leverage it. Then comes the data-classification and governance stage, so establishing what data is appropriate for inclusion in the model — and how to go about including that data. In fact, Insight is working with clients encountering those same obstacles right now, in the Insight Azure OpenAI Immersion Workshop, for example.

The workshop features several components across a series of five sessions. However, for the purposes of this overview, it makes more sense to describe the engagement as consisting of three critical parts as the client begins to explore the adoption of AI models to boost productivity:

1) Customer education on the cutting-edge capability of OpenAI, how its models work and how they are applied to real-world problems.

2) Specific outcomes and use cases regarding your business goals in particular, i.e., how these AI models can be applied in real life in your organization.

3) What do you need as prerequisites to take your first/next steps towards adopting and ultimately taking advantage of AI in your organization.

Your journey to adopt AI models

Ultimately, every organization (and every organization’s journey) is different. Those just-mentioned prerequisites include first assessing your proper cloud-adoption posture. Are you in the cloud today? Do you have all your files sitting internally in network-attached storage on the premises?

In the case of the latter, you’ll probably need some cloud-adoption fundamentals under your belt, before entertaining the data side of the story, which would be Step 2: first ensuring your data estate is in a cloud-ready location that is accessible to these AI models and then determining the maturity level of that data. Do you have a curated area where data can be ingested into the model? [SR1]

You’re in the 1% of organizations if you do. If not, Insight can help you get started by developing a Minimum Viable Product (MVP)… along with any of our other available services in this domain.

What the right data foundation sets you up for in terms of AI

It’s hard to minimize AI’s business potential here, as the possibilities are endless, especially as we’ve just scratched the surface of what the technology can do. There are near-out-of-the-box solutions out there, where you simply need to make a few connections to your existing data repositories, enabling users (or customers) to simply enter a request of the AI model into a prompt… similar to with ChatGPT.

One possible application we’ve been circulating internally is of a hypothetical bike manufacturer. The user, someone in the marketing department, can for just one example ask the large language model to write an email to send to advanced riders, advertising available bikes that haven’t sold in the last month.

Seeing as that last sentence was a bit of a mouthful, there’s clearly a lot would go into that one request. Here’s a brief summary:

· The user chats with the application, making their request.

· The request goes through the model and makes an enquiry of the different data sources on the back end, like Azure SQL/Cognitive Search.

· The data sources pulls the necessary data out of Azure Storage.

That data includes product and review tables from your eCommerce database, along with marketing data and photos from your stock repository. The model then finds a relevant email template in the connected marketing platform and crafts the content.

The difference between this solution and ChatGPT is clear. A model like ChatGPT is the foundation on which a business can build their own using domain-specific data, data which ChatGPT wouldn’t be able to leverage. Only that organization would, making this model tailormade to do everything it and only it needs to — and that’s true of companies all over now leveraging AI models. Potentially even yours.