ChatGPT seems to suddenly make AI mainstream for businesses, but in fact, almost every organization has likely been using software with AI incorporated for years—or possibly even decades. Thanks to successive evolutions like machine learning, big data applications, and the rise of LLMs since the launch of the Transformer architecture, more and more advanced AI applications have become available.
Since the introduction of ChatGPT in November 2022, almost every company seems to be exploring relevant possibilities. With LLMs like these, an incredible range of capabilities is possible. Companies are discovering new applications daily, whether through integrations with other applications or the ability to securely interact with their own data.
Because LLMs like ChatGPT can do so much—and often do it very well—the technical solutions are not the problem. The real challenge lies in asking the right questions. What our customers struggle with the most is identifying the pain points that we can solve with tools like ChatGPT. In addition to identifying the challenges you want to address, it's important to consider factors such as how complex these issues are, how much (high-quality) data you have available, how much time and money you want to invest, and whether you should opt for an out-of-the-box, hybrid, or fully custom solution.
Why make things complicated when it can be simple, fast, and proven effective? The Microsoft Azure stack offers numerous AI solutions that might be suitable for you—and for 101 other companies facing a similar problem. Consider, for example:
The best part: all of this can be done out-of-the-box. It's almost plug-and-play. Some customization may be required, such as integrating the output into a local application.
The benefits of such out-of-the-box implementations are evident: it's relatively easy to integrate, requiring minimal investment and expertise. Last but not least: hardly any data is required to train the model; you use pre-trained models. However, there is a risk: your data must match what the model was trained on. Nevertheless, the more frequently the use case occurs, the more likely this is.
Our advice: If you have a common problem that isn't too complex and you want to create value quickly, first check if there's an out-of-the-box AI tool available.
On the other end of the spectrum is developing everything custom. We only recommend this if you have a unique AI-based use case that hardly any other companies are grappling with. Such a custom solution requires significant investment, a lot of (preferably high-quality) data, and considerable time for model training and validation. However, it offers two significant advantages: the output is entirely tailored to your organization, and your recurring costs are low since everything belongs to you.
The custom aspect can encompass various elements, including data integrations with other applications, architecture, feature engineering, model development, training and validation, frontend development, and security. Cegeka can handle all of this, resulting in a solution that is as reliable, secure, compliant, and effective as it can be.
Examples of such custom solutions include:
For such a chatbot, we use the organization's own data, which may come from sources like Sharepoint or the HR system. It utilizes ChatGPT (or a custom-trained LLM) via API calls. This is done in a security and privacy-conscious manner because all data remains within the organization's Azure environment. It's worth noting: an HR chatbot can also be relatively out-of-the-box, but in cases like very technical companies, such a solution may not perform well.
With such a system, you can integrate ChatGPT into your ticketing system, like TOPdesk, via API calls. Customization can be achieved by training your own model on historical ticket data on top of ChatGPT. Additionally, you can develop scripts to clean up ticket data and establish business rules that determine how tickets are routed and handled. Thanks to a feedback system, employees can evaluate the quality of ChatGPT's responses.
A hybrid AI solution can be a kind of middle ground. You start with an out-of-the-box AI solution as the foundation. However, because it doesn't align perfectly with your use case, you apply some custom development to it. This way, you leverage the speed and efficiency of an out-of-the-box solution and don't require a lot of (high-quality) proprietary data. The downside is that you incur both initial development costs and ongoing expenses.
Consider solutions like "legal buddy" in this context. With this, legal professionals can search for information in old cases, using cognitive search technology to retrieve relevant documents from the company's database. The information is then fed as input to an out-of-the-box Large Language Model (LLM), which can extract and present the key information. Through few-shot learning, the model quickly improves with only minimal additional data.
Matthias Verlinde, a data scientist at Cegeka, states, "During brainstorming sessions, we often distill several use cases that may be interesting for a client, either because they address specific pain points or accelerate processes. These use cases can be entirely out-of-the-box, entirely custom, or anything in between. Customers often hesitate when they hear about the investments involved in more custom work, in part because you often need people like data scientists, data engineers, and architects. The tricky part about AI solutions is that you can never guarantee success in advance."
Therefore, Cegeka works with a proof of value via a fast track approach. Verlinde explains, "This means we first do an iteration with an out-of-the-box tool or a fairly standard implementation of, for example, a custom computer vision or NLP model. This gives the customer a good sense of the potential performance of the solution. Then we ask: is this sufficient? Or does it require more custom development to better align with the actual business needs?"
This article was created following the Cegeka webinar "Integration of ChatGPT into Business Processes". Want to learn more? You can watch the webinar here.