Never Miss a Beat: Get a snapshot of the issues affecting the IT industry straight to your inbox.
What Makes a Perfect AI Pilot?What Makes a Perfect AI Pilot?
Successful pilots are critical to creating AI use cases. Here is the recipe to achieve it.
January 13, 2025
4 Min Read
Image by Gerd Altmann via Pixabay
Sponsored by Lenovo
How do you get your AI use case up and running? A pilot is the essential starting point. However, if you’re like most organizations, you may run up against barriers at this early stage. In research conducted by Lenovo earlier this year, 60% of companies told us that they struggle to get stakeholder support for pilot programs, while the majority reported substantial financial constraints at both pilot (88%) and deployment stage (83%). How can you overcome these hurdles and put a successful pilot together?
Like good cooking, it's about getting the ingredients, people, and timing right. Let’s start with ingredients.
Bring Together the Data and Tools
Data is the key component of any AI application. What’s imperative is to first find the data that’s needed to drive your use case. While each one is different, what GenAI projects have in common is that the data is likely to be unstructured – for example, a mix of documents, PDFs, and web content – rather than relational data. To be maximally effective, AI solutions typically need as much of this data as possible.
This brings us to the second ingredient – the tools to run and manage the pilot. These break down into three broad areas:
Data evaluation and screening. First up, is there enough data to work with? It’s important to know what’s there upfront. Tools are available now that can evaluate documents and ensure they have enough quality data to support a GenAI use case.
Data cleansing. Documents need to be screened for low-quality documents, duplicate information or personally identifiable information that should be masked, for example. So, you’ll need a tool that’s capable of doing this across unstructured data, which is a different challenge to cleaning up relational databases.
GenAI operations. If an AI pilot is not properly proven, stakeholders will rapidly lose trust. Solutions have been known to “hallucinate,” give out incomplete or incorrect answers, or even make missteps like recommending a competitor’s products. GenAI operations put in place practices that check for and correct issues such as model drift and bias, guarantee accuracy, and generally make sure the solution is ready for use in a production environment. This might include, for example, using a set of baseline questions to benchmark the AI’s performance.
Assemble the Right People
Next, the people. The first and most obvious are the individuals who own the use case that the GenAI pilot is designed for. Engaging with them – and understanding their definition of success – is a critical first step.
Then, there are a range of people that our experience shows are pivotal to making any AI pilot a success:
Data steward. This is the subject-matter expert who owns and understands the data needed to drive the solution.
Vendor data expert. They work with the data steward to make sure data is cleansed and ingested into the GenAI system, making it ready to use.
IT team. The pilot needs people who understand the systems that are being integrated with the vendor’s AI solution and have ownership of them from a technology point of view.
Security team member. The AI pilot also needs a security expert on hand to ensure that the solution meets your security requirements. This individual will also be able to define upfront the levels of integration required, as well as the solution’s balance between the cloud and your secure data center.
Internal domain experts. While a vendor can of course help with testing an AI pilot, it’s essential to bring in subject-matter experts to validate whether it is providing the right responses. In an insurance use case, for example, these would be the people who know whether the AI was recommending the correct policies.
Get the Timing Right
Finally, timing. GenAI pilots typically take between two and six months, depending on the extent of the use case. However, like any good recipe, they need careful monitoring throughout the process, in this case to check for “model drift” and to correct for any changes in data or the model itself.
And to save time in the future, what’s important is to use one Gen AI platform for every use case, to avoid having to start from scratch every time.
Proven Solutions
Lenovo, in collaboration with NVIDIA, offers tried-and-tested tools and frameworks for building AI pilots to help you get GenAI solutions into production. Lenovo AI Advantage with NVIDIA, a full-stack solution for building and deploying AI capabilities, combines Lenovo’s services and infrastructure with NVIDIA accelerated computing and NVIDIA AI Enterprise software. To build real-world proofs of concept, Lenovo AI Fast Start delivers live solutions to demonstrate generative AI deployment and showcase business, operational, and technology results. Businesses can accelerate and quickly scale AI using full-stack NVIDIA-based technologies through Lenovo AI Fast Start for NVIDIA AI Enterprise, which also includes NVIDIA NIM microservices, NVIDIA NeMo, and NVIDIA Blueprints.
To learn more about GenAI pilots, and how Lenovo and NVIDIA technologies can help, contact [email protected].
You May Also Like
More Insights
