Build vs. Buy vs. Bespoke: Navigating AI Solutions

AI is here to stay. How should Medical Affairs teams think about integrating this technology into their tech stack?

As interest and adoption of AI solutions continue to surge across all industries, pharmaceutical companies are once again facing a familiar question: build or buy? When it comes to leveraging this technology for medical affairs in particular, the accuracy of AI output is crucial. Medical language is highly complex, with nuances that can significantly alter meaning; for example, asking AI to interpret whether patient-reported information is a symptom or a side effect requires a deep understanding of context and clinical judgment. Medical affairs teams make data-driven decisions that can impact patient lives, and having unreliable or incorrect AI output can have serious – and potentially fatal – consequences.

Build: a Custom – but Complicated – AI Strategy
Teams may feel as though building a custom AI system from the ground up is the best approach. This strategy gives companies control over their data and system security. It also provides full authority over system development, enabling companies to create solutions and add new features tailored precisely to their needs. Additionally, custom-built systems can seamlessly integrate with existing infrastructure, potentially providing a significant competitive advantage.

 Building a custom system is extremely expensive, requiring significant upfront funding and ongoing maintenance costs. 

But despite these few positives, choosing to build a custom AI system has many drawbacks. For one, building a custom system is extremely expensive, requiring significant upfront funding and ongoing maintenance costs. AI is evolving at such a rapid rate that just keeping up with the latest progress would be challenging - likely necessitating a dedicated team and an even bigger investment compared to working with an external AI vendor.

Not to mention the challenge of staffing resources to build a custom AI solution. Given the AI field is still relatively new - yet technically advanced, sourcing top AI talent who also understand the life sciences would be difficult. Such obstacles are no secret, and often make it hard for internal teams to secure buy-in for these large, uncertain custom technology initiatives, delaying progress even further.

Regulatory compliance adds another layer of complexity. While building an in-house AI solution might seem like the best way to adhere to regulations, it actually creates additional headaches. Companies must consistently monitor changing regulations and ensure their AI solution meets any new requirements.

Buy: Off-the-shelf AI is Easy, But Not Necessarily Better
Some organizations have taken a different approach, choosing to forgo the challenges of custom-built solutions in favor of off-the-shelf AI products. As solutions like ChatGPT, Microsoft Copilot, and Google Gemini have skyrocketed in popularity over the last two years, these pre-built AI tools have become commonplace in many organizations. It's easy to see why they've gained traction: they are fast to implement, easy to use, and relatively inexpensive compared to building in-house. But for Medical Affairs teams, the promise of generic AI tools has underdelivered. So much so that some big pharma companies are clawing back decisions to invest in generic AI products.

 The real challenge is finding AI products that can serve hyper-specific and complex life sciences and medical affairs use cases. 

While this trend to move away from AI may seem surprising, the real challenge is finding AI products that can serve hyper-specific and complex life sciences and medical affairs use cases. Generic solutions are built and trained for broad application across many industries, activities, and use cases – from business writing to recipe development. It’s no surprise that without a deep understanding of life sciences terminology and concepts, many generic generative AI products struggle to produce comprehensive, accurate answers.

Here’s a common example to illustrate this: in the context of many medical decisions, the word “negative” is actually a positive signal. If a patient undergoes a breast cancer screening and the results show negative for breast cancer, this is in fact a positive outcome. When AI models are trained primarily on definitions of language outside medical contexts, it would assume that negative is bad and positive is good – in this case, misinterpreting crucial diagnosis information.

Training AI models on life sciences terminology is no easy feat. Doing so requires access to comprehensive, structured training data from a vast range of specific sources – published literature, clinical trial outcomes, HCP feedback, and much more. Gathering and structuring this information is just the first step. AI tools must also use advanced, domain-specific natural language processing (NLP) ontologies to train models in the context of scientific domain knowledge. Without applying such a rigorous approach to model training for this specialized industry, it's likely that generic generative AI products will hallucinate (generate false information that’s presented as fact) – which can provide incorrect information that potentially hurts patient outcomes.

Not to mention, many generic AI products fail to meet the stringent data security and regulatory standards required by pharmaceutical companies. This presents an unfortunate trade-off: while life sciences companies might be able to train generic models with sufficient high-quality data, the lack of robust security in these solutions means doing so is risky. As industry regulations around AI continue to evolve, companies that use generic models must take on the burden of continuously monitoring AI solutions to ensure they’re complying with the latest rules.

Bespoke: The Best of Building and Buying AI
At the crossroads of build or buy, medical affairs teams are left with two less-than-perfect options. But the surge in generic AI products has also sparked a surge in AI solutions specifically built for life sciences – providing a third pathway for companies to invest in AI: the bespoke approach.

 Bespoke AI solutions combine the best of building in-house with the benefits of buying generic – bringing the customization and industry-specific architecture of built solutions with the speed and flexibility of off-the-shelf, easy-to-buy tools. 

Bespoke AI solutions combine the best of building in-house with the benefits of buying generic - bringing the customization and industry-specific architecture of built solutions with the speed and flexibility of off-the-shelf, easy-to-buy tools. Unlike generic AI tools, industry-specific AI solutions are trained on life sciences concepts from the start, employing advanced medically tuned ontologies and data labeling to properly categorize information and connect data in ways that's meaningful for medical affairs.  

These bespoke solutions also consider the vast range of data that’s needed to train the AI solution with a comprehensive understanding of life sciences knowledge. This includes sources ranging from published literature to a company’s product information to clinical trial data to information locked up in unstructured formats like advisory board presentations or hand-written notes. For companies with established medical strategies and an extensive amount of internal data, this wealth of information becomes a valuable asset, further training these AI models on the company's unique business context and objectives.

Bespoke AI solutions prioritize ingesting and structuring data from these numerous sources, as well as providing secure, dedicated environments that maintain appropriate levels of data privacy. When it comes to ensuring compliance and upholding other life sciences and medical affairs regulations, the vendor takes on full responsibility. This saves organizations the time and burden of navigating these requirements on their own.

While bespoke AI solutions aren’t custom-built for each team, they are validated with companies within the life sciences, and therefore provide a best-in-class, industry-proven solution. Additionally, vendors building in this space must hire top engineering and product talent who deeply understand both AI and life sciences. Given the highly competitive and saturated AI market, it’s the only way to have a chance of delivering a successful product. Therefore, any companies using bespoke AI solutions benefit from the collective expertise of the team – who, in addition to their technological skills – can also help guide companies in areas like AI strategy and rollout, change management, and driving adoption. 

Considerations When Choosing a Bespoke AI Solution
At the crossroads of build or buy, despite the clear advantages of a bespoke solution, not all life sciences AI products offer the same quality. When assessing potential vendors, it is important to keep a few considerations in mind. First, teams should find a vendor that will work with them to implement a flexible system that fits the company’s unique needs, instead of force-fitting data and workflows into their structure. To assess this, ask the vendor how they developed their features and the type of customer service they provide. Companies that offer their customers a direct line for feedback, and focus on developing new features in partnership with customers, typically have a positive track record of meeting customer expectations.

 Teams should find a vendor that will work with them to implement a flexible system that fits the company’s unique needs, instead of force-fitting data and workflows into their structure. 

Lastly, while companies approach security and ownership differently, those purpose built for life sciences know the crucial importance of confidentiality when dealing with sensitive healthcare information and proprietary data. Vendors must disclose how they train their AI models and engines, and provide customers a private, dedicated tenant to ensure sensitive data is protected.

To learn more about navigating the decisions to build, buy, or use bespoke AI solutions,  .