Using AI Archetypes in Designing Digital Transformations in Life Sciences and Healthcare

Exactly how will artificial intelligence (AI) help life sciences and healthcare leaders develop more valuable companies?

The latest AI tools are rapidly growing in their capabilities, and the speed at which vendors are integrating AI into their products – clinical and lab systems, office productivity, graphical content development, meeting scribes, back office, and much more – is unprecedented.  But to many industry leaders, AI can still seem like a collection of questionable parlor tricks – creative slights of hand with data that may be interesting but also disconnected from corporate strategy, patient impact, and investor interest.

Real-world experience can certainly fuel skepticism as well: how many times have executives been told software will totally revolutionize drug development or healthcare delivery?  A realistic understanding of the barriers (i.e., emerging technology maturity, data curation, bias mitigation, skills availability, cost of talent, change management, IP protection) that exist between pretty slides and practical solutions can also contribute to a lack of confidence.  As an experiment, ask ChatGPT to generate a graphical image of the drug development process; it’s apparently a lot easier to draw lovable animals than spell “manufacturing” (see Figure 1).

Figure 1.  ChatGPT responses to the following prompts: “Generate a photorealistic diagram of the drug development process. Make sure all words are spelled correctly.” and “Draw a photorealistic unicorn that enjoys science”. (ChatGPT 4o, 2024).

Product maturity issues aside, for many leaders, it’s not a question of trust or optimism, it’s a question of focus.  Exactly what processes will benefit from AI-driven digital transformation, and how will that work?

I recently published a CREO white paper on the major life sciences AI trends.  In looking across use cases shaping this industry landscape, patterns in the intended uses of AI are increasingly evident.  These patterns – which I think of as “archetypes” – reflect the role(s) AI can serve in enterprise workflows.

AI archetypes – such as “Conversationalist” – offer a non-technical perspective on how AI can function collaboratively in workflow alongside humans.  And though the archetypes I describe below include the current market fixation on generative AI, they actually function independently from any underlying methodological or computational approach.  This agnostic stance to methods offers flexibility for the use of predictive analytics, simulations, prescriptive analytics, and other data-driven logic to support digital transformations.

Seven AI Archetypes

So, what are the recurring archetypes?  I see at least seven in the market today:

1. The Conversationalist. Software agents powered by natural language models can serve as a human-friendly front end to business processes, backend systems, and other AI models.  For example, a pharmaceutical research company might use AI agents to answer frequently asked questions related to their research studies or drugs.

2. The Advisor. These models are tuned to very specific problem spaces and can offer decision support assistance.  Advisors today are showing up in areas such as in-application collaborators (i.e., guiding users with completing basic tasks, presumably better than Clippy) and embedded functionality within enterprise applications (i.e., helping users detect and assess emerging issues in workflow execution and data).

3. The Expert. AI models can aggregate, organize, and synthesize extremely large volumes of complex, domain-specific information on demand.  Experts are often positioned alongside a Conversationalist or Advisor to surface expertise, though that is not the only deployment model.  And though we have already seen the emergence of models with tailored expertise in areas such summarizing scientific literature, many non-generative AI models offer expert insights as well.

4. The Creator. With input from a user, models can generate de novo content – documents, software code, models, audio, video, etc. – that would otherwise require a human resource to develop.  Many authors are already using AI to help draft reports, industry articles, soundtracks, voiceovers, and full-motion online videos.

5. The Modeler. AI models can simulate the structure and behavior of real-world phenomenon, either to help humans understand the problem space or to help develop new solutions (often paired with a Creator).  Industry models such as AlphaFold are a great example of this approach in the biotechnology sector; others include simulating patient traffic patterns in hospitals, developing synthetic patient cohorts, modeling financial performance, predicting disease progression, and other forms of advanced analytics.

6. The Orchestrator. Models can serve as intelligent machines (both physical and software), performing tasks that would otherwise require a human set of hands to operate.  Applications such as autonomous driving, military drone operations, robotic process automation, and other tasks fall into this category.

7. The Monitor. Using information from software systems, cameras, and sensors, models can continuously analyze real-time data feeds to detect signals, defects, or other notable events.  Uses of Monitors include manufacturing quality, cybersecurity, performance measurement, and regulatory compliance.

There are probably other AI archetypes that might come to mind as well.  Of course, these archetypes are not mutually exclusive; many popular AI tools today offer a creative blend of these capabilities.

Putting Archetypes into Practice

AI archetypes can be useful when brainstorming process re-engineering concepts within digital transformation programs.  Though AI can be overlayed on existing business processes without re-engineering, it is often unadvisable.  Many existing processes are inefficient and error-prone; adding automation (or training an AI model to be inefficient) adds cost, complexity, and risk without fully exploring opportunities for optimization.  From a digital transformation perspective, AI presents an opportunity to reconsider the most effective ways of getting a task done, and archetypes offer a simple way of thinking about how AI can contribute to streamlined processes.

For example, if I wanted to develop a new customer service experience where clients had easier, immediate access to product information, I might envision an interface that starts with a Conversationalist that knows how to talk to customers in my particular industry.  I might want that Conversationalist to connect to a deeper Expert specifically trained on our products – a model also used by internal teams at my company so there is a single source of truth.  I might also want the Conversationalist to know how to take guidance from an Advisor that specializes in troubleshooting.

Figure 2. A three-archetype solution example

Note that in this simple example, I could have engineered the solution like a “search” process, as many existing software interfaces do today.  But using this archetype approach, I’m able to explore potentially offer clients a higher service level while lowering more expensive call center activity resulting from failed searches.  I’m also able to potentially improve service quality because customers have access to the same deep well of product insights and troubleshooting expertise as internal company teams.

Picking a more complex example, perhaps I have a workflow where I need a more controlled, regulatory-compliant solution.  In manufacturing, for example, I might start with an Orchestrator that is charged with overseeing a controlled quality management process.  The Orchestrator receives regular feedback from a Monitor that is aggregating and analyzing manufacturing line data in real time and knows what should happen when quality issues emerge.  It also knows how to tap into an Advisor to determine what corrective actions should be generated based on emerging quality issues.  That Advisor leverages an Expert in the regulations as well as a regulated document Creator to automate some of the steps required.

Figure 3. A five-archetype solution example

Of course, it’s possible to perform all of these functions within one AI model, and in some cases that may make the most sense.  But it’s worth considering some of the advantages for treating each archetype as a modular asset:

1. Tuning.  By constraining the scope of each model to a more discrete problem space, we’re able to more finely tune the behavior of the model.  Though we can always constrain model inputs by controlling our input parameters, modularized AI models allow better control of model outputs as well.

2. Variability.  Tuning also helps us reduce variability in performance.  When models are engineered to support more generalized tasks, the increased breadth of responsibility can undermine the precision, accuracy, and repeatability of performance.  In highly regulated areas of the business, that variability may be undesirable.

3. Complexity.  From a solution development perspective, it is often useful to decompose complex problems into smaller chunks of work.  When development teams can more easily characterize use cases, not only does development velocity increase, but code quality and performance often increase as well.

4. Manageability.  Different aspects of our solution may have different scalability and security requirements.  For example, our monitor needs to efficiently process high-volume data feeds, but other elements of the solution rely on more static and carefully curated data repositories.

5. Reusability.  Ideally, we want to repurpose these investments in the future.  Our Expert could be used to answer internal staff questions about changing regulations.  Our Creator could be used to develop documentation in other regulated areas of the business.  By looking at AI capabilities as domain-specific archetypes, we can accelerate future improvement opportunities that need comparable capabilities and expertise.

If any of this approach seems familiar, software architects will immediately recognize this design ethos: service-oriented architecture.  This proven design strategy has been powering non-AI software for many years.  AI solutions can easily inherit these software engineering best practices even as AI changes the way we think about the future of software itself.

The Road Ahead

For the foreseeable future, digital transformations will feature creative combinations of humans leveraging both traditional software and machine intelligence.  In most settings, we won’t be fully relinquishing the steering wheels of our scientific or business processes – the operating model will look more like human-machine teaming.

With the software and data capabilities that exist today, we can reasonably ask how AI can contribute to accelerating the important work of improving patient lives.  If someone had asked me to draw a picture of a science-loving unicorn, it would have never even occurred to me to include those safety goggles.  But my own drug development diagrams are way better.  Developing AI-driven solutions requires "pressure testing" the performance of various AI approaches, but archetypes offer us a mechanism for exploring what role(s) AI could potentially serve.

What do you think?  Are AI archetypes a useful way of thinking about AI solution designs?  What other archetypes can you imagine?