Generative AI agents: The next frontier in enterprise transformation

Generative AI has the potential to revolutionize enterprise processes, but most current applications are far from realizing their full potential. While companies have made significant strides, many still confine themselves to basic, single-step solutions for content generation or question-and-answer style chatbots. However, the next frontier, where generative AI surpasses these initial solutions, is within reach and several good use cases are feasible today.

Imagine a future where AI can handle complex, open-ended questions requiring deep reasoning and intelligently source information from a diverse array of databases, including both web-based and internal systems? And what if, when the initial output falls short, AI could ask another model to refine and improve the response, ensuring consistently high-quality results? Furthermore, could AI code or leverage predictive models to forecast outcomes, what if it could make decisions, or even intelligently pause and wait for human input when necessary?

To achieve these ambitions, a new vision is starting to take shape: a world of generative agents that, much like humans, specialize in specific tasks, and collaborate to tackle complex challenges. Generative agents have the potential to revolutionize enterprise applications and fundamentally transform how work is accomplished.

Generative agents are specialized instances of large language models (LLMs). Generative agents are fine-tuned, or prompt-tuned using instructions, to perform specific tasks, and can be equipped with various capabilities to enhance decision-making and enable action-taking (see figure). 

Each agent is equipped with instructions that allow it to perform specific tasks, including:

• Role or persona: The LLM is instructed to assume a particular role or persona, sometimes with guidelines to plan in specific ways, recruit other expert agents as needed, and reflect on and improve its own outputs.
• Context: These agents operate within a specific context relevant to their assigned tasks, which might be derived from the user’s immediate environment or other sources within a workflow.
• Memory: Generative agents retain memory of previous interactions and actions, allowing them to maintain continuity and coherence in complex problem-solving tasks.
• Knowledge: Agents can access structured data or vector databases containing embedded information, enabling them to draw upon a wide base of knowledge when generating responses or making decisions.
• Tool integrations: Agents can be equipped with integrations that allow them to use various tools and take actions using those tools. This might include accessing a calendar, updating databases like enterprise resource management or client relationship management systems, performing web searches, using predictive models or interacting with other agents.

The combination of these features creates a generative agent capable of acting as a powerful and versatile problem-solver.

Consider a Master Data Management (MDM) agent tasked with resolving duplicate entries in a database. Imagine you have a global list of clinical trial locations compiled from multiple sources. This agent could autonomously identify and resolve data duplication issues by grouping similar entries and using external sources, such as the internet or other sources of information, to clarify ambiguities based on addresses or other identifying details.

For example, if there are 12 different entries for UCLA Health’s Department of Cardiology, the agent could deduce that these records should be consolidated. It might search online to confirm addresses, look at historical trials and verify duplicates, updating your data accordingly. If unresolved issues remain, the system could escalate these to a human data steward, significantly enhancing productivity and accuracy in data management.

As is evident from this example, generative agents aren’t mere tools; they’re intelligent partners that work across various patterns of use cases, whether it’s identifying sources of information, analyzing content, generating text or code or interacting with external tools and humans. LLMs help these agents reason, plan and execute tasks, transforming the way enterprises approach problem-solving. 

The complexity of real-world use cases often requires the coordinated efforts of multiple agents working together to achieve transformative outcomes. Akin to a symphony orchestra, each agent plays a distinct role but all work harmoniously toward a common goal. 

A good multiagent system will start with a thorough plan to address the tasks in a workflow. This plan includes collaboration with existing stakeholders, enlisting the most suitable agents for each task, leveraging the right tools, taking decisive action and continuously evaluating and improving system performance.

Consider a scenario where nontechnical business users can query large databases using natural language—a feature often called “Talk to your data.” This process might first involve a planning agent that analyzes the user’s query and devises a strategy to address it.

The planning agent could come up with an initial plan and recruit various other agents: one to identify which database tables and columns are relevant so the planning agent could in turn refine the initial plan or pseudo code, another to generate the actual code, and perhaps a code verification agent to ensure the code’s functionality and efficiency. A narrative generation agent might then articulate the results, providing insights derived from the data.

This dynamic and collaborative approach ensures workflows remain efficient and adaptable to a wide range of user needs. In fact, this is a simplified summarization of what our own “Talk to your data” product does and has proven to be very effective at bringing back answers from complex databases for our clients, with minimal burden of configuration. 

The next frontier of multiagent workflows will bring both classical AI and generative AI together to manage processes and decision-making. In some cases, this might even be autonomous, and in others, guided by humans.

Consider the following examples of enterprise processes and workflows that could be significantly disrupted and transformed through multiagent collaboration:

1. Enhancing supply chain resilience: Agents can analyze supply chain data, gather external intelligence, identify bottlenecks, leverage models to predict disruptions and demand, and leverage integrations to place orders and adjust inventories, ensuring a more resilient and adaptive supply chain.
2. Streamlining the regulatory document authoring process: Agents can retrieve data from various databases and documents, synthesize insights, draft initial versions of regulatory documents, critique and refine these drafts, and incorporate human feedback, effectively orchestrating the entire authoring process.
3. Optimizing clinical trial execution: Agents can predict potential patient drop-offs, execute interventions to minimize attrition, assess the need for additional trial sites to ensure timely completion, and make strategic decisions with human oversight about which sites to potentially add, thereby optimizing the clinical trial process.
4. Facilitating marketing content creation: Agents can analyze customer engagement patterns, identify gaps, create targeted marketing content within established guidelines, and execute campaigns using marketing automation platforms.
5. Orchestrating patient support and services: Agents can deliver personalized educational emails to patients, especially those not adhering to prescribed therapies (or those predicted to deviate). They can also schedule patient support services and nurse education, thereby enhancing patient adherence to therapy, ensuring the right support reaches those who need it most to ultimately improve overall health outcomes.
6. Supporting software product development: Agents are a great complement to the growing use of code co-pilots in software development. They can gather and summarize user requirements under human oversight, help synthesize these into functional and nonfunctional requirements, draft user stories and provide leadership with summaries of overall progress across agile workstreams.

As you can imagine, we see countless other processes and workflows to transform with multiagent systems.

From a technical perspective, groundbreaking open-source frameworks like LangGraph and AutoGen are rapidly advancing this field by adding new capabilities at an impressive pace. These frameworks facilitate the creation and management of multiagent systems by offering modular components and integration tools that simplify development.

Moreover, major cloud providers are working to simplify the design, configuration and management of these systems, enhancing the reusability of agents and their integration across different platforms.

Regarding the brains behind these agents, leading LLMs such as GPT-4o are of course being widely used. We’re also seeing increased use of smaller models. Fine-tuned open-source smaller models are especially becoming an increasingly attractive proposition for specific tasks within agentic workflows, given their performance, higher cost-effectiveness and lower latency.

As these technologies continue to mature, we expect to see innovative applications of multiagent systems proliferate further.

While the potential benefits of generative AI and multiagent systems are vast, it’s also important to address potential challenges and ethical considerations. For instance, issues related to decision-making biases, emergent unintended and hazardous behavior, potential misuse of tools and human overreliance on automated systems must be carefully managed. 

With the ability to analyze, reason and execute plans, generative agents and multiagent systems will soon become an essential tool for businesses looking to create competitive advantage. To capitalize on these advancements, start exploring and experimenting with agentic workflows today. Beginners can start with quick-to-implement use cases. Or you can contact us for more ideas on how to get started.