Beyond the Hype: Making AI Agents a Reality
AI agents hold immense potential, poised to reshape how we work by turning labor into software—a market estimated to be worth trillions of dollars. Despite rapid advancements in AI capabilities, there’s a noticeable gap between technological progress and real-world adoption. This article explores the critical layers needed to bridge that gap and unlock the full potential of AI agents, examining the shift from RPA to APA.
Many organizations are eager to implement AI, with 63% of leaders viewing it as a high priority, yet 91% don’t feel prepared. The key to unlocking this potential lies in understanding the missing pieces of the AI agent stack:
- The Accountability Layer: Ensures transparency, verifiable work, and reasoning.
- The Context Layer: Unlocks company knowledge, culture, and goals.
- The Coordination Layer: Enables seamless collaboration among agents through shared knowledge systems.
- Empowering AI Agents: Providing tools and software for maximizing autonomy in the B2A sphere.
As we address these challenges, we can build the infrastructure necessary to tackle more complex and valuable tasks with AI. The future will see a shift from Robotic Process Automation (RPA) to Agentic Process Automation (APA).
Unlocking Autonomy: From RPA to APA
Robotic Process Automation (RPA), a multi-billion dollar industry, has already demonstrated the willingness to adopt automation for key tasks. RPA excels at handling rule-based, structured processes across various business systems. Its strengths lie in efficiently capturing company knowledge within set rules, making automations reliable when underlying systems remain static.
However, RPA has limitations. It struggles with tasks that lack clearly defined, repeatable steps, and lacks flexibility and understanding of dynamic contexts. The RPA’s dependence on process mapping and strict repeatability limits its adaptability.
The rise of LLMs marks a major shift towards cheap, adaptive intelligence that can define and collate the context required for handling complex problems. LLMs are suitable with unstructured business data, and can handle reasoning, but they also have shortcomings. LLMs can be a “black box,” and produce inconsistent results, and may even hallucinate their reasoning. This lack of certainty hinders enterprise implementation and adoption, even when a user wants more creativity in results. The solution lies in addressing the strengths and weaknesses of both RPA and LLMs.
The answer for agents and APA: We need the reliability of an RPA system, coupled with the flexibility and affordability of an LLM. This is possible with an auditability and context layer in the AI agent stack—something that builders in this business should be focusing on if they want to achieve widespread adoption.
The Accountability Layer: Adoption, Learning, and Supervision
Just as students must “show their work” in math, AI systems need to provide auditable trails of their actions to verify reasoning and to ensure understanding of the process. This is particularly key because AI systems do not always generate the action or train of thoughts.
Maisa, a company highlighted in the article, developed a concept of “Chain of Work.” This illustrates how important it is for AI agents to be implemented in the workforce. Maisa uses its Knowledge Processing Unit (KPU), a proprietary reasoning engine, to orchestrate each AI step as code rather than relying on ephemeral “chain-of-thought” text, which facilitates deterministic and auditable outcomes. This means every action is logged in an explicit “chain-of-work.” This technology increases trust and accountability, providing enterprises with clarity on how specific AI actions are taken, which can then allow them to assess, correct, and refine each step.
The Accountability layer helps enterprises reduce risk and improve the likelihood of successful AI deployment. This approach is also built to help employees be more successful when integrating AI.
The Context Layer: What Makes a Great Employee?
A great employee is much more than just credentials or experience; success is dependent on adaptability, communication, and an understanding of company specifics. GPT-4, for example, is not a great employee precisely because it doesn’t understand any one organization. The AI lacks the nuance and contextual understanding you’d find in a human employee.
Much of the knowledge required for success is not written down, or is located in unstructured data that most AI tools cannot access. This is where the context layer becomes critical, allowing AI agents to access and incorporate the “unwritten stuff” that distinguishes a great from a good employee.
Maisa’s Virtual Context Window (VCW) functions as an OS-like paging system. This system allows them to “load” and “navigate” only the data they need which provides zero collisions. The VCW serves as a long-term know-how store that can adapt to new instructions. This contextual layer is critical for allowing a customer to “onboard” an AI worker into their organization.
The Coordination Layer: Managing the Agentic Workforce
Businesses will ultimately manage multiple AI “employees” and agents. AI will be tasked with customer service, sales, HR, accounting, and other roles. Agents will need to communicate with humans and each other; as well as requiring proper permissions and security.
As AI agents proliferate, there are two possible scenarios: companies tightly control these systems, creating a winner-take-all system, or independent developers solve the communications and permissioning problems faster than incumbents can. A thriving agent ecosystem provides the customer with a diverse pool of potential AI talent, and it provides founders with the opportunity to enhance their products through network affect.
Flexibility is essential in a future where foundational models continuously improve. With AI agents, it will be necessary to provide systems to safely exchange and share knowledge.
The Frontier: Giving AI Agents Tools for the Job
Once we tackle accountability, context, and coordination, we move into a domain of tools created by B2A (business-to-agent) developers, to make AI agents better at their jobs. This is a major step in making AI more autonomous. Agents are set up for success by providing them with access to software and tools.
We have already seen examples of AI agents utilizing web browsers and giving them a voice. These examples will get 10X better, and give agents the ability to take action in the world.
What it Takes to Onboard AI Agents
For AI agents to be an integral part of workflows, businesses must create the necessary foundational layers to empower their AI agents. This article concludes that AI agents will play a growing part in the future. By addressing challenges in the new computing paradigm, adoption gaps can be overcome. Companies that recognize these challenges and innovate will be at the forefront of the AI agent revolution.
