Sentius White Paper
Executive summary
- Generative AI and large language models enable AI agents to reason, plan, and execute complex tasks autonomously, surpassing the limitations of traditional automation and Symbolic AI.
- By adopting autonomous agents, organizations can offload routine tasks to AI, allowing humans to focus on strategic, creative, and empathetic roles, thereby creating a new type of organization – Agentic Enterprise.
- The Sentius Teach & Repeat Platform empowers organizations to create autonomous AI agents that learn and replicate company-specific processes and knowledge — using tools like the Task Recorder, Multi-Agent Systems, Knowledge Graphs, and specialized agents — enhancing efficiency, reliability, and adaptability in business operations.
Humanity's cooperative evolution has continually pushed the boundaries of what's possible — from the pyramids to space stations, our drive to build and innovate has only intensified. Today, we enter another transformative era, driven by artificial intelligence. AI agents are emerging as a revolutionary force, pushing the development of the Agentic Enterprise — a new organizational paradigm enabling companies to thrive amid a rising wave of disruption and unprecedented opportunities.
In 2023, Bill Gates predicted that AI agents would reshape the tech industry within five years. Recently, Vinod Khosla forecasted that AI would perform 80% of the work in 80% of jobs. This vision is already materializing as organizations integrating AI are five times more likely to make fast, data-driven decisions. As of early 2024, 72% of business leaders report high productivity due to extensive AI integration. And this is just the beginning. The global AI market is expected to exceed $1.8 trillion by 2030, with AI agents projected to increase global GDP by 26% in the same timeframe. Adaptability is now critical; organizations leveraging AI agents will endure and flourish, while those that resist risk fading into obscurity.
Evolution of Work
From Lever to Digital Enterprise
Throughout history, we have adapted our environments, accumulated knowledge, and solved increasingly complex challenges. Ancient innovations like the lever allowed us to accomplish feats previously unimaginable. And each technological breakthrough — from stone tablets to the specialization of labor — has acted as a lever for progress.
Money, invented about 7,000 years ago in Mesopotamia created infrastructure for scaling human cooperation beyond a neighborhood. In ancient Egypt, the construction of the pyramids show the power of coordinated specialization: stone quarrying, transportation, and assembly were organized into an efficient system. This capability to divide labor laid the groundwork for larger, structured organization. Centuries later, the industrial loom brought a comparable shift, drastically reducing costs and creating global shifts in the production and trade of raw materials. These innovations, while boosting efficiency, have consistently introduced new layers of complexity to human labor structures.
The Information Age marked a defining shift in work structures. Acceleration of information flows within organizations made them more efficient. Bill Gates envisioned the "digital nervous system" carrying information flows across companies. Today, platforms like Microsoft 365 and Google Workspace have made this vision a reality, enabling organizations to operate "at the speed of thought."
However, cognitive overload remains a limiting factor. As technology advances rapidly, only those will win, who will delegate increasingly complex tasks to machines. The "ChatGPT moment" in 2022 marked a significant milestone, where assigning cognitive tasks to AI on a personal level — once considered science fiction — became a practical reality. Today, the capabilities of Generative AI enable individuals to achieve higher productivity and adaptability. This leap forward brings us closer to the vision of the Agentic Enterprise, where not just everyone has their own AI assistant, but a hybrid workforce of humans and autonomous agents collectively tackles organizational challenges.
Waves of Automation: From RPA to GenAI
The evolution of automation in modern organizations has unfolded in waves. Standardized software made document creation, calculations, and data sharing routine. Custom software enabled companies to address unique challenges, though it required engineering support. Robotic Process Automation (RPA) added further autonomy to business processes but was often cost-prohibitive and fragile.
Tasks within an organization vary: some can be performed by machines with minimal cognitive input, while others require complex coordination across multiple roles. Consider customer service in logistics. To book a flight or ship goods internationally, a company must coordinate between departments, consider tariffs, schedule transport, and manage legal compliance. Each step requires specialized expertise and the capacity to adjust to changing circumstances. While standardized processes exist, they inevitably involve variability and necessitate real-time adaptability.
Customer experience managers, for instance, are entrusted with overseeing client interactions, ensuring quality, and managing exceptions. Their role requires autonomy and a degree of "agency" to adapt to each scenario. As Generative AI continues to mature, these capabilities will extend to autonomous agents, capable of independently managing complex workflows across teams and departments. This is the foundation of the Agentic Enterprise.
The Rise of the Agentic Enterprise
Generative AI has unlocked the potential for autonomous agents to perform actions needed to achieve specific goals. Unlike prior RPA systems limited to predefined responses, these agents will be able to interpret dynamic situations, performing sequences of actions that span people, teams, and organizations. This independence will enable them to assist in tasks from gathering insights to interacting with systems on behalf of their human counterparts.
Consider a browser agent capable of autonomously navigating and interacting with websites, a prompt agent transforming unstructured data into insights, or a document agent with long-context capabilities to answer questions within procedural guidelines. Each agent brings unique capabilities, supporting faster, more informed decision-making while easing cognitive burdens on human team members.
Over time, these agents will collectively reshape organizational workflows. Teams of autonomous agents can collaborate much like human specialists, delivering tailored solutions in complex situations. As this AI workforce grows, so will organizational efficiency, scalability, and resilience. Humans will shift into more strategic roles, focusing on creativity, innovation, and empathetic decision-making, while agents handle routine but vital tasks. This heralds a transformation of traditional structures into what we can call an "Agentic Enterprise."
The Agentic Enterprise represents more than just an evolution in technology; it's a fundamental shift in how we perceive work and collaboration. By integrating autonomous agents into our operations, we unlock unprecedented levels of efficiency and adaptability. These agents will democratize expertise, making high-level skills and knowledge accessible across all tiers of an organization.
This transformation will redefine what it means to work and, ultimately, what it means to be human. Freed from routine tasks, we can focus on pursuits that inspire us, fostering a society that values creativity, empathy, and shared success.
Sentius Agentic Enterprise Stack
Teach & Repeat Platform
Autonomous AI agents are being built based on the ability of Generative AI to autonomously reason, plan, and act. Advancements in base models capable of doing web browser management (OpenAI Operator) and desktop applications (Anthropic "Computer Use") illustrate rapid progress in their ability to find solutions to "never-seen-before" tasks (aka "zero-shot mode"). This allows them to solve microtasks in many businesses. However, these models were trained on interactions with publicly available websites or applications and were not trained on organization-specific information like processes, best practices, and policies. This means these models have to try to find solutions while interacting with business-specific sites - both provided by major SaaS providers like Salesforce, Microsoft, or Google, and those developed in-house. These systems are customized to the needs of organizations, and interactions with them haven't been available to public Generative AI models. But why seeing these internally deployed systems essential for building reliable autonomous AI agents?
MOTIVATING EXAMPLE
AI agents can be compared to self-driving cars. Indeed, self-driving car is provided with a goal to reach which is exactly what an AI agent is asked to do, too. Both an AI agent and a self-driving car are asked to execute their task autonomously, i.e., without human intervention.
Tesla's strategy revolves heavily around its fleet of vehicles on the road. Tesla is leveraging real-world data collected from its customer base to create maps of the real world and routes taken by its customers. Waymo's strategy also collects data; but as Waymo doesn't sell cars to the general public, it collects similar data using its autonomous machines.
Both solutions have to ensure that they can handle numerous edge cases and anomalies encountered in real-world driving. This requires a continuous iterative process that ensures that resulting system is constantly updated and rigorously tested to address every possible scenario.
Self-driving solutions require mapping territories, remembering routes within them, mechanisms to trace these routes and finding the optimal ones, sets of rules and regulations that are needed to be taken into account during real-time operations, mechanisms to predict the next few moments and to change planned actions to alter cars' behaviors to adapt to the predicted changes in the environment.
Similar to self-driving cars, autonomous AI agents need to map out their "territories". But, unlike with Operator or "Computer Use", "territories" that are to be mapped out are not publicly available: web or desktop apps built by independent software vendors or built in-house are tailored to the needs of the companies they are deployed in. Typical midsize organization uses 100 to 200 web and desktop apps. Larger organizations use from 300 to 3000 applications (third-party and in-house).
While self-driving cars operate in the terms of routes, autonomous AI agents that control web and desktop apps operate in the terms of workflows. The difference between road routes and workflows is huge: routes (except in rare cases) can be taken in both directions while many workflows cannot be "undone". For example, you can drive on a car from San Mateo to Mountain View and back, while in the app you can send money to your supplier, but you cannot easily get that money back without spending time on processing this exception with them. The cost of error in not making a given workflow executed, hence, is much bigger.
Therefore, to facilitate a smooth transition from human-led to agent-led action execution, we believe that a mechanism for teaching our autonomous agents the organizational knowledge is essential. This "Teach & Repeat" methodology ensures that agents are not just intelligent but also know "how things are done" and are aligned with the unique operational paradigms of each enterprise.
TEACH & REPEAT METHODOLOGY
The Teach & Repeat platform functions by first employing a Task Recorder to observe and capture how an employee performs a specific task. Based on this recorded behavior, instructions for agents are created to mimic the employee's solution for that task. These instructions are then integrated into the Multi-Agent Systems & Workflow Engine, allowing it to be seamlessly merged into business processes and reused for similar tasks in the future. Furthermore, the organizational knowledge gained during the interaction with the user is stored in Knowledge Graphs and Dynamic Ontology, creating an organizational memory that enhances future performance and adaptability. Finally, a combination of Guardrails and Policies ensures that agents perform within the requirements of the organizations and only execute explicitly allowed actions.
PRIVACY & PRICING
At Sentius, our goal is to provide our customers with the right balance between privacy and pricing. Mapping of applications and workflows spanning across them can be done only within the boundaries of a given customer tenant in our cloud environment, or even bound to the on-prem deployment.
Let's explore the essential elements of the Sentius Teach & Repeat Platform that bring the Agentic Enterprise vision to fruition.
Task Recorder
The Task Recorder observes and analyzes user interface activity and application logs. It generates workflow configurations and agent instructions, effectively learning from human actions to automate future tasks. This not only accelerates the automation process but also ensures that the agents are tailored to the specific workflows of the organization.
Multi-Agent Systems and Workflow Engine
Our platform supports multi-agent systems that can collaborate across multiple users/devices to execute complex workflows. Workflows can be composed manually or generated and can be further automatically optimized for a given task. The Workflow Engine orchestrates these agents, managing dependencies and optimizing task allocation as well as healing entire workflows when unexpected changes come. This mimics human teams working in unison, but with the added benefits of machine efficiency and scalability. Workflow Engine can be used as part of our cloud-based solution or can be deployed in the secure customer's environment.
Knowledge Graphs & Dynamic Ontology
Organizations maintain structured representations of their essential business data—whether about products, customers, orders, or other key entities. To reduce AI hallucinations and ensure decision-making transparency, we anchor LLM responses to these data representations, forming knowledge graphs that allow clients to retain full control over their information. By combining knowledge graphs with dynamic ontology, we provide our agents with not only raw data but also a contextual understanding of relationships, hierarchies, and business-specific semantics within the information. This layered approach boosts reasoning capabilities, enhancing decision-making processes and making AI outputs more reliable and explainable.
No-Code Tools
To democratize access to AI automation, Sentius Teach & Repeat Platform offers no-code tools. Users without technical backgrounds can design, deploy, and manage agents, fostering innovation and enabling rapid adaptation to changing business needs.
Guardrails and Policies
The importance of having AI agents perform under strict control cannot be overstated, as it ensures that their actions are aligned with ethical standards and user expectations. By being guided by well-defined guardrails and steered using human-formulated policies, AI systems can operate safely and effectively, reducing the risk of errors or misuse. For instance, consider a policy that prohibits AI agents from transferring money while interacting with a banking app unless they have explicit user authorization. Such a policy ensures that sensitive financial transactions are conducted only with the user's direct consent, preventing unauthorized access and potential fraud. This not only protects users' financial security but also builds trust in the AI systems, demonstrating a commitment to responsible and transparent AI usage.
Conversational AI Orchestrator
The Conversational AI Orchestrator enables people to interact with the platform's components in a natural, conversational manner. It includes guardrails that have been battle-tested in Amazon Alexa Prize competitions, ensuring safe and effective communication between humans and agents. This fosters a more intuitive user experience, lowering the barriers to adopting AI solutions within the organization.
Sentius Agents
Autonomous AI agents provided by Sentius can be set up to run on the end-user devices (to operate within a given security context and/or under human supervision) or in the VDI infrastructure provided by customer or Sentius.
Browser Agent
The Browser Agent understands tasks defined in natural language and performs them on the web via the user's web browser. Imagine delegating routine online tasks—such as data entry, information retrieval, or even complex transactions—to an agent that executes them flawlessly and efficiently. This not only saves time but also reduces the potential for human error.
Prompt Agent
The Prompt Agent is designed to execute lightweight AI tasks on input data. A basic example includes sharing pre-existing knowledge, such as common-sense information about well-known products and services. The Prompt Agent can provide results either as text or in a structured format adhering to a specified data schema. This agent acts as a bridge between raw data and actionable insights, enabling faster decision-making processes.
OpenAPI Agent
The OpenAPI Agent uses natural language instructions to connect with internal and external APIs. This enables seamless integration across various platforms and services, breaking down silos and fostering interoperability. By translating human intentions into machine-executable actions, the OpenAPI Agent simplifies complex integration tasks that traditionally require specialized engineering expertise.
Document Agent
The Document Agent is engineered to answer in-domain questions by reasoning and connecting multiple facts over very large contexts. Utilizing Sentius' Associative Recurrent Memory Transformer technology, it demonstrates successful processing of tasks up to 50 million tokens, and it's potentially scalable to any length. It can be deployed in-house or on a private cloud, ensuring confidentiality and security. This agent revolutionizes how organizations handle vast amounts of textual data, making information retrieval and analysis more efficient than ever before.
Other Agents: Desktop, Code, SQL etc.
In addition to the core agents described above, we are actively developing and integrating top-tier autonomous agents that specialize in manipulating desktop applications, managing SQL databases, and generating code. These agents enhance our platform's versatility, making it capable of addressing a diverse array of tasks with high efficiency and precision. By automating interactions with desktop software, optimizing database operations, and facilitating code generation, we ensure that our solutions can seamlessly adapt to various business needs. Our commitment to innovation and excellence drives us to continually expand our suite of agents, ensuring our clients benefit from cutting-edge technology that streamlines operations and fosters productivity.
Custom Action Models
Reliability is paramount for enterprise systems. To achieve robust yet adaptive workflow automation, we are developing action models for target domains. As organizations start to use the Teach & Repeat platform, they begin to accumulate business know-how and other organizational knowledge. We are developing training methods to convert this accumulated knowledge into company-specific business action models. These in-domain action models will improve prediction quality and enhance the reliability of business process automation. Moreover, these models will be updated continuously and evolve alongside the company's business processes. This continuous evolution allows agents to replicate user behavior accurately, ensuring consistency and reliability in task execution.
AI Agents vs RPA
Workflow automation systems have been around for decades, even before the Generative AI era. These systems, influenced by Symbolic AI, relied heavily on explicitly defined rules and symbolic representations to perform computations and solve problems. While they offered high precision and reliability, they demanded extremely high accuracy in the information entering and exiting the automation systems. This rigidity made them less adaptable to the dynamic, real-world environments where data is often semi-structured or unstructured.
Generative AI paradigm shift allows systems to understand and handle semi-structured and unstructured information. Unlike the traditional Symbolic AI approach, Generative AI-powered systems can interpret nuanced data, though this comes with higher computational costs. At Sentius, our core philosophy revolves around harmonizing Symbolic AI with Generative AI and finding an equilibrium between human involvement and autonomous agents.
We advocate for equipping modern organizations with hybrid tools that leverage the strengths of both methodologies. For instance, when we can attain a level of accuracy ideal for Symbolic AI solutions, we utilize it for efficiency. Conversely, if processes driven by Symbolic AI falter due to changes, we harness the adaptability of Generative AI to dynamically fix these processes. This symbiotic relationship ensures that we maintain high reliability while also being agile enough to adapt to new challenges.
The Future is Agentic
The Agentic Enterprise represents more than just an evolution in technology; it's a fundamental shift in how we perceive work and collaboration. By integrating autonomous agents into our operations, we unlock unprecedented levels of efficiency and adaptability. These agents will democratize expertise, making high-level skills and knowledge accessible across all tiers of an organization.
This transformation will redefine what it means to work and, ultimately, what it means to be human. Freed from routine tasks, we can focus on pursuits that inspire us, fostering a society that values creativity, empathy, and shared success.
