3. How SenAi differs from other AI agent systems

Most AI agent systems are designed primarily to perform specific tasks. Agents receive instructions, use available tools, analyse information and carry out defined processes.

SenAi starts from a different premise: the system should not only perform work, but also learn about the enterprise in which it operates, learn its processes and gradually develop its own knowledge of how they are carried out.

This approach rests on an architecture inspired by the Axiomatic Theory of Self-Consciousness (AToS), combining LLMs, PAS scenarios, internal process initiation mechanisms, change recognition and controlled improvement of operating methods.

Individual elements can also be found in other AI solutions. SenAi’s intended distinction is their combination in a single system where task execution, knowledge acquisition and process development are parts of the same mechanism.

3.1. SenAi should learn from employees, rather than employees learning SenAi

Traditional information system implementations require prior analysis of business processes, documentation, configuration and user training.

SenAi proposes reversing this approach.

Employees perform their everyday duties with an AI agent. They explain their objectives, identify documents, make decisions, correct proposals and approve results.

The agent does more than execute instructions. It also analyses how the task is performed, its context, the rules applied and dependencies between activities.

SenAi can then propose recording the acquired knowledge and creating or extending a process scenario.

An enterprise need not therefore begin implementation by describing all its processes in detail. The system is intended to learn them gradually through participation in the organisation’s actual work.

This does not mean abandoning oversight. Acquired information does not automatically become a binding rule. Its meaning, correctness and potential for reuse are subject to verification.

3.2. PAS — processes recorded as knowledge rather than program code

PAS are a fundamental element of SenAi’s architecture.

In SenAi’s implementation, PAS serve as declarative records of knowledge, rules, process scenarios and execution conditions. Their concept refers to the formal notion of PAS in AToS.

In traditional software, much business logic is directly tied to application code. Changing a process may therefore require analysts, programmers and testers.

SenAi separates these concerns.

PAS specify what should be done and under which rules. AI agents interpret scenarios and organise their execution. ENGINE provides the technical functions needed to perform operations.

If an enterprise changes document approvals, the order of activities or decision conditions, the change can be introduced in PAS without modifying ENGINE’s code — provided the necessary technical functions already exist.

Process knowledge can thus develop independently of the execution infrastructure.

Scenarios can also be reused. Knowledge acquired during one task can become the foundation for subsequent actions.

3.3. The LLM as a mechanism for interpreting stimuli and initiating actions

In SenAi, language models are not merely a conversational interface.

Within the AToS-inspired architecture, they perform two functions:

InspFu — the inspiration function, triggered by external stimuli such as user instructions, documents, messages or information from other systems.

ImFu — the imagination function, triggered by internal stimuli arising during SenAi’s own operation.

This distinction matters.

An agent system need not stop after executing a single instruction. One analysis may reveal a problem, missing information, a contradiction or the need for another activity.

SenAi can then initiate the next stage using its operational state and recorded knowledge, without waiting for a new human instruction.

ImFu and InspFu need not be two different LLMs. They are distinct architectural functions whose operation depends on the origin and nature of the stimulus.

3.4. StSe — a mechanism for initiating internal processes

StSe have particular significance in SenAi’s architecture.

They are PAS stored in a vector database that serve as stimuli initiating further internal processes.

Suppose an employee asks SenAi to prepare a customer quotation.

During analysis, the agent discovers that the current price list is missing. This situation can activate an appropriate StSe, initiating a process to find and verify pricing data.

The resulting information may indicate a need to check contractual conditions and then prepare suitable clauses.

This creates a chain of related processes, initiated by internal stimuli and executed according to applicable PAS.

SenAi is therefore intended not only to answer questions and execute instructions, but also to recognise the need for further actions arising from work already in progress.

This does not imply unlimited autonomy. Further processes must respect applicable scenarios, permissions, technical constraints and required user decisions.

3.5. Delta — change recognition as a foundation for improvement

Another AToS-derived element is Delta (Δ), the state-change recognition function.

Executing a task alone is insufficient for improving how the system operates. The system must also recognise what has changed.

Delta provides the basic mechanism for detecting differences between states.

Additional analytical processes can then examine the change’s significance, possible causes and consequences.

For example, SenAi may observe that a particular process repeatedly requires employee corrections. Analysis across executions may lead to a proposal to revise its PAS scenario.

The system should not independently treat every observed change as an improvement.

Detecting a change, establishing its cause, assessing correctness and approving a new method are separate activities.

Process improvement in SenAi is intended to be controlled: agents can identify problems and propose solutions, while changes requiring approval remain under human supervision.

3.6. Full enterprise control over its knowledge

A fundamental design principle of SenAi is that all enterprise knowledge resides within a secure, isolated area of the client’s IT infrastructure (the Client Zone) and does not leave its boundaries.

This covers documents, business data, customer and employee information, PAS scenarios, process histories, decisions and knowledge acquired by AI agents.

These resources are stored in infrastructure controlled by the enterprise, particularly in the AToS-DB knowledge base.

SenAi is designed for information processing inside the protected Client Zone. Organisational data is not transferred to external AI model providers or stored in their cloud services.

The enterprise retains control over data access, storage, use, sharing and deletion.

Enterprise knowledge is entirely independent of the LLMs being used.

Language models are tools for interpreting information, analysing situations and realising InspFu and ImFu. They are not the primary repository of organisational knowledge.

Replacing or updating an LLM does not cause the loss of accumulated data, PAS scenarios or process histories. A new model can use the same knowledge, provided its behaviour meets the system’s requirements.

This independence is both technological and organisational. The enterprise need not tie its process knowledge to one AI producer, cloud platform or model version.

Knowledge belongs to the enterprise, remains in its infrastructure and is independent of language model technology.

Meeting this principle requires appropriate protection of the entire environment, including integrations, network communications, permissions and administrative mechanisms.

3.7. Small, independent and scalable LLMs running locally

SenAi does not assume that a single very large and expensive language model is necessary.

Instead, the architecture favours small, economical and scalable LLMs running locally in the client’s infrastructure (on-premises).

The preferred configuration comprises three different LLMs, available to SenAi’s agents for their assigned functions.

Using models from different technological families enables the design of independent analysis, result comparison, discrepancy detection and cross-checking mechanisms.

Three models do not automatically guarantee correct decisions. Their diversity is intended to reduce shared-error risks and provide additional quality control.

Scalability is another important architectural feature.

Depending on user numbers, workload and process complexity, the enterprise can increase computing resources or model instances without rebuilding the knowledge base and PAS scenarios.

Separating LLMs from process knowledge also permits gradual replacement with newer, more efficient or cheaper models.

SenAi does not require all enterprise knowledge to be encoded in the parameters of a huge AI model.

Knowledge resides in AToS-DB and PAS. LLMs interpret, analyse and use it within defined processes.

This approach has three main objectives:

• Security — data and knowledge processing remain within the client’s protected infrastructure.
• Technological independence — the enterprise is not dependent on one supplier or a particular LLM version.
• Economy and scalability — smaller models can be used and computing resources adjusted to actual workload.

SenAi is ultimately intended to combine contemporary language models’ capabilities with the control, security and predictability of systems maintained within an enterprise’s own infrastructure.

3.8. System development without continually programming processes

One of SenAi’s central goals is to reduce the dependence of business development on the traditional software development cycle.

In the conventional model, a new business need often triggers analysis, design, programming, testing and deployment.

SenAi assumes that many such changes can emerge through collaboration between employees and AI agents.

An employee describes the need, demonstrates how the task is performed and verifies the result. The agent analyses existing knowledge, finds relevant PAS and, where necessary, proposes a new scenario or a revision to an existing one.

Once approved, the scenario can be used in subsequent executions.

This does not eliminate programming. New integrations, technical functions, security mechanisms and infrastructure changes may still require IT specialists.

SenAi’s ambition is that developing business logic should not require a team of programmers and analysts on every occasion.

If validated in practice, this principle could significantly change the economics of implementing and maintaining systems supporting enterprise processes.

3.9. AToS as the conceptual foundation of the architecture

AToS — the Axiomatic Theory of Self-Consciousness — provides formal concepts for describing states, stimuli, transformations and change detection.

SenAi uses these concepts as inspiration for its architecture.

PAS describe knowledge and operating scenarios. InspFu and ImFu, realised through LLMs, interpret external and internal stimuli. StSe enable subsequent processes to be initiated, while Delta recognises changes.

Together, these elements form a system concept connecting action, observation of results, knowledge acquisition and proposed changes.

This is not a mathematical proof that SenAi’s implementation satisfies all the self-awareness conditions specified by AToS. Such a claim requires separate formal analysis and verification.

Independently of that question, AToS provides the conceptual foundation for SenAi’s proposed internal activity and knowledge-development mechanisms.

3.10. What is the proposed innovation?

SenAi’s innovation is not simply the use of multiple AI agents, a vector database or LLMs. These technologies are already used elsewhere.

The core concept is their combination in an architecture where:

• everyday user activity becomes a source of enterprise knowledge;
• acquired knowledge can be transformed into approved PAS scenarios;
• processes can initiate further processes through StSe and ImFu;
• action results can be analysed using the Delta mechanism;
• business logic can change without changes to ENGINE code;
• all organisational knowledge remains within the client’s protected IT zone;
• knowledge and process logic are independent of LLMs and their providers;
• the system can use three different, small, scalable LLMs running on-premises;
• people retain control over approval of significant decisions and changes.

The approach’s actual advantages must be established through practical deployments, primarily by measuring process time, cost, quality and security.

The most important difference nevertheless lies in the design premise itself.

A traditional information system is implemented so that an enterprise operates according to the processes programmed into it.

SenAi is intended to learn how an enterprise works, learn from its employees and then help the organisation improve its own processes.

Knowledge gained through this collaboration remains the enterprise’s property, is stored in its secure infrastructure and does not depend on a particular AI model.

Employees should not have to learn SenAi.

SenAi should learn from employees.