inLABs — research projects

“A partial or negative result will be a valid research outcome”

“The system should clearly show its limitations instead of replacing human judgement with apparent certainty”

“CDT can provide the context needed, but it does not make decisions and does not relieve a person of accountability”

“This description does not mean a finished integration, a shared operating environment or data flows between these solutions”

“This page does not announce a product, a pilot programme, availability or a commercial schedule”

“Detailed procedures, configurations, data and results are not part of this public description”

CDT (Cognitive Digital Twin) is a concept being developed in inLABs to organise knowledge related to a particular person's tasks and everyday work context. We study how this knowledge, together with its rationale and information about its currency, can be collected, organised and used responsibly. The goal is to support continuity of work, not to reproduce a person's identity or take over their decisions and accountability

The technology is not intended to imitate a person or present uncertain information as certain. Above all, we want to distinguish honestly between what can be described and passed on and the knowledge that still requires experience, conversation and practice

Why it is being developed

Important organisational knowledge is usually dispersed across procedures, decisions, instructions, other documents, conversations and everyday practice. When a role, team composition or priorities change, it is easy to lose not only information about what was done, but also the reasons for choosing a method, the conditions considered and the precautions taken

Simply collecting files does not solve the problem. Information unconnected to a specific task, person, source and creation date can be difficult to find, out of date or used incorrectly. Conversely, overgeneralising an expert's experience can create a false sense of certainty, especially in unusual situations that require years of practice

That is why we study how to preserve knowledge continuity without reducing it to a list of facts. This includes consent to use information, access rules, knowledge ownership, the ability to correct and delete it, currency and responsible use. The system should clearly show its limitations instead of replacing human judgement with apparent certainty

Potential application scenarios

The examples below show possible research directions. They do not describe finished deployments, available features or services

Role in the inLABs ecosystem

Research approach

We assume that useful knowledge must stay linked to the task, context, source, date of creation and level of certainty. The system should help recognise not only facts and procedures, but also the conditions under which they apply, their rationale and signals that information needs to be confirmed again

We are studying three core capabilities: acquiring and organising knowledge covered by consent, linking it to a specific task and presenting it to the person who still makes the decision on their own. The result is not meant to be a digital copy of a person, but a clearly labelled supporting tool that helps people understand, check and pass on knowledge

Not all knowledge is easy to describe. Some experience can be written down as facts and procedures, some can be drawn out in conversation, and some can only be learned through observation and practice in a specific situation. We do not promise to fully reproduce this last category. The more honest solution is to indicate when contact with an expert or practical preparation is needed

Research methods and validation

Developing CDT requires checking both the quality of the knowledge collected and the rules for using it. We are planning controlled research cases, expert reviews and voluntary studies with people of different levels of experience, conducted with appropriate data protection. We will compare ways of describing context, communicating uncertainty and preparing materials for the person taking over tasks

The assessment will cover, among other things, usefulness in a specific task, consistency with the sources, how current the information is, clarity about limitations and the quality of the information passed to a person for further assessment. In parallel, we are studying the rules for consent, access, identifying the author of the knowledge, and correcting and deleting it. Detailed procedures, configurations, data and results are not part of this public description

HCG (Human Competence Gate) is a way of checking, before an AI-supported decision is approved, that the right person understands the situation, can assess the recommendation and consciously takes responsibility for the decision made

In many processes the organisation uses AI recommendations, but a person still decides what happens next. The methodological foundations of HCG have already been developed, and the corresponding control stage is described in the CDF implementation methodology. Within inLABs we are now studying how to apply such a check in specific processes and how to assess its effectiveness

Why approval alone is not always enough

A person being present in the process does not in itself guarantee the quality of oversight. When recommendations appear frequently and their content is complex, approval can become a reflex. The decision-maker may not have a full picture of the premises, exceptions or limits of the recommendation. In such a case the system has a formal record of approval, but the organisation still does not know whether the decision was actually considered

This problem is particularly important where a decision affects a client, an employee, operational safety, resources or the continuity of a process. Speed is valuable, but it cannot lead to a situation in which a person signs off a result without being able to understand what it means in the case at hand

A second challenge is the difference between general knowledge and preparation for a specific decision. A competent person does not always have the current context of the matter, and a person who knows its details may need support in assessing the limitations of an AI recommendation. The organisation therefore needs an approach that does not reduce accountability to the user being present in front of the screen

HCG answers the question of how to make a person’s decision really matter: before an action is approved, it makes it possible to make sure that the person responsible understands the specific matter and has the competence needed to assess it

It is important to keep a sense of proportion: the amount of attention and support should match the significance of the specific decision, rather than creating the same obstacle for every action

Where HCG may matter

The scenarios below are possible directions of application, not a description of deployments or available features

In methodological terms, approving an action may require confirmation that the person responsible has the appropriate competence and understands the matter. HCG does not, however, determine which decision is right; that remains with the person responsible for the process

Role in the inLABs ecosystem

Principle of the approach: from recommendation to an informed decision

The starting point of the work is a simple hypothesis: formal approval has limited value if the approver does not understand what the decision concerns and what its limits are. HCG studies how the elements requiring informed assessment can be presented before a decision, so that the person remains a real participant in the process

The approach can be described through three elements. The first is the context of the decision: what it concerns and what information is important. The second is understanding: whether the approver can relate to the assumptions, limitations and possible consequences. The third is responsible action: approval, consultation or withdrawal. This model refers to the requirements for human competence and oversight described in the AI Act and ISO/IEC 42001, but it is not a declaration of conformity or legal advice

The aim is not to replace human judgement with an automatic assessment. The goal is a way of organising decisions that helps a person consider a specific matter carefully. HCG therefore differs from general training or a universal checklist: it takes into account the situation, the responsibility assigned to the role and the limitations of the given recommendation

The approach must remain understandable to the person using it. The research will therefore concern not only the moment of decision itself, but also whether the context presented helps people see uncertainty rather than seemingly removing it

Research and validation

Further research on HCG may include analysis of different decision situations, expert reviews and a comparison of several ways of supporting a person before an action is approved. We will assess, among other things, how clearly the context is presented, whether an additional check is justified, how useful it is for the decision-maker and whether the person’s role in the whole process can be explained clearly

The limits of this approach also need to be defined. An additional check should improve the quality of decisions, not become an empty formality or an excessive burden on the process. The research must therefore take into account different roles, levels of risk and the real conditions in which the organisation operates

It will be important to assess whether the form of support leaves the decision-maker with a clear basis for their own judgement, instead of suggesting automatic certainty. Depending on the scenario, both qualitative analyses and comparisons of how the process behaves before and after the control point is introduced may be needed

Subsequent work may focus on distinguishing the decisions in which a human control point is genuinely justified, on ways of presenting the limitations of recommendations and on keeping a clear boundary between AI support and human accountability. Every future scenario requires a separate assessment of the context, risk and organisational rules, and an accountable decision owner

AICF (AI Compliance Framework) is an inLABs research and development project for which we are seeking funding under the European Funds for a Modern Economy (FENG) programme. We want to study how organisations can continuously oversee the compliance of their use of generative AI with regulations, standards and their own rules. We do not want to reduce compliance to a one-off checklist. We plan to examine how to describe the organisation’s requirements clearly, relate them to a specific task and retain the evidence later needed by the person responsible for the assessment. Work will begin only once funding has been granted; the project cannot be launched without a positive outcome of the application assessment

Why this research matters

Generative AI increasingly supports business processes in which data, decisions and obligations change. A policy or assessment prepared before deployment is still needed, but it does not always explain what to do after a change of task, model, data source, tool or organisational conditions. AICF is meant to focus precisely on this gap between a written requirement and the way AI is actually used

Potential application scenarios

The examples below show possible research directions. They do not describe client deployments, product commitments or the current availability of a solution

Role in the inLABs ecosystem

From requirements to verifiable evidence

At a public level, the research can be summed up as a simple relationship:

We want to check whether requirements can be written down in a way that makes it easier to assess AI-supported actions in a specific context. We also plan to study whether significant changes can be signalled without creating a false sense of certainty, and whether the evidence collected helps the person responsible understand and challenge the basis of a decision. AICF is meant to concern ongoing compliance control; it will not be a universal security tool, a legal opinion or a system that approves decisions automatically. This description does not constitute legal advice

Detailed mechanisms, data models and decision procedures are deliberately left out of this public description. What matters is the direction of the planned research: control that is more continuous, interpretable and linked to real accountability within the organisation

Research methods and validation

AICF is a research and development project whose assumptions will require evidence, not claims, once it has been launched. The work is to include limited test cases, controlled comparisons with suitably chosen reference approaches and qualified expert assessment. The assessment will concern the comprehensibility, consistency, traceability and robustness of the approach in a defined context. It will also take into account situations in which requirements overlap and the available information is incomplete

The results will require careful interpretation. A positive result in one scenario will not establish universal compliance, legal sufficiency or suitability for another organisation. A partial or negative result will be a valid research outcome: it may show where human verification, a clearer requirement or a different approach to control is needed. Detailed research materials and unpublished results will not be part of this page

SAVANT-AI (Semantic Autonomous Versatile Advanced Network Technology — Artificial Intelligence) is an inLABs research and development project. We are studying how to bring together organisational knowledge, expert experience, AI agents and business systems in a single, controlled working context. It is not about replacing people or announcing a finished product. We want to check whether complex organisations can make better use of scattered information while preserving human accountability, access rules and the ability to trace information back to its sources

Why this research is needed

Important organisational knowledge is created in documents, operational systems, expert conversations and everyday decisions. Each source shows only part of the situation, often describing it in different language and from a different perspective. To answer one important question, a team has to find the information, check how current it is, establish what it means and reconcile it across departments. This is necessary, but time-consuming and prone to losing context

A second problem is knowledge that depends on specific people. When experience is not visible to other teams, or it is unclear where to look for the right expert, the organisation finds it harder to learn from its own decisions and changes. The risk grows when processes are complex, data is sensitive and the consequences of a mistake go beyond one department

Searching documents alone is not enough. The question also has to be understood in the context of the task, information has to be combined carefully, access rules have to be taken into account and a person has to be able to assess the conclusion received. SAVANT-AI explores this space without assuming that every answer can or should be produced automatically

Potential application scenarios

The examples below are research directions, not a description of deployments or available features, nor a promise of results

Role in the inLABs ecosystem

General principle of the approach

The starting point is simple: an organisation can make better-founded decisions when the right sources, the user’s question and expert knowledge are analysed together instead of remaining in separate places. The system receives only information approved for use and a question embedded in a specific task. It then helps organise the context, indicate uncertainty and determine what still needs to be checked. The result is meant to be material that supports a person’s assessment

The difference compared with simple search lies in studying the relationships between information, roles and the consequences of using knowledge. This does not mean, however, that the system automatically understands the organisation, makes decisions or determines whether conclusions are correct. The scope, reliability and usefulness of the answers remain the subject of research and depend on the quality of the sources and the context of application

Research methods and validation

Work on SAVANT-AI may include comparisons with suitably chosen baseline approaches, expert assessment and qualitative and quantitative study in controlled cases. The properties analysed may concern the relevance of the context, usefulness for the user, the ability to find the basis of reasoning, robustness to data changes and the limits of the solution’s use

An important part of validation is also checking when an answer should not be treated as sufficient: when the sources are incomplete, contradictory, unavailable or require a decision by an authorised person. Security, privacy and compliance will be assessed in relation to a specific environment, data and way of use, not declared as a universal property. On this page we do not present experimental results, benchmarks or non-public procedures

GENESIS-AI is an inLABs research and development project for which we are seeking funding under the European Funds for a Modern Economy (FENG) programme. The project concerns the use of generative AI in software engineering. We plan to examine how to move from approved business and technical requirements to a consistent package of application elements that can be reviewed and assessed. Responsibility for engineering decisions will remain with people

The research is to cover selected types of web applications. Mobile and desktop applications and real-time systems remain out of scope. We do not want to replace the product owner, the architect or the development team. We want to check whether better organisation of work, regular verification and the ability to trace the link between a requirement and the result increase the reliability of AI-supported work. Delivery depends on funding being granted; the project cannot be launched without a positive outcome of the application assessment. GENESIS-AI is a research and development project, not an announcement of a finished product or service

Why this research is needed

Generative AI can quickly produce fragments of software. It is much harder to build a complete application that still serves its real purpose. That is why we plan to focus GENESIS-AI research on four related problems, provided the project receives funding

Possible applications

Role in the inLABs ecosystem

Approach and operating principle

GENESIS-AI is meant to study the relationship between three levels:

Approved requirements are to be the point of reference. On their basis, related application elements would be produced in a controlled way and then checked. The finished package would go to the people responsible for technical assessment and for the decision on further development or possible deployment

This is a general description of the planned way of working, not of the solution’s architecture. We want to check whether better organisation of requirements and explicit checking of results can reduce the gap between the original intent and the outcome. We also plan to study whether shared assessment criteria for code, tests and documentation lead to a more consistent result than producing these elements independently. These are hypotheses to be tested only after the project has been launched; the experiments have not yet begun and depend on funding being granted

Research methods and validation

The reliability of GENESIS-AI is to be assessed through complementary, planned classes of research: controlled comparisons with reference approaches, reviewed cases, testing of repeatability across different classes of problems and qualitative and quantitative assessment in a defined context. The analysis is to cover, among other things, consistency with requirements, functional and integration correctness, reliability, maintainability, portability, security, controllability and usefulness for the engineering team

The assessment will not be based solely on demonstrations or the volume of material produced. Automated checks will require expert interpretation, and repeated trials should distinguish a lasting effect from chance. A partial or negative result will be a valid research outcome; evidence of failure will be equally important: unresolved ambiguity, divergence between artefacts, limitations of the review or the need for additional human work

The planned direction of further work, to be carried out once funding is granted, includes studying fidelity to requirements, managing ambiguity, consistency across the artefact life cycle, reliability across different classes of problems and ways of conducting reviews without losing human accountability