Responsible AI Framework

Contents

1. Purpose, Scope and Fundamental Principles

1.1. Purpose

The purpose of this Framework is to establish the fundamental principles governing the design, implementation and monitoring of AI-supported products and services developed, provided or operated by Next4biz in a reliable, controlled, transparent and human-centric manner.

This Framework sets out the corporate practices and the approach adopted by Next4biz in order to increase the benefits provided by AI systems, to mitigate foreseeable risks and to ensure that such systems are operated in accordance with their defined intended purposes.

1.2. Scope

This Framework covers the entirety of machine learning, deep learning, natural language processing, generative AI, large/small language models, forecasting, classification, recommendation, anomaly detection and AI-supported decision mechanisms that are developed by Next4biz or used within the scope of Next4biz products.

The Framework applies irrespective of whether the relevant service is operated as SaaS, in a private cloud, in a hybrid model or on-premise in the customer's environment. The scope and level of the controls to be applied are determined by taking into account the intended purpose of the system, its technical architecture, customer requirements and the risks that it may give rise to.

1.3. Nature of the Framework

This Framework explains Next4biz's general approach to the development and use of AI systems.

Product- and project-specific technical requirements, customer responsibilities and service conditions are governed separately in the relevant agreement, project document, technical specification or solution design report.

2. Responsible AI Principles

2.1. Human-Centricity and Human Oversight (human-in-the-loop)

Next4biz AI systems are developed for the purpose of supporting human decisions and business processes. The level of automation granted to AI systems is determined in proportion to the nature of the use case and its potential impacts.

Results that are uncertain, high-impact, or that the customer requires to be subject to human assessment may be routed to the relevant user or group. Routing to users and control are provided through the tools and techniques offered by Next4biz products.

2.2. Limitation to Purpose and Scope

For the AI system provided by Next4biz, the intended purpose of use and the situations in which it must not be used are defined. The systems are not used to make decisions or carry out transactions outside the defined scope.

Customer-specific limitations relating to subject matter, products, brands, communication, recommendations, and transactions may be determined within the scope of the project. These limitations are implemented by means of system instructions, knowledge sources, business rules, access controls, and output controls.

2.3. Reliability and Measurability

AI systems are validated with data sets, test scenarios, and evaluation methods appropriate to their intended purposes. The data used in the training or model configuration process and the validation and test data by which performance is measured are separated from one another to the extent possible. Care is taken to ensure that the test data is representative of the classes, user expressions, transaction types, and exceptional situations that the system will encounter in the production environment.

In the testing process, not only standard examples in which the system is expected to produce a correct result, but also incomplete or ambiguous inputs, similarities between classes, rarely occurring examples, out-of-scope requests, erroneous content, and situations in which the system should not produce a result or should route the matter to human review are evaluated. In generative AI systems, the consistency of the generated responses with the provided knowledge sources, the accuracy of the content, compliance with the scope rules, and whether prohibited responses or transactions are produced are examined through automated measurements and, where necessary, expert assessment.

System performance is evaluated not solely on the basis of an overall success rate, but by taking into account the error types required by the use case, class-based results, false positive and false negative rates, confidence scores, the need for human intervention, and end-to-end business outcomes. The performance indicators to be used and the acceptance thresholds are determined before testing is carried out, taking into consideration the intended purpose of the system and the potential impact of erroneous results.

2.4. Privacy and Data Protection

The data used in AI systems is processed in accordance with the applicable data protection obligations and with Next4biz information security policies and procedures. Data protection controls are applied so as to cover the entire data life cycle, including the ingestion of the data into the system, its preparation, its transfer to the model, its storage, its logging, and its deletion.

The principle of data minimization is observed in the design of the systems. It is ensured that personal or sensitive data that is not necessary for the intended purpose of use is not transferred to the model. In order to detect personal data such as name, surname, telephone number, e-mail address, identification number, address, financial information, and the like in model inputs, regular expressions, dictionary-based controls, Named Entity Recognition (NER) models, and data loss prevention mechanisms may be used in combination.

Personal data that is detected is transformed by means of masking, de-identification, pseudonymization, tokenization, or field-based deletion methods, depending on the use case. Where reversible tokenization or pseudonymization is used, the mapping information is kept separate from the environment in which the model operates and is protected so as to be accessible only by authorized services.

It is essential that data is protected with up-to-date encryption methods during transmission and in the environments in which it is stored, that access is restricted on a role and authorization basis, and that data belonging to different customers is logically or physically segregated. Data access is recorded together with user, service, time, and transaction information according to the nature of the operation performed.

Training, validation, test, and production use data are separated from one another to the extent possible. The use of production environment data for the purposes of model training, fine-tuning, or retraining is carried out taking into account the project scope, the data use authorization, and the relevant approval processes. In test and development environments, the use of masked, de-identified, or synthetic data is preferred instead of real personal data.

Where third-party models or services are used, the scope of the data to be transferred to the model is technically restricted. The service provider's data retention, model training, logging, and data processing conditions are assessed. Where possible, configurations that prevent customer data from being used by the service provider for model development purposes are preferred.

Only the data necessary for error analysis, traceability, and operational monitoring is retained in input and output logs. Personal or sensitive data in log records is masked, de-identified, or removed prior to recording where necessary. Access rights to logs are restricted, and records are deleted or irreversibly de-identified at the end of the specified retention period.

Data sources, the transformations applied, access rights, data transfer points, and retention periods are recorded. When the purpose of use of a data source ceases to exist or its retention period expires, processes are operated for the deletion of the relevant data from active systems, from temporary storage areas, and, to the extent applicable, from backups.

2.5. Transparency and Explainability

Next4biz takes care to ensure that the intended purpose of use, the basic manner of operation, and the known limitations of AI systems are understandable to the relevant parties.

To the extent required by the use case, the information on which the class, score, recommendation, or response produced by the system is based, the confidence level, the influential variables, the decision rules, and the relevant system records may be made available to the user or to the authorized review unit.

2.6. Security and Resilience

AI systems are developed on the basis of a secure design and layered defense approach. Security risks such as unauthorized access, data leakage, malicious or manipulative inputs, the neutralization of system instructions, the creation of unauthorized transactions, the use of services outside their intended purpose, and the like are assessed during the design, development, and operation stages of the system.

In this context, technical controls such as authentication, role- and authorization-based access control, segregation of customer data, input and output validation, secure configuration management, protection of secret keys, transaction limitation, request rate control, the keeping of security logs, and the monitoring of anomalous use may be applied.

The output produced by the AI model is not accepted directly as a trusted or authorized command. Model outputs are validated at the application layer in terms of authorization, parameters, data types, business rules, and transaction scope before any transaction is executed in external systems. For transactions that are irreversible, that give rise to financial consequences, or that create significant business impact, additional validation or human approval may be applied.

It is essential that systems transition to a safe state in the event of an error, an outage, a security suspicion, or uncertainty. In such cases, it may be ensured that the system does not produce a transaction, suppresses the output, provides a limited response, informs the user, records the transaction, or transfers the process to an authorized person.

The effectiveness of the security controls applied is evaluated, according to the nature of the system, by means of security tests, access control tests, malicious input scenarios, API tests, vulnerability scans, and controlled attack simulations. Where security vulnerabilities are identified, they are remediated on a prioritized basis according to their risk levels.

2.7. Accountability

For each AI system, a business owner, a technical owner, and, where necessary, approval authorities are designated. The responsibilities relating to the development, deployment to production, modification, and monitoring of the system are defined.

Where a result produced by AI is subsequently modified by business rules, threshold values, or user actions, the model output and the final business outcome are monitored separately from one another to the extent possible.

3. Governance and Responsibilities

3.1. AI Governance

AI systems developed within Next4biz or integrated into its products are evaluated in terms of their intended purposes of use and information security criteria.

During the development of a new AI system, the significant modification of an existing system, or its introduction into a new area of use, the technical, operational, information security and, where necessary, legal implications are evaluated.

Use cases assessed as carrying significant risk are reviewed with the participation of the relevant product and R&D units, the information security committee, legal counsel and management.

3.2. Team Responsibilities

The project team ensures that the intended purpose of use of the system and customer expectations are clearly defined.

The R&D and software teams ensure the implementation of data, model, technical controls, tests, versioning and monitoring mechanisms.

The infrastructure team evaluates the security requirements relating to the architecture and data flow of the system. The legal and compliance units provide their opinion on the applicable obligations in cases where the use case so requires.

The customer, in addition to conveying its project expectations in full, ensures the accuracy, currency and authorization for use of the content it provides to the system, as well as the correct and complete communication of customer-specific business rules.

4. AI Project Lifecycle

4.1. Use Cases and Assessment

Before an AI system is developed or deployed in a new use case, the purpose of the system, its target users, the data to be used, the results it is expected to produce, and the impacts of possible failure conditions are assessed.

The assessment takes into account matters such as whether the system provides recommendations, makes automated decisions, communicates directly with the customer, or initiates a transaction in an external system.

At the assessment stage, controls such as human intervention and approval, confidence thresholds, withholding a response, a preliminary review pool, content filters, transaction limits, or additional monitoring are applied.

4.2. Data Management

The suitability of training, validation, test, and production data for the system is assessed. The source of the data, its currency, its scope, and the extent to which it is representative of the intended purpose of use are taken into account.

It is acknowledged that significant changes in the data distribution may affect system performance. Data that has become outdated or that does not represent current customer behavior is not used for model development or performance validation purposes without the necessary assessments having been carried out.

Information and documents provided by the customer are processed solely within the scope of the relevant project and contract.

4.3. Testing and Acceptance

AI systems are subjected to functional, technical, and scenario-based testing before being placed into production.

Testing evaluates not only the expected correct results, but also misclassification, inappropriate responses, deviation from the subject matter, ambiguous input, incomplete information, unauthorized transactions, and situations in which the system should not provide a response.

Where necessary, customer end-user testing is conducted for customer-specific configurations. The decision to go live is taken with due regard to the assessment results.

4.4. Change Management

Significant changes made to the model, knowledge source, system instructions, threshold values, business rules, or integrations are implemented in a controlled manner.

Where a change may significantly affect system behavior, the relevant tests are repeated. Where necessary, the ability to revert to the previous version or to disable the relevant feature is provided.

5. Specific Principles Applicable to Next4biz AI Systems

5.1. Classification and Routing Systems

In classification and routing systems offered by Next4biz, such as Issue Intelligence and PRIME, class definitions, category hierarchies, confidence scores, decision thresholds and workflow rules are determined according to the use case. Care is taken to ensure that the classes are distinguishable from one another, operationally meaningful and capable of being represented by a sufficient number of examples. Classes that overlap with one another, that are ambiguous or that have an insufficient number of examples are examined separately during the data preparation and model development phases.

In systems such as PRIME, in which ambiguous records are routed to a preliminary review pool, the pool is treated as a separate and valid system output. Records sent to the preliminary review pool are not automatically deemed to be incorrect predictions. System performance is measured across all defined classes, namely active, passive and the preliminary review pool, and by means of an evaluation method appropriate to the use case.

The initial result proposed by the model, the confidence score, the alternative classes, the applied threshold value, the business rule and the final routing are recorded separately from one another. In systems operating asynchronously, a transaction or process identifier is generated for each prediction, thereby ensuring that the content ultimately saved by the user is matched with the relevant model prediction. This prevents interim predictions produced before the text has been completed from being treated as the final system result, and prevents model performance from being confused with the result arising at the application layer.

During live use, class distributions, confidence score distributions, corrections made by humans, the rate of routing to the preliminary review pool and requests for new categories are monitored. The change between the training data and the live data may be assessed by means of distribution comparisons and data drift analyses. In the event of a decline in performance in certain classes, the emergence of new forms of expression or a change in the category structure, the dataset, the threshold values, the class definitions and the need to retrain the model are reviewed.

5.2. Dialogue Systems

Generative AI systems offered by Next4biz, such as AI-Bot, Auto Responder and similar systems, are configured in line with the intended purpose of use, the target users, the communication channels, the information sources, the customer rules and the operations the system is permitted to perform, as defined within the scope of the project.

At the outset of the project, an AI-Bot Solution Design Document is prepared together with the customer. This document determines the topics on which the AI-Bot will provide responses, the areas to be excluded from scope, the information sources to be used, the target user profiles, the language and tone of communication, the conditions for transfer to an agent, prohibited responses, product recommendation rules, sensitive topics, external service calls and operations requiring human approval. Customer-specific rules are implemented through system instructions, information retrieval filters, business rules, output controls and integration authorizations.

In order for the system to be able to generate accurate and up-to-date responses concerning the customer's processes and products, the necessary information, documents and content must be provided by the customer. Such content may include product and service descriptions, frequently asked questions, user manuals, procedures, campaign information, product catalogues, business rules, prohibited topics, contractual limitations and situations that must be transferred to an agent. The accuracy, currency, consistency and authorization for use of the content provided by the customer are the responsibility of the customer.

Rules are defined such as not recommending products designated by the customer, not providing advice regarding certain product components, not generating directive responses in sensitive areas such as health, law, finance or similar fields, not comparing particular brands or products, and not answering questions in the nature of personal advice such as "which product is most suitable for me".

Where the system determines that a user input relates to a topic on which responses are prohibited or which requires human assessment, it may generate a standard informational message, suppress the response or transfer the conversation to an agent. In such cases, the reason why the system did not generate a response may be recorded in association with the identifier of the relevant customer policy or business rule.

User inputs and system instructions are separated from one another as a safeguard against prompt injection, jailbreak attempts and requests aimed at rendering system instructions ineffective. Content provided by users is not accepted as a trusted system instruction or as an authorized operation command. In order to ensure that instructions contained within texts retrieved from information sources do not alter model behaviour, the content and instruction layers are separated, known attack patterns are examined through input controls, and the precedence of system instructions is preserved.

Model inputs and outputs may be checked in terms of personal or sensitive data. Data that is not necessary for the intended purpose of use is masked, de-identified or removed before being transferred to the model. Access filters and output controls are applied in order to prevent the model from disclosing system instructions, access keys or unauthorized document content to the user.

Before the system is put into live operation, in addition to normal use scenarios, it is tested with off-topic questions, incomplete information, contradictory documents, prohibited product recommendations, requests for health advice, prompt injection, inputs containing personal data, questions whose answers are not present in the source, and unauthorized operation requests. In line with the test results, the system instructions, information retrieval thresholds, the document chunking method, the re-ranking settings, the output controls and the conditions for transfer to an agent are updated.

Additions, removals and updates made to the knowledge base are recorded together with version information. Content that has become outdated, that is contradictory or that the customer has notified as having been withdrawn from use is removed from the retrieval index. Following material changes to the model, the system instructions, the information retrieval method or the customer rules, the relevant test scenarios are re-run.

These controls are intended to reduce the risk of generating content that is outside the source, inappropriate, unauthorized or contrary to fact. Owing to the probabilistic nature of the model used, the scope and currency of the information provided by the customer, the nature of user inputs and the behaviour of third-party models, they do not constitute a guarantee that all errors will be eliminated under all circumstances.

5.3. Automated Response Systems

In systems in which an AI output is sent directly to the customer without agent review, additional verifications relating to the scope, content and communication rules determined by the customer are applied before the response is sent.

The topics, transactions and types of customer request to which automated response systems may respond within the scope of the project must be clearly defined within the scope of the project. For requests that are not included within the scope of automated responses, that have no adequate counterpart in the information source, that involve uncertainty, or that require personal assessment or expert opinion, the system may refrain from generating a response, may present the response as a draft, or may route the process to agent review.

Certain topics, such as health, law, finance, product advice, product comparison, campaign commitments, or price or stock verification, may be excluded from the scope of automated responses by the customer. Matters such as not recommending certain products or product components, not using certain expressions, providing only a standard informational message on certain topics, and the situations that must be transferred to an agent are defined within the customer rules.

These rules are implemented through topic and intent classification, information source scope controls, product and category metadata, confidence thresholds, lists of prohibited topics and expressions, output verification mechanisms and rule-based suppression controls.

In cases where the response is not found suitable for automated sending, the operation may be suppressed, a draft response may be created, additional information may be requested, or the record may be routed to human review.

The message length, format, form of address, prohibited content, use of links and reply rules specific to the email, web, WhatsApp, marketplace or other written communication channels through which the automated response is to be sent are incorporated into the system configuration upon the customer's request and approval.

5.4. Text Analysis, Sentiment and Quality Systems

The sentiment, quality, summary, topic and root cause results generated by systems offered by Next4biz, such as Sentiment analysis and InsightX, are of a decision-support nature.

The models and methods to be used in these systems are determined by taking into account the language and length of the text, the channel characteristics, the class structure and the intended purpose of use.

Model outputs are evaluated together with confidence scores, decision thresholds are determined on validation data, and class-based performance results are monitored.

The results of these systems must not be used as a conclusive determination in isolation and independently of context, particularly in assessments that may give rise to significant consequences for employees or customers.

In systems that perform summarization and root cause inference, care is taken to preserve the critical information contained in the source text, to ensure that the generated result is consistent with the source content, and to ensure that a cause not present in the source is not presented as a conclusive determination. Where necessary, summary, topic or root cause results are assessed by means of rule-based controls, a second model or user verification.

5.5. Anomaly (Outlier) Detection and Prediction Systems

The results generated by Anomaly.Net and similar anomaly detection systems offered by Next4biz are of an indicative nature with regard to unusual process instances that require examination. The flagging of a record as an anomaly does not, in itself, mean that an error, breach or misuse exists.

During live use, changes in the process structure, in anomaly scores and in time series distributions are monitored. Where new process types, changed user behaviours, sudden increases in traffic or events not present in historical data are detected, the model, the alarm thresholds, the prediction variables and the training data are re-assessed.

6. Customer-Specific Controls and Shared Responsibility

6.1. Customer Control List

In projects where the behavior of the AI system must be restricted in accordance with the customer's commercial, legal or operational requirements, a Customer AI Control List shall be prepared.

This list sets out the topics on which the system may and may not respond, the information sources it will use, the products that may or may not be recommended, the conditions for transfer to a human agent, the transactions requiring human approval, and the customer-specific communication rules.

6.2. Content Provided by the Customer

The customer is responsible for the accuracy, the currency and the authorization to use the information, documents, rules and product content that it provides to the system.

Any changes made to the content or any new restrictions must be notified to Next4biz and the necessary configuration work must be completed. The customer shall be informed that content which is no longer current may affect the system's results.

6.3. Limits of Use

AI systems must not be used outside their defined purpose of use. The necessary human oversight must be ensured.

The effects that may arise as a result of the customer modifying the system configuration, the information sources, the integrations or the security controls without the knowledge of Next4biz shall be assessed separately.

7. Monitoring, Logging and Incident Management

7.1. Monitoring of Live Systems

AI systems operating in the live environment are monitored through technically and operationally appropriate indicators.

The scope of monitoring may include indicators such as system performance, user corrections, transfers to agents, instances of non-response, service latencies and resource consumption.

Where a significant change in performance is observed, the infrastructure, data, model, knowledge source, threshold value or business rules may be reassessed.

7.2. Logging and Traceability

The transaction data (log records) generated by AI systems are recorded taking into account the nature of the relevant system and data protection requirements.

It is essential that only the data necessary for operational purposes be retained in log records and that sensitive data be protected. Log records do not contain personal data unless necessary, and where necessary such data is retained in masked form.

8. Third-Party Systems and Vendors

8.1. Use of Third-Party Models and Services

Third-party AI model services may be used in Next4biz solutions. In addition, open-source models, embedding models, re-ranking services, vector databases, content safety services, speech-to-text, text-to-speech, OCR, translation, observability and model monitoring components may also be incorporated into the system architecture.

Before a third-party service is put into use, the scope of the inputs and documents to be transferred to the service, whether the generated outputs are retained, whether the data is used by the provider for the purposes of model training or service development, the retention period of abuse monitoring records, the regions in which the data is processed and stored, sub-service providers and data deletion capabilities are examined. This assessment is carried out on the basis of the API endpoint, feature, account type and project configuration used.

Only the data necessary for the performance of the relevant operation is transferred to third-party services. Personal, sensitive or commercial data that is not necessary for the purpose of use is masked, de-identified or excluded from the scope of the input prior to transfer. API keys and access credentials are not kept in application code or in model inputs; they are stored in secure secret key management systems, and access to service accounts is granted in accordance with the principle of least privilege.

Third-party model outputs are not directly accepted as reliable or verified information. Depending on the use case, model outputs are checked in terms of consistency with the source, scope, format, content safety, customer rules and transaction authorization.

In the event that the provider changes the model version, the security policy, the data processing terms, the API behavior or the usage limits, the impact of the change on the Next4biz solution is assessed. Where possible, fixed model versions are used, automatic version transitions are monitored, and acceptance tests are re-run following significant model changes.

In projects where dependency on a single model or provider constitutes a risk to service continuity, an alternative model, provider or local deployment option may be identified. In the event that the primary service becomes unavailable, its performance falls below acceptable limits, its cost changes significantly, or its data processing terms become incompatible with project requirements, a migration plan to an alternative service may be implemented.

The use of third-party services is recorded in the technical documentation together with the provider name, the model and version used, the API feature, the scope of data transfer, retention settings, deployment region, security configurations and the alternative service plan.

9. Training, Audit and Continuous Improvement

9.1. AI Awareness

Employees involved in the development, sale, implementation or operation of AI systems are ensured to possess knowledge of AI risks, data protection, information security and system limitations at a level commensurate with their roles.

Customer users may be provided with the necessary information regarding the manner of use and the limitations of the relevant product.

9.2. Review and Improvement

The controls applied to AI systems under this Framework are reviewed taking into account technological developments, customer feedback, performance results and changes in the relevant regulations.