AI is a risk management project, not just an IT project
Artificial intelligence has become part of everyday business in almost every industry. Banks use it for credit scoring and fraud detection, insurers for claims assessment, manufacturers for predictive maintenance, and hospitals for diagnostic support. Each of these applications brings benefits, but also risks that are not always visible in the business case.
Experience shows that projects rarely fail because of poor technology alone. More often they fail because of unclear accountability, insufficient understanding of the model within business functions, or employees who accept AI outputs without checking them. That is why it is useful to look at AI risks in two groups.
Hard risks are measurable and technical: model errors, data quality, cyber threats, vendor dependency and regulatory non-compliance. They can be quantified, tested and covered by controls.
Soft risks are human and organisational: governance and accountability, culture, competencies, ethics, reputation and over-reliance on automation. They are harder to measure, but they are often the root cause of hard risks.
Where AI is used today
The most mature applications are in banking and the financial sector, but the same types of risk arise in every industry.
| Industry | Typical AI applications | Key risk |
|---|---|---|
| Banking | Credit scoring, fraud detection, AML monitoring, chatbots, document processing | Model risk, biased credit decisions, lack of explainability, legal, reputational, regulatory risk |
| Capital markets and funds | Algorithmic trading, portfolio management, robo-advisors | Model risk, systemic and market risk, reputational risk |
| Insurance | Pricing, claims assessment and handling, fraud detection | Model risk, pricing discrimination, exclusion of high-risk groups, legal risk, reputational risk |
| Leasing and microfinance | Automated approval, asset valuation | Data risk, model risk, consumer protection non-compliance, regulatory, legal risk |
| Telecoms and retail | Personalised offers, dynamic pricing, customer support | Manipulative practices, data protection, legal risk, reputational risk |
| Healthcare | Medical image analysis, diagnostic support, triage | Patient safety, liability for errors, legal risk |
| Manufacturing and energy | Predictive maintenance, quality control, grid management | Operational and safety risk to critical infrastructure |
| Human resources (all industries) | Candidate screening, performance evaluation | Model risk, discrimination, high-risk under the AI, reputational risk Act |
| IT and outsourcing | Developing AI solutions for EU clients | Provider obligations under the AI Act |
This article is intended for risk managers, compliance functions, internal audit and management boards of organisations in Bosnia and Herzegovina and the region that are introducing or developing AI solutions. Most examples come from the financial sector, where regulatory practice is the most developed.
Regulatory framework
The EU AI Act is not only a matter for organisations in the EU. Its application is extraterritorial, similar to the GDPR, so it also binds companies from Bosnia and Herzegovina and the region that place their AI solutions or services on the EU market.
EU AI Act (Regulation (EU) 2024/1689). The Regulation classifies AI systems by risk: prohibited practices, high-risk systems, limited-risk systems and minimal-risk systems. High-risk systems (Annex III) include, among others, creditworthiness assessment of natural persons, risk assessment and pricing in life and health insurance, recruitment and evaluation of employees, access to education and management of critical infrastructure. Prohibited practices and the AI literacy obligation for staff have applied since February 2025, and the rules for general-purpose AI (GPAI) models since August 2025.
Who is in scope outside the EU. The AI Act applies to:
- providers that place an AI system on the market or put it into service in the EU, irrespective of where they are established;
- providers and deployers in third countries where the output produced by the AI system is used in the EU.
This is particularly important for BiH because of its strong IT and outsourcing sector. A software company from Sarajevo or Banja Luka that develops a credit scoring model for a bank in Germany, or a candidate screening tool for a client in Austria, has the obligations of a provider of a high-risk system. These include a risk management system, technical documentation, data governance, conformity assessment and the appointment of an authorised representative in the EU. Fines reach EUR 15 million or 3% of total worldwide annual turnover for breaches of high-risk obligations, and EUR 35 million or 7% for prohibited practices.
AI Omnibus (Regulation (EU) 2026/1744). It entered into force on 27 July 2026 and moved the application of obligations for stand-alone high-risk systems under Annex III from 2 August 2026 to 2 December 2027. For AI embedded in products covered by EU product safety legislation (medical devices, machinery), the deadline is 2 August 2028. Transparency obligations, such as informing users that they are interacting with an AI system, were not deferred and have applied since 2 August 2026. The deferral is not a pause: the remaining 13 months should be used for inventory, classification and documentation of systems.
Sector-specific rules in the financial sector.
- DORA (Regulation (EU) 2022/2554) has applied since 17 January 2025 to banks, insurers, investment firms and other financial entities. AI systems are ICT systems and are subject to the ICT risk framework, resilience testing, incident reporting and ICT third-party risk management.
- EBA Guidelines on loan origination and monitoring (EBA/GL/2020/06) require institutions to understand and document the automated models used for creditworthiness assessment, including their limitations.
- EIOPA has set out expectations for the insurance sector on the governance of AI systems, with an emphasis on fairness, explainability and human oversight.
Horizontal regulation. The GDPR (Article 22) gives individuals the right not to be subject to a decision based solely on automated processing that significantly affects them. For critical infrastructure in energy, transport and healthcare, the NIS2 Directive is relevant, and for medical devices the Medical Devices Regulation (MDR).
BiH and the region. Bosnia and Herzegovina currently has no specific law on artificial intelligence. AI systems are nevertheless covered by existing rules: in the financial sector by the regulations of the FBA and ABRS on governance, risk management, information systems and outsourcing, and by the rules of the insurance agencies; across all sectors by personal data protection and consumer protection legislation. In practice, the regulator or supervisor will assess AI through the questions it already asks: who is accountable, how the model is validated, how the vendor is managed and how customer data is protected. Croatia, as an EU Member State, applies the AI Act directly, while organisations in Serbia and Montenegro, especially those in EU groups or with EU clients, follow the same framework.
Hard risks: measurable, technical and regulatory
Hard risks can be identified, quantified and controlled with existing risk management tools. What sets them apart in AI systems is speed: an error that takes months to develop in a traditional model can emerge overnight in an AI model.
- Model risk
AI models, especially machine learning models, introduce new forms of model risk:
- Lack of explainability (black box). A bank must be able to explain to the customer and the supervisor why a loan was declined, and an insurer why a premium is higher. A model that cannot do this is not acceptable for decisions with a material impact on individuals.
- Model drift. Changes in customer behaviour or the economic environment can quickly erode accuracy. Fraud detection models are particularly sensitive because fraudsters constantly change their patterns. Continuous monitoring is required, not just annual validation.
- Overfitting. The model performs very well on the development sample and poorly in production. Out-of-time and out-of-sample testing is essential.
- Hallucinations in generative AI. A chatbot can give incorrect information about an interest rate, policy terms or a regulation, with complete confidence. The organisation is accountable for what its chatbot tells a customer.
- Data risk
An AI model is only as good as the data it learns from. The key questions are completeness, accuracy, timeliness and data lineage. Historical data may contain bias against certain groups, which the model then reproduces and amplifies: a credit model learns historical discrimination, and a candidate screening model favours the profile of existing staff. In the financial sector, the BCBS 239 principles on risk data aggregation gain a new application here.
- ICT and cyber risk
AI systems open new attack vectors:
- Data poisoning: deliberately injecting distorted data into the training process.
- Prompt injection: manipulating a generative system to bypass restrictions or disclose confidential data.
- Data leakage: employees entering confidential customer data, patient data or trade secrets into public AI tools.
- Deepfake fraud: a fake voice or video of an executive used to approve payments or bypass KYC checks. This risk affects every organisation that makes payments, not just banks.
- Third-party and concentration risk
Most organisations in the region will not develop AI themselves; they will buy it or use it as a cloud service. This means dependency on a small number of global vendors, limited insight into the model, restricted audit rights and questions about data location. Contracts must cover access rights, audit, exit strategy and notification of model changes. In the financial sector this is required under DORA and domestic outsourcing rules; in other industries it is good practice.
- Legal and compliance risk
This includes non-compliance with the AI Act, the GDPR and consumer protection rules, discrimination in credit, insurance or employment decisions, copyright infringement by generative models, and unclear liability for damage caused by an AI system’s decision. For IT companies exporting AI solutions to the EU, a particular risk is becoming a provider of a high-risk system without having anticipated it in contracts and processes.
- Operational risk and business continuity
The more automated a process, the greater the impact of an outage. In manufacturing and energy, an error in an AI control system can also have physical consequences. The organisation must know how the process works when the AI system is unavailable and who makes decisions in that case.
Soft risks: people, culture and accountability
Soft risks are hard to quantify and rarely appear in the risk register. Yet they most often determine whether hard controls actually work.
- Governance and accountability
The most common problem is the question: whose system is this? IT develops it, the business function uses it, the vendor maintains it, and risk is supposed to control it. Without a clearly named owner of the AI system, accountability becomes diluted. The management board and supervisory board must understand AI well enough to ask the right questions and approve the risk appetite for AI.
- Skills gap
The market for AI and data science specialists in the region is small and competitive. The other side of the problem is less visible: risk managers, auditors and compliance officers often lack the knowledge needed to effectively challenge an AI model. This is why the AI Act introduces an AI literacy obligation for staff.
- Over-reliance on automation (automation bias)
When a system gives the right answer 95% of the time, people stop checking. Human oversight then exists only on paper: the credit officer signs off whatever the model proposed, the claims handler approves the assessment without review, and HR never looks at candidates the algorithm ranked low. This is one of the most dangerous soft risks because it creates a false sense of control.
- Shadow AI
Employees are already using public AI tools to write emails, analyse documents and summarise contracts, often without the organisation’s knowledge. Bans rarely work. What works is a clear policy, approved tools and training.
- Ethical risk and fairness
A model can be lawful and still unfair. Examples include credit models that indirectly, through postcode or type of employment, discriminate against certain groups, or insurance pricing models that penalise customers based on data with no real link to risk. The question is not only “may we?” but also “should we?”.
- Reputational risk
A single publicised chatbot error or discriminatory credit decision can damage customer trust more than any regulatory fine. In small markets such as BiH, such stories spread quickly.
- Resistance to change and culture
Employees may see AI as a threat to their jobs and passively sabotage or bypass it. Conversely, an “AI solves everything” culture leads to systems being introduced without a real business need and without a risk assessment.
How soft risks become hard losses
A soft risk is usually the cause, and a hard risk the consequence. Controls that cover only the technical side therefore leave the root of the problem untouched.
| Soft risk (cause) | Hard consequence | Key control | Regulatory basis |
|---|---|---|---|
| Unclear ownership of the AI system | Model not validated or monitored, drift goes undetected | Named owner and RACI for each system | AI Act (provider and deployer obligations); FBA/ABRS governance regulations |
| Skills gap in risk and audit | Weak model challenge, missed errors | AI literacy programme, independent validation | AI Act, Art. 4 (AI literacy) |
| Automation bias | Human oversight exists only on paper | Measuring the rate of human overrides, test cases | AI Act, Art. 14; GDPR, Art. 22 |
| Shadow AI | Leakage of confidential data and trade secrets | Usage policy, approved tools, DLP | GDPR; DORA (financial sector) |
| Ethical negligence | Discriminatory credit, insurance or employment decisions, litigation | Fairness testing before and after deployment | AI Act (Annex III); EBA/GL/2020/06 |
| Pressure for rapid rollout | Inadequate vendor contract, lock-in | Due diligence and exit strategy before signing | DORA; outsourcing regulations |
| Unrecognised exporter role | Provider obligations for a high-risk system without prepared documentation | AI Act role assessment for every contract with an EU client | AI Act, Art. 2 and Art. 22 (authorised representative) |
The takeaway for risk managers: for every system, the AI risk register should record the human cause, not just the technical symptom.
Practical recommendations for risk management
Managing AI risk does not require an entirely new framework. It requires extending existing frameworks for model risk, ICT risk and outsourcing, with a clear focus on people.
- Build an inventory of AI systems. Include purchased solutions, features embedded in existing software and tools that employees use on their own initiative.
- Determine your role under the AI Act. Are you a provider, deployer, importer or distributor? IT companies developing AI for EU clients are usually providers and need to know this before signing a contract.
- Classify systems by risk. Use the AI Act logic: prohibited, high-risk, limited risk, minimal risk. Treat credit scoring, risk assessment in life and health insurance, and AI in recruitment as high-risk, regardless of whether the AI Act applies to the organisation directly.
- Integrate AI into the model risk management framework. Development, independent validation, monitoring of performance and drift, and periodic revalidation.
- Define human oversight that actually works. Specify when a person must intervene and measure how often they actually do.
- Strengthen vendor management. Audit rights, model transparency, notification of changes and an exit strategy must be in the contract.
- Introduce an AI usage policy. What is permitted, which data must never be entered into public tools, and which tools are approved.
- Invest in AI literacy. From the management board to employees who use AI every day, with deeper knowledge for risk, compliance and internal audit.
- Include AI in the risk register and reporting. Management should regularly receive an overview of AI systems, their performance and incidents.
Checklist before putting an AI system into production
Checklist before putting an AI system into production
0 / 12Conclusion
Successful implementation of AI systems in any organisation depends as much on the quality of the model as on the quality of the people overseeing it. Hard risks are addressed through validation, controls and contracts. Soft risks are addressed through clear accountability, knowledge and a culture that allows the question “is the model right?” to be asked.
For organisations in BiH and the region, this is a moment of opportunity. The financial sector has the most developed practice and can serve as a model for other industries. IT companies exporting to the EU are already in scope of the AI Act, and compliance is becoming a competitive advantage for them. Organisations that build an AI risk governance framework now will not have to rush when the regulator or supervisor starts asking questions, and they will be able to show customers and partners that their technology is both reliable and fair.
Sources
- EU Digital Omnibus on AI Enters Into Force, The National Law Review (K&L Gates), 31. juli 2026
- EU AI Act: AI Omnibus formally adopted, Addleshaw Goddard, 1. juli 2026
- Uredba (EU) 2024/1689 (AI Act), uključujući član 2 o području primjene;
- Uredba (EU) 2022/2554 (DORA);
- EBA/GL/2020/06;
- GDPR, prema objavljenim tekstovima propisa
Prepared with AI assistance
