Artificial Intelligence (AI)
Capabilities, Limitations and the Challenges of Responsible AI Adoption
AI has progressed from specialised research and experimental applications into a broad range of commercial, scientific, engineering and professional activities. The development of Generative AI, Large Language Models (LLMs) and increasingly capable AI agents is fundamentally changing how organizations analyse information, develop software, automate processes and support decision-making.
Skills is the first of these barriers. Business and IT leaders alike acknowledge that AI will change the skills needed to accomplish AI jobs. Currently, for example, AI can evaluate X-rays like human radiologists. Similarly, Machine Learning ML also has a learning curve.
However, understanding what AI can and cannot reliably achieve remains essential.
AI can accelerate complex tasks, identify patterns in substantial volumes of information, generate software and documentation, and assist specialists in exploring alternative solutions. It does not, however, automatically guarantee accuracy, factual correctness, sound reasoning or compliance with legal and regulatory requirements.
The successful application of AI requires a combination of technical capability, reliable data, appropriately designed processes, human expertise and independent verification.
From Machine Learning to Generative and Agentic AI
Artificial Intelligence encompasses several approaches, including conventional Machine Learning (ML), Natural Language Processing (NLP), computer vision, knowledge-based reasoning and Generative AI.
Traditional Machine Learning systems generally learn statistical relationships from training data to support classification, prediction or decision-making.
Generative AI systems, including LLMs, can produce text, software code, images and structured information. They can also assist with document examination, technical investigations, research and complex problem-solving.
A further development is Agentic AI, in which AI-enabled software can plan and execute sequences of actions, interact with applications, retrieve information and use external tools.
These capabilities introduce significant opportunities for automation. They also introduce new risks, particularly where an AI system is permitted to modify information, execute code, access sensitive systems [internal or outside], or initiate actions without adequate supervision.
The distinction between generating a plausible answer and producing a correct, independently verified result is therefore fundamental.
The Principal Challenges of AI Adoption
1. Skills, Expertise and Organizational Readiness
AI is changing the expertise required across business, engineering, healthcare, legal and professional services.
Organisations increasingly require personnel who understand not only how to operate AI systems, but also how to assess their limitations, evaluate outputs, recognise errors and determine when human intervention is necessary.
The effective implementation of AI requires collaboration between domain specialists, software engineers, data professionals, security practitioners and business management.
AI should be regarded as a capability that augments professional expertise rather than as an automatic substitute for it.
2. Reliability, Accuracy and Verification
One of the most important limitations of contemporary Generative AI is that a convincing output is not necessarily a correct output.
LLMs can generate inaccurate statements, misinterpret documents, produce defective software, invent references or arrive at conclusions that are unsupported by the available evidence. Such failures are often described as hallucinations, although errors may also arise from incomplete information, incorrect reasoning or unsuitable system design.
For business-critical applications, AI-generated outputs should be subject to proportionate verification.
Appropriate safeguards may include independent validation, automated testing, source verification, traceability, reproducibility assessments and human review.
In software and hardware engineering, mathematical techniques such as formal verification, theorem proving and model checking can provide stronger assurances about specified system properties than testing or AI-generated explanations alone.
AI can assist in discovering a solution, but the correctness of that solution must be established through appropriate evidence.
3. Data Quality, Security and Confidentiality
Data quality remains central to many successful AI applications, although contemporary foundation models have changed the relationship between training data and business deployment.
Organisations can increasingly use pretrained AI models and connect them to authorised internal information through techniques such as Retrieval-Augmented Generation (RAG), without developing a completely new model.
Nevertheless, incomplete, inaccurate, inconsistent or outdated information can significantly reduce the reliability of an AI-enabled system.
Businesses must also consider data provenance, access control, intellectual property, confidentiality, personal information and the risks of exposing sensitive material to external services.
Data volume is not the only or most important factor to consider, however. Many successful Use Cases UC can be achieved using a reasonable amount of data, provided that the dataset is good quality (i.e. normalized, complete and diversified). A lack of volume can be compensated by a reduction in project scope, while a lack of data quality will invariably lead to Proof of Concept PoC failure.
It is important to remember that “a reasonable amount of data” will mean different things, depending on the AI technique being considered. Probabilistic reasoning techniques such as Machine Learning rely heavily on data to deliver insights. This is where data quality problems are most acute.
The use of AI introduces additional security considerations, including prompt-injection, malicious source material, data leakage and inappropriate execution of automated actions.
4. Commercial Value and Return on Investment ROI
The widespread availability of AI does not mean that every implementation will create measurable business value.
Organizations should distinguish between successful demonstrations, Proofs of Concept (PoCs), operational deployments and systems that deliver sustained commercial benefits.
A successful AI project should have defined objectives, measurable performance indicators, realistic operating costs and appropriate acceptance criteria.
Potential benefits include improvements in productivity, software development, research, operational efficiency, knowledge management and customer services.
However, implementation costs must also account for integration, infrastructure, licensing, testing, security, supervision and ongoing maintenance.
AI adoption should therefore be evaluated against actual business outcomes rather than the perceived sophistication of the technology.
5. Governance, Accountability and Regulatory Compliance
As AI becomes embedded within business processes, governance and accountability become increasingly important.
Organizations must understand who is responsible for AI-generated decisions, how errors are detected, whether an outcome can be explained or challenged, and what safeguards apply when automated systems cause harm.
International existing legal obligations concerning data protection, intellectual property, confidentiality, discrimination, professional responsibility and information security may apply to AI-enabled activities.
International organizations must also consider applicable AI-specific regulatory requirements and technical standards.
Effective governance requires documented responsibilities, appropriate risk assessments, testing, monitoring, audit records and mechanisms for human intervention.
AI and the Future of Professional Work
The impact of AI on employment is more complex than a simple prediction of jobs created or eliminated.
Some repetitive and information-intensive activities are increasingly capable of automation. Other responsibilities are changing as professionals use AI to perform research, examine evidence, develop software and evaluate possible solutions.
New requirements are also emerging in AI systems engineering, independent assurance, security, evaluation, governance and technical oversight.
In professional and high-consequence environments, human judgement, accountability and specialist-expertise remain essential.
The practical objective should be to combine the speed and analytical capabilities of AI with the reliability, experience and critical judgement of appropriately-qualified professionals.
From AI Experimentation to Verified Deployment
The next phase of enterprise AI adoption will increasingly concern the transition from experimental systems to reliable, maintainable and independently assessable operational capabilities.
A sound implementation process should establish:
A clearly defined business or technical objective.
The suitability of the proposed AI architecture and underlying data.
Objective criteria for accuracy, performance, reliability and acceptable failure.
Appropriate data protection, security and governance controls.
Independent testing, validation and, where applicable, formal verification.
Human oversight, exception handling and accountability.
Ongoing monitoring of performance, operating costs and changes in system behaviour.
The central challenge is no longer simply whether AI can produce an answer, recommendation or action.
It is whether that result can be trusted, verified and relied upon for its intended purpose.
Successful organizations will be those that recognise both the considerable capabilities of modern AI and the engineering, commercial and professional disciplines necessary to deploy it responsibly.