In the Small Firms Association’s 2026 survey [1], 92% of respondents reported using AI tools such as ChatGPT or Microsoft Copilot. For Irish businesses, the challenge already extends beyond getting access to the technology. Attaching a rocket to a racing car adds extraordinary power, but it does not improve the steering, strengthen the brakes or teach the driver how to handle it. AI presents a similar challenge: as it expands what we can do, we need to develop our ability to judge its work and keep its use within limits we understand. The survey’s findings on shortages of time and technical expertise make that a practical concern.
The opportunity is substantial. A small company can attempt analysis, compare proposals, write code or prepare a report that might previously have required outside help. Experienced staff can work faster, while people with less experience can attempt tasks that were previously beyond their reach. For a business with limited resources, that could mean pursuing an idea or improving a service it could not otherwise afford to explore.
But the ability to produce the work does not automatically bring the knowledge needed to assess it. An owner can obtain a convincing analysis in minutes and still be unable to tell whether it rests on a mistaken assumption. A junior employee can produce a polished report without understanding what it leaves out. As businesses make greater use of AI, they therefore need to develop people’s ability to question its work alongside their ability to generate it. That requires more than learning to write a good prompt.
Capability is getting cheaper
Generative AI can handle work that conventional automation struggled with because every rule does not have to be specified in advance. It can interpret an email, compare documents, summarise an ambiguous situation or suggest a response. That opens useful work to people and organisations that previously lacked the time, staff or expertise to attempt it, and it can make experienced people dramatically faster at work they already understand.
The difficulty is that competence does not transfer neatly from one task to another. Research involving 758 Boston Consulting Group consultants [2] found large gains in speed and quality on tasks that suited GPT-4, but worse performance on another task. The researchers described a “jagged technological frontier”: tasks that appear similar to us can sit on different sides of what the technology handles well.
Existing expertise gives people a starting point, much as an experienced driver brings skills to an unfamiliar car. An engineer reviewing AI-generated engineering advice can recognise omissions that a novice might miss. But knowing the work does not mean knowing how the AI will behave. The engineer still needs to discover where it is dependable, which questions expose its limitations and when to stop relying on it. We are all learning to handle this technology, although we bring very different levels of experience to the work we ask it to do.
Judgement becomes the bottleneck
The difference matters because producing an answer and establishing whether it deserves to be used can have very different costs. An experienced manager may assess a proposed rewrite in seconds. Confirming whether one technical product can safely replace another may require specifications, context and several independent checks. Both outputs may take the AI seconds to generate; the second may still require substantial human expertise before anyone can act on it.
This is more difficult than checking ordinary software output because poor AI work does not necessarily look poor. NIST identifies confidently generated false information as a characteristic risk of generative AI, including plausible explanations and citations that can themselves be wrong [3]. A polished answer is therefore weak evidence that the underlying work is sound. The practical value of an AI use case depends partly on whether a person can check the result reliably and at a reasonable cost.
Businesses also need to consider how people acquire that understanding. A junior employee who receives a finished analysis may have little opportunity to learn why one assumption is reasonable and another is not. Training should therefore include examining sources, comparing possible answers and explaining why the final recommendation is justified. Experienced staff need opportunities to examine failures and share what they learn. Reviewing AI-generated work is itself a skill that needs practice.
Delegation changes when AI can act
The consequences increase when AI moves from producing work to influencing or taking action. Drafting a customer reply, recommending a refund and issuing that refund are different levels of delegation. The underlying model may be the same, but each step reduces the distance between a mistaken output and a real-world consequence. A system can therefore be useful enough to advise long before it has earned the right to act.
That distinction matters because AI is increasingly being connected to business systems rather than used only as a standalone chat interface. Access to email, files, customer records or operational tools can make an assistant much more useful, but it also means that mistakes can travel further.
The sensible limit is therefore not simply what the model is technically capable of doing. It is what we can evaluate well enough to trust, and what we are prepared to let happen when the model is wrong. That may mean allowing AI to retrieve information without changing it, draft a recommendation without approving it, or suggest an action without executing it. As evidence about performance improves, those boundaries can move. They should move because we have learned more, not because the technology has become more impressive.
For Irish businesses, this makes staff development part of the investment in AI. Employers need to give people time to practise checking its work, discuss failures and recognise when they need specialist help. They also need controls that prevent a mistake from becoming an action the business cannot undo. We are still on L-plates with this technology, however experienced we may be in our own professions. As more power becomes available, the investment in learning, judgement and control needs to keep pace. That is what will allow businesses to make useful progress rather than simply move faster.
[1] SFA survey
