Artificial intelligence has sparked predictions of sweeping productivity gains, transformed business models and widespread job losses. But the commercial excitement surrounding AI is running ahead of what the technology can reliably deliver.
Richard Firth | image supplied
Today's AI systems are powerful tools, but they have significant technical and economic limitations. The assumption that continued investment will soon produce artificial general intelligence (AGI), or that AI can replace large sections of the workforce, deserves greater scrutiny.
The current AI boom could face a significant correction in the coming years. If it does, that would not mean the technology has failed. It could instead mark a shift away from inflated expectations towards practical applications that deliver measurable value.
The dot-com crash offers a useful comparison. The internet survived the collapse of the companies built around unrealistic expectations. AI could follow a similar path.
The limits of today's AI
Today's large language models (LLMs) generate responses by predicting the next token based on patterns in their training data and the context available to them. They can produce sophisticated results, but this does not mean they understand information or reason in the same way humans do.
Their limitations become particularly apparent in complex business environments, where systems depend on extensive data, interconnected processes and institutional knowledge.
An AI model can process large amounts of information, but it still operates within a finite context window. Even when that window accommodates hundreds of thousands of tokens, it may not capture all the dependencies within an enterprise system or software codebase.
The result can be inconsistent outputs, overlooked dependencies and confident but incorrect answers.
Techniques such as prompt engineering and retrieval-augmented generation (RAG), which supplies models with relevant external information, can improve performance. However, they do not eliminate the underlying limitations.
Meanwhile, simply increasing model size and computing power is becoming an increasingly expensive way to pursue better results. More parameters and larger training budgets do not automatically translate into more reliable intelligence.
Bigger models are not the whole answer
These constraints are helping to shift attention towards more efficient, specialised AI systems designed for particular tasks and environments.
This could prove more valuable to businesses than trying to apply one general-purpose model to every operational challenge.
Data will also play a central role. Organisations hold structured operational information that is not publicly available online, including records of transactions, processes and customer interactions. When appropriately governed and prepared, this data can help AI systems produce more relevant and useful outputs.
However, access to data alone is not enough. Its quality, structure, security and relevance all influence the results an AI system can deliver.
Businesses must therefore look beyond the model itself and consider the wider infrastructure required to make AI useful. That includes data management, integration with existing systems, ongoing maintenance and the cost of computing resources.
The risk of replacing people too soon
Predictions about AI-driven job losses often assume that current capabilities will advance quickly enough to automate entire roles. In practice, many jobs involve more than completing individual tasks. They require judgement, contextual understanding, accountability and the ability to respond when something goes wrong.
Software engineers and support specialists, for example, may use AI to accelerate parts of their work without being fully replaceable by it.
The risk for employers is that they restructure their workforces based on anticipated capabilities rather than demonstrated performance. Removing experienced employees before AI systems can reliably perform their work can also mean losing institutional knowledge that is difficult to recover.
Salesforce has become part of this wider debate about AI-led workforce changes and the practical limits of automation. More broadly, businesses experimenting with AI need to distinguish between automating selected tasks and replacing the people responsible for entire functions.
The economics also deserve closer examination. AI does not necessarily eliminate operational costs. It can introduce new expenses involving computing, implementation, oversight and integration, while existing systems and staffing requirements remain in place.
For some organisations, AI may initially add another layer of operational spending rather than reduce costs.
A more practical approach to AI
Businesses should approach AI as they would any other significant technology investment. They need to understand its limitations, assess its costs, test how it performs under real operating conditions and establish clear accountability when it fails.
Human oversight remains important wherever decisions require judgement or carry significant consequences.
The technology can still deliver meaningful benefits. The challenge is identifying where it improves productivity, reduces errors or helps people make better decisions, rather than adopting it simply because competitors are doing so.
The AI market may be entering a period in which expectations are tested against commercial reality. Some investments will not deliver the returns anticipated, and some applications will prove less useful than their advocates predicted.
That correction would not necessarily be a setback for AI. It could help separate viable applications from speculative projects and encourage businesses to focus on outcomes rather than promises.
When the hype subsides, the organisations best positioned to benefit will be those that have integrated AI into their operations where it makes sense, measured its results and recognised where human expertise remains essential.
