The year kicked off with a breakthrough coming out of China with Deepseek, which seriously dented the US’ ambitions to dominate the market for large language models (LLMs). What Deepseek showed the world, beyond causing a big hiccup in the financial market valuation of the US AI tech giants, is that China, which the US had attempted to undermine by only allowing the export of less powerful AI acceleration hardware, was able to produce a model that could outperform US LLMs that benefited from the most powerful Nvidia chips.
Its significance goes beyond geopolitics: Deepseek’s R1 model demonstrated that it is certainly not necessary to throw vast amounts of computational resources and spend a huge amount of money on AI acceleration hardware to achieve good results. The financial results from the hyperscalers show that the trend is to invest heavily in gigawatt datacentres, which they anticipate will be needed to support the most powerful AI acceleration hardware.
But for everyone else, including corporate IT, such infrastructure is certainly overkill, especially as smaller AI models are able to combine the expertise of the public LLMs with more focused training to deliver outcomes that can outperform the major players when deployed within a business context.
Agentic AI has become the most hyped technology trend of 2025. Corporate IT and business leaders are having to deal with the aftermath of a feeding frenzy among enterprise technology providers to sell AI-enabled products.
There are numerous reports showing that AI is delivering a low return on investment (ROI); most projects are failing to get past the pilot stage and yet more AI is embedded into corporate IT systems. This has meant that while a corporate AI strategy may have been based on standardising on a few AI engines, every piece of enterprise software is being sold with standalone AI capabilities.
Given the poor ROI being achieved by the majority of corporate AI projects, the industry has pivoted towards agentic AI to join the dots between the enterprise AI systems that have been bolted onto commercial enterprise software. The goal is to drive up efficiency by enabling disparate AI systems to act as specialist AI workers that have been tuned to handle specific parts of a business workflow.
The question then becomes what happens to the parts of the workflow that need to be done by a human worker. It is this interface between workers and AI systems that is now receiving a lot of attention. If AI is being sold to improve efficiency, then at some point, people’s jobs will change and some may find they are surplus to requirements. Those who remain in employment will have AI agents as co-workers.
Business leaders are pondering how to balance human work with tasks that can be achieved easily by AI agents. Rather than being mere digital tools, there are discussions being had that look at treating an AI agent as a resource that improves over time and gains experience through training aka machine learning. There will be societal ramifications as agentic AI moves beyond hype to something that can do useful work in an organisation.
Here are Computer Weekly’s top 10 AI stories of 2025.
If used correctly, large language models (LLMs) promise to revolutionise software development – but they do not easily fit some corporate IT use cases, with the vagaries of natural language raising some challenges. The majority of programs are written in English-like programming languages that are deterministic, which means the programmer effectively tells the computer exactly what it needs to do. However, using natural language in vibe coding can lead to problems when trying to describe something unambiguously.
The availability of the DeepSeek-R1 LLM shows it’s possible to deploy AI on modest hardware. Matthew Carrigan, a machine learning engineer at Hugging Face, suggested a system for running AI inference based on DeepSeek could be built using two AMD Epyc server processors and 768 Gbytes of fast memory. The system he demonstrated in a series of tweets could be put together for about $6,000.
The Ada Lovelace Institute examines how “market forces” can be used to drive the professionalisation of artificial intelligence assurance in the context of a wider political shift towards deregulation. It recommends that frameworks for AI regulation need to distinguish between AI systems generally and those used for narrower contexts, in terms of both the practical technical and legal competencies needed to assure each type of system, as well as the standards that should be applied to each.
The companies promoting AI fail to mention it’s often underpinned not by code but by humans tagging data and viewing unsavoury content – AI could not exist without cheap labour largely outsourced to the Global South. Then there is the “cloud”, which has a larger carbon footprint than the airline industry and is decidedly physical, as manifested in water-guzzling datacentres and extractive mining in environmentally challenging locations.
We speak to Chris Loake, group CIO at Hiscox, about the roll-out of Microsoft Copilot and how to succeed with AI projects. For Loake, an AI strategy is like the North Star, which broadly stipulates an AI-enabled business.
“We believe that AI is a generational technology which will underpin many, many things,” he says.
The phrase “don’t believe the hype” has never been more apt – there are growing warnings of an AI investment bubble that could affect everyone if it bursts. For example, Thinking Machines Lab, an AI startup, recently raised $2bn funding on a valuation of $10bn – the company has zero products, zero customers and zero revenues. The only thing it made public to its investors was the resume of its founder, Mira Murati, formerly chief technology officer at OpenAI.
We find out how organisations can take automation to the next level using agentic artificial intelligence. Analyst firm Forrester uses the term “process orchestration” to describe the next level of automating business processes, using agentic AI in workflow to handle ambiguities far more easily than the programming scripts used in RPA.
AI job disruption was among the hot topics at the Gartner Symposium in Barcelona. We speak to Gartner’s Helen Poitevin about AI job chaos, with Poitevin stating that employees will see that certain tasks they do will start to go away. She recommends IT and business leaders take a human-first approach to design AI systems that people want to use to do their jobs more effectively.
We speak to security experts about how IT departments and security leaders can ensure they run artificial intelligence systems safely and securely. If you think of an AI model as a new employee who has just come into the company, do you give them access to everything? No, you don’t. You trust them gradually over time as they demonstrate the trust and capacity to do tasks.
Organisations are starting to look at where artificial intelligence fits into business workflows. IT leaders can get their organisations ready for workflows that may be split between internal staff, external contractors and AI agents by capturing the knowledge using structured data ontologies to make expertise machine readable.