Five AI trends to watch in 2025

AI continues to be the number one topic in tech, especially in boardrooms and VC circles. By Q3 2024, EY reported that AI companies made up 37% of all VC investments. But it’s not just about the money—on the employee side, 93% believe AI is positively impacting their work and even helping reduce workplace anxiety, according to data from The Access Group.

Throughout the tech sector, AI is becoming the engine behind thorny issues such as data management and security and leaders are taking a keen interest in the technology to meet their organizational goals. But as the AI landscape matures,

Jeff Watkins, CTO at CreateFuture, warns that the inflated valuations in the AI market won’t last forever. When the bubble bursts, many of the “chancers” will exit, leaving behind only those ready to focus on cost efficiency and tighten their belts. At the same time, some CIOs are still excitedly waiting for the “Ford moment in IT,” when automation and productivity truly take off.

2025 might finally be the year we start getting answers. For now, though, here’s a quick rundown of the AI trends in IT that are likely to make waves over the next 12 months.

1. IT adopts small language models (SLMs)

Steven Webb, chief technology & innovation officer at Capgemini UK, predicts that we’re about to see a shift from large language models like GPT to smaller, more nimble models—small language models (SLMs). These SLMs are cheaper to run, need less computing power, water, and energy, and can be tailored to understand industry-specific jargon, pain-points, and technical language. SLMs are quickly gaining traction with 24% of organizations are already using them, while 54% planning to jump on board within the next few years.

This year has already seen some interesting developments in SLMs. Nvidia launched NIM, its own version of SLMs for easy enterprise deployment. At the same time, IBM introduced Granite 3.0 SLMs, which could actually outperform LLMs like Llama for tasks such as game design, drug discovery, or code generation due to use of a limited 3-4 billion parameters. This is far fewer than the tens or many hundreds of billions used by LLMs. Salesforce Research too is preparing to roll out SLMs that can operate offline, grounded in the data stored on user device to deliver a more tailored, cost-efficient experience.

As the pressure around data security, privacy, and sovereignty increases, the key to managing these challenges will be improving how we track data lineage and restrict access to only what’s necessary. Right now, that’s a limitation of LLMs, but SLMs are expected to fill that gap soon enough.

2. Shadow AI will continue to grow but AI will also step up to block threats

Shadow AI—often called “bring your own AI”—is when employees use unapproved AI tools in the workplace without IT or security approval. Salesforce reports that more than half of generative AI users rely on unapproved tools, and 7 in 10 workers globally have no training on how to use these tools safely or ethically. Unsurprisingly, this trend isn’t slowing down.

“The rise of Shadow AI wil create new attack vectors for security tems to fight and detect,” says Nicole Carignan, VP of strategic cyber AI at Darktrace. She expects 2025 to be marked by an explosion of AI and generative AI tools, spreading across enterprises and even employees’ personal devices. Her biggest concern, however, is the financial, reputational, and technological consequences businesses might encounter as regulations like the EU AI Act begin to take effect.

Shadow AI was at the center of Spotify’s decision to block GPT usage by employees and remove “tens of thousands” of Boomy-created tracks. According to the Financial Times, the move aimed to protect artists’ royalties after suspicious streaming patterns were flagged. Similarly, Samsung banned GPT after employees mistakenly uploaded sensitive code to ChatGPT, risking insider attacks.

Nicole argues that curbing these issues requires two things: continuous AI-powered asset discovery in real-time and better management-level training. How these solutions will shape 2025 is anyone’s guess. Jeff adds that AI won’t just be a part of the problem—it will also be part of the solution. He points to Protective AI, which is already in use but will take huge leaps forward when combined with agentic AI in 2025.

Protective AI helps by analyzing patterns in data, network traffic, and user behavior to detect threats like phishing, malware, ransomware, and data breaches in real time. Once a threat is identified, Protective AI can immediately act—blocking malicious IPs, shutting down compromised accounts, or isolating infected systems. It can also notify security teams for manual intervention if necessary.

3. Agentic AI will be the start of AI 3.0

Ann Maya, EMEA CTO at Boomi, is bullish on the future of Agentic AI, and she has the data to back her optimism. “Think about an average expense report,” she shares. “An AI financial agent could handle everything—from error checks and policy compliance to flagging exceptions and managing approvals. It’s all possible with no human intervention, connecting departments seamlessly through APIs tied to sales systems, ERP platforms, HR software, and even deeper functions like entity matching in data.”

AI agents are designed to plan and make decisions on their own, even when working with real-time information. Unlike traditional AI systems like LLMs, which follow “fixed” paths, AI agents can dynamically adjust their actions based on training and context. As Ann explains, with the right safeguards – like compliance checks and failsafes – these agents could introduce a more decentralized AI model for businesses. For CIOs, this means better operational efficiency and unprecedented control over data quality, security, and timeliness through smarter, context-driven decisions.

Agentic AI won’t have an easy road to success, however. “The key is for businesses to set measurable goals for their AI agents, such as reducing costs or increasing efficiency, and to align those goals with their broader priorities,” Ann tells ITPro. “That’s what will make the difference.”

4. AI will modernize legacy infrastructure

Updating legacy technology often means dealing with the high costs and risks of doing a massive overhaul all at once, which is why it can be a tough pill for many companies to swallow. But there’s a smarter way forward—companies will increasingly use AI to make incremental upgrades that are more manageable, cost-effective, and low-risk.

“AI is no longer a question of ‘if,’ but ‘when,’” says Burley Kawasaki, global VP of strategy at Creatio, when talking about AI’s role in IT. He points out that global businesses are prioritizing legacy modernization, with spending expected to reach $3.4 trillion by 2026.

According to Burley, businesses will start using a combination of AI and no-code platforms to update their legacy stack in the coming year. No-code tools can slowly build and integrate new features with older systems and keep things stable while workflows move to the no-code environment. Meanwhile, AI adds intelligence – like predictive analytics and decision-making capabilities – to make the process faster and smarter.

5. Multimodal systems will become the new “ChatGPT”

The number of multimodal AI models is increasing. Claude can now interpret images in PDFs and documents and ChatGPT is now capable of analyzing image prompts, even those embedded in files. These developments make one thing clear: multimodal systems are on the rise, and by 2025, they’ll be everywhere.

Multimodal AI brings together different data types – text, images, audio, video, and more – to create a richer understanding of information. Imagine a technician uploading a photo of an error screen and an AI model being able to provide text-based troubleshooting guides to resolve the issue.

Nasuni’s CIO of data intelligence and AI, Jim Liddle, sees multimodal AI becoming a dominant trend in enterprise data management. Its ability to interpret and visualize different data types without back-and-forth human input will help to cut costs, unify scattered data sources into a single strategy, and strengthen governance by keeping everything compliant with regulations.

While it’s hard to say exactly how these solutions will play out in 2025, Jeff sees it as a story of two halves: plenty of challenges, but also a lot of progress. “Expect some surprises,” he adds with a laugh, “but if I could predict them all, I’d probably be sailing on my megayacht right now.”