AI in 2025: is it an agentic year?

2024 was the GenAI year. With new and more performant LLMs and a higher number of projects rolled out to production, adoption of GenAI doubled compared to the previous year (source: Gartner). In the same report, organizations answered that they are using AI in more than one part of their business, with 65% of respondents mentioning they use GenAI in one function.

Yet, 2024 wasn’t just a year of incredible GenAI results, it was also a year of challenges – security, ease of use and simplified operations remain core obstacles that organizations still need to address. So what do we foresee in the AI space this year? Where is the community going to channel their energy? Let’s take a quick look at 2024 and then dive into the expectations for AI in 2025.

Sponsored

class=”wp-block-heading”>AI in 2024 at a glance

At the beginning of last year, we said 2024 was the year of AI, and it’s safe to say we were right – admittedly, it was a pretty safe prediction. In the last 12 months, in Canonical we focused on enabling AI/ML workflows in production, to help organizations scale their efforts and productize their projects. We worked on an array of AI projects with our partners, including Nvidia, Microsoft, AWS, Intel, and Dell, for instance integrating NVIDIA NIMs and Charmed Kubeflow to ease the deployment of GenAI applications. 

We also made good on our commitment to always work closely with the community. We had 2 releases of Charmed Kubeflow almost at the same time as the upstream project, running beta programs so the innovators could get early access. As the difficulty of operations is still a challenge for most Kubeflow users, hindering adoption in enterprises, we’ve been working on a deployment model that only takes a few clicks. You can sign up for it here, if you haven’t already.

Retrieval Augmented Generation (RAG) is one of the most common GenAI use cases that has been prioritized by a large number of organizations. Open source tools such as OpenSearch, Kserve and Kubeflow are crucial for these use cases. Canonical’s enablement of Charmed Opensearch and Intel AVX is just an example of how our OpenSearch distribution can run on a variety of silicon from different vendors, accelerating adoption of RAG in production. In the case of highly sensitive data, confidential computing unblocks enterprises and helps them move forward with their efforts. During our webinar, together with Ijlal and Michelle, we approached the topic, covering some of the key considerations, benefits and most common use cases. 

Is 2025 the year of AI agents?

As for 2025, one of the hottest topics so far is AI agents. These are systems that can independently perform self-determined tasks, interacting with the environment to reach pre-determined tasks. NVIDIA’s CEO, Jensen Huang, declared that “AI agents are going to be deployed” (source), signaling a higher interest in the topic and a shift from generic GenAI applications to specific use cases that organizations would like to prioritize.

Enterprises will be able to quickly adopt AI agents within their business function, but that will not solve or address all the expectations that AI/ML has created. AI agents will still face many of the same challenges that industry has been trying to overcome for some time:

• Security: whether we talk about the models, infrastructure or devices where AI Agents run, ensuring security will be critical to enabling organizations to roll them out to production and satisfy audits.
• Integrations: the AI/ML landscape is overall scattered and the agentic space is no exception. Building an end-to-end stack that enables not only the use of different wrappers, but also provides fine-tuning or optimized use of the available resources is still a challenge.
• Guardrails: the risk of AI agents is mostly related to the misleading actions that they can influence. Therefore, organizations need to build guardrails to protect any production-grade environment from putting them at risk.
• Operations: whereas experimentation is a low hanging fruit, running any AI project in production comes with an operational overhead,which enterprises need to simplify in order to scale their innovations.

Security: at the heart of AI projects

Let’s drill down into that security challenge. According to Orca, 62% of organizations deployed AI packages that had at least one vulnerability. As AI adoption grows in production, security of the tooling, data and models is equally important. 

Whether we talk about the containers that organizations use to build their AI infrastructure, or the end-to-end solution, security maintenance of the packages used remains a priority in 2025. Reducing the number of vulnerabilities is turning into a mandatory task for anyone who would like to roll-out their projects in production. For enterprises which consider open source solutions, subscriptions such as Ubuntu Pro are suitable, since they secure a large variety of packages that are helpful for AI/ML, including Python, Numpy and MLflow.

As the industry evolves, confidential computing will also grow in adoption, both on the public clouds and on-prem. Running ML workloads that use highly sensitive data is expected to become a more common use case, which will challenge AI providers to enable their MLOps platforms to run on confidential machines.

AI at the edge

Interest in Edge AI continues to rise. With a growing number of models running in production and being deployed to edge devices, this part of the ML lifecycle is expected to grow. The benefits of AI at the edge are clear to almost everyone, yet in 2025 organizations will need to address some of the common challenges in order to move forward, including network connectivity, device size and security.

Deploying AI at the edge also introduces challenges around model maintenance that go beyond model packaging. Organizations are looking for solutions that support delta updates, auto-rollback in case of failure, as well as versioning management. 2025 will see accelerated adoption of edge AI and an increase in the footprint of models running on a wider variety of silicon hardware.

Canonical in the AI space: what is 2025 going to look like?

Canonical’s promise to provide securely designed open source software continues in 2025. Beyond the different artifacts that we already have in our offering, such as Opensearch, Kubeflow and MLflow, we’ve significantly expanded our ability to help our customers and partners in a bespoke way. Everything LTS will help organizations secure their open source container images for different applications, including edge AI and GenAI. 

Throughout 2025, you can expect to see plenty more thought leadership spanning all things AI/ML. We’ll keep sharing best practices based on our customers’ and our own experience. And we will continue to pioneer in providing a fully open-source end-to-end solution that helps organizations run AI at scale.

Learn more about our solutions for edge AI and genAI infrastructure.

If you are curious where to start or want to accelerate your ML journey, optimize resource usage and elevate your in-house expertise, our MLOps workshop can help you design AI infrastructure for any use case. Spend 5 days on site with Canonical experts who will help upskill your team and solve the most pressing problems related to MLOps and AI/ML. Learn more here or get in touch with our team: https://ubuntu.com/ai/mlops-workshop