{"id":9213,"date":"2024-12-18T09:50:47","date_gmt":"2024-12-18T09:50:47","guid":{"rendered":"https:\/\/weareoakland.com\/?p=9213"},"modified":"2025-01-15T09:33:53","modified_gmt":"2025-01-15T09:33:53","slug":"the-future-of-ai-is-ai-agents","status":"publish","type":"post","link":"https:\/\/weareoakland.com\/blog\/the-future-of-ai-is-ai-agents\/","title":{"rendered":"Agentic AI and AI Agents: Shaping the Market in 2025"},"content":{"rendered":"\n

What a difference a year makes! As 2024 comes to a close, and for everyone in the world of AI, it\u2019s been a blast! It’s been a pivotal moment in the evolution of artificial intelligence<\/span><\/a>, particularly generative AI. The year has been defined by a number of major milestones in deploying large language models (LLMs) into production. As we finish the year, we can\u2019t help but feel we had our fingers right on the pulse when we launched our intelligent agent solutions, which seems like years ago. <\/p>\n\n\n\n

We start 2025 with a new term, but the same premise Agentic AI and AI Agents <\/strong>are where the market is heading. Autonomous systems or \u2018AI agents\u2019 perform all of those mundane (and increasingly not so mundane tasks that we never get around to doing). If 2024 was about figuring out how to take models into production, 2025 will be the year when we see agents emerging as first-class citizens in enterprise environments.<\/p>\n\n\n\n

\n
Get in touch<\/span><\/a><\/div>\n<\/div>\n\n\n\n

Where we started: The AI landscape at the beginning of 2024<\/h2>\n\n\n\n

At the outset of 2024, the AI world was still buzzing from the release of GPT-4, which solidified OpenAI\u2019s leadership in the generative AI space. GPT-3.5 Turbo was being phased out as GPT-4 became the new standard. In the open-source arena, Meta’s LLaMA 2<\/strong> was gaining traction, offering businesses an alternative to closed, proprietary models. Meanwhile, Anthropic<\/strong> continued to develop its Claude series, and Google\u2019s Gemini 1.0<\/strong> arrived, signalling serious competition in the race to dominate AI capabilities.<\/p>\n\n\n\n

At the same time, Retrieval-Augmented Generation (RAG)<\/strong> started to gain serious traction in enterprise contexts. This approach\u2014which combines large language models with external data retrieval\u2014became the default architecture for deploying AI systems that required domain-specific knowledge. Tools like Azure OpenAI\u2019s implementation of RAG and Google\u2019s NotebookLM<\/strong> began offering practical solutions for integrating business data into AI systems. By the end of the year, deploying an AI solution without integrating it with your organisation\u2019s data became unthinkable.<\/p>\n\n\n\n

Production the key focus of 2024<\/h2>\n\n\n\n

While 2023 was the year of experimentation, 2024 was the year of making AI operational. Enterprises moved beyond proof-of-concepts (PoCs) to attempt production-grade implementations of LLMs. Yet, this process proved anything but smooth. The Gartner statistic that 85% of generative AI projects fail to make it into production<\/strong> rang true, illustrating the unique challenges that this technology presents.<\/p>\n\n\n\n

The Production Problem<\/h3>\n\n\n\n

Deploying LLMs into production is fundamentally different from traditional software. In 2024, businesses grappled with challenges such as:<\/p>\n\n\n\n

Cost considerations<\/h4>\n\n\n\n

Dedicated AI budgets, in the main, are few and far between. Data teams are being challenged to do more with less. Add to this the double whammy that most AI projects are far more expensive than initially projected. Gartner reported that many projects cost 300 to 500 times more<\/strong> than anticipated, due in large part to infrastructure, security, and integration costs.<\/p>\n\n\n\n

Data integration<\/h4>\n\n\n\n

LLMs need access to enterprise data to add value. Whether via APIs, RAG pipelines, or custom integrations, aligning AI capabilities with business data became a priority.<\/p>\n\n\n\n

Infrastructure limitations<\/h4>\n\n\n\n

AI requires massive compute resources. Many organisations found themselves constrained by global hardware shortages. For example, in the UK, models have to be deployed in Sweden because the necessary GPUs are simply not available locally.<\/p>\n\n\n\n

Security and governance<\/h4>\n\n\n\n

The biggest challenge we see and one that the biggest tech vendors in the globe are wrestling with. Moving your enterprise data into AI workflows raises serious concerns around data sovereignty and governance. Even Microsoft\u2019s Copilot could not guarantee that enterprise data would stay within a specific geographic region; an issue that created significant friction for regulated industries.<\/p>\n\n\n\n

The SaaS boom<\/h2>\n\n\n\n

To address these production challenges, 2024 saw an explosion of Software-as-a-Service (SaaS) offerings targeting AI deployment. These tools aimed to simplify everything from RAG implementation<\/strong> to software development augmentation<\/strong>. For instance:<\/p>\n\n\n\n

Copilot<\/strong> emerged as the dominant enterprise AI tool, thanks to Microsoft\u2019s vertically integrated stack.<\/p>\n\n\n\n

Third-party solutions<\/strong> like Cursor<\/strong>, Devin<\/strong>, and Windsurf<\/strong> tackled specific pain points for developers, automating workflows and improving coding efficiency.<\/p>\n\n\n\n

Numerous RAG-focused SaaS platforms offered plug-and-play solutions for integrating enterprise data into LLMs, such as vector database providers like Pinecone and open-source alternatives like Weaviate.<\/p>\n\n\n\n

The result? Companies could adopt AI solutions faster, but they also became increasingly dependent on these third-party tools to abstract away the complexity of production deployment.<\/p>\n\n\n\n

The Physical Infrastructure Bottleneck<\/h2>\n\n\n\n

A major theme of 2024 was the physical infrastructure constraints facing AI. As global demand for compute power skyrocketed, manufacturers like TSMC<\/strong> and NVIDIA<\/strong> struggled to produce enough GPUs to keep up. Cloud providers like Azure<\/strong>, AWS<\/strong>, and Google Cloud<\/strong> were unable to provision sufficient hardware in all regions, forcing businesses to look for creative solutions.<\/p>\n\n\n\n

This scarcity brought back challenges reminiscent of the 1990s on-premise era: organisations had to think carefully about where their workloads would run and how physical infrastructure limitations would impact their deployments. A system that worked seamlessly in Sweden, for example, might be impossible to deploy in the UK. As AI models become increasingly powerful, ensuring access to hardware remains a critical factor for enterprise success.<\/p>\n\n\n\n