The great Yogi Berra once said, “It’s difficult to make predictions, especially about the future.” Truer words have never been spoken, especially in the rapidly evolving world of AI.
With the new Gartner 2024 Hype Cycle for Emerging Technologies providing a roadmap for the near future, it’s time to look ahead to 2025. Generative AI has crested the Peak of Inflated Expectations and moves ever closer to the Trough of Disillusionment. 2024 was a year of generative AI experimentation, and IT leaders are learning what should and shouldn’t go into production at scale based on the tangible reality of improved business outcomes.
Meanwhile, AI agents gain awareness as they move up the Innovation Trigger slope. For IT leaders who are unclear about the differences between generative AI and agentic AI, it can be broken down as generative AI creates, and agentic AI acts.
Humans partner closely with generative AI to generate and use outreach emails, source code, business plans, and marketing strategies, but the human is the one who activates the generative AI – and it’s also the human who acts upon the generative AI’s output.
In the case of agentic AI, agents can activate generative AI, as well as external workflows and systems of record. Agents can also figure out and execute the next steps, which may or may not require human intervention.
While the business value of generative AI may be difficult to assess, the benefit of agentic AI is more straightforward as there’s a human cost of “jobs to be done,” which can be compared with the cost of the job being done by an agent. Furthermore, agentic AI is elastic, which allows for surges of AI agents to cost-effectively scale up and down based on predictable seasonality, e.g. Black Friday sales, Tax Day support or unexpected events like major service outages or natural disasters.
5 AI predictions for 2025
The landscape of artificial intelligence is shifting rapidly, with new advancements and trends emerging constantly. As we approach another year, it’s crucial for technology leaders to anticipate and prepare for these changes.
Here are five AI predictions that will equally excite and challenge technology leaders in 2025:
- AI agent success will hinge on integration
- Large action models (LAMs)
- Small language models (SLMs)
- Energy efficiency
- Multiple AI model rings
1. AI agent success will hinge on integration
For the past year, industry leaders have been repeating the mantra, “Your AI strategy is only as good as your data strategy.” This is true because if your data is trapped and inaccessible by your AI, your generative AI results will be generic and disappointing at best and harmfully hallucinated at worst. As our AI journey expands to the world of agents, we’ll quickly learn the corollary: “Your agent strategy is only as good as your integration strategy.”
Why is this the case? Just as a well-oiled machine relies on the seamless interplay of its parts, the success of AI agents will depend heavily on their integration within the broader technological ecosystem.
AI agents, with their ability to automate tasks and make decisions, are poised to revolutionize various aspects of business operations. However, their true potential can only be unlocked when seamlessly integrated with existing systems and data sources across the enterprise.
Imagine an AI agent tasked with customer service. To be truly effective, it needs access to customer data, order history, product information, shipping information, and even real-time inventory levels. Without this integration, the agent’s usefulness would be severely limited, resulting in fewer case deflections and higher support costs.
What IT leaders should do in 2025
- Prioritize integration: Don’t treat AI agents as standalone solutions. Instead, view them as integral components of your overall IT infrastructure.
- Invest in robust integration platforms: Explore and implement tools and technologies that facilitate seamless data exchange and communication between AI agents and existing systems.
- Break down data silos: Ensure that data is readily accessible to AI agents across different departments and functions.
By prioritizing integration, IT leaders can entrust AI agents to perform at their best and maximize business value for their organizations.
2. LAMs will be lions
Move over LLMs – there’s a new acronym in town. While large language models (LLMs) excel at understanding and generating human language, large action models (LAMs) take it further by translating that understanding into action.
Think of it this way: LLMs are like the brains of the operation, capable of processing information and generating insights. LAMs are the hands, putting those insights into action.
Why LAMs are critical to agentic AI success
LAMs are poised to be the driving force behind the success of agentic AI, which focuses on creating AI agents that can autonomously perform tasks and achieve goals. By integrating with APIs across the enterprise, LAMs can interact with various systems and applications, enabling them to carry out complex actions.
For example, an LAM-powered AI agent could automate the onboarding process for new employees. It could gather the necessary information, create accounts, grant access to relevant systems, and even schedule introductory meetings.
Integrating LAMs with APIs
The key to unlocking the full potential of LAMs lies in their integration with APIs. APIs (application programming interfaces) act as the connective tissue between different software applications, allowing them to communicate and exchange data.
By integrating LAMs with APIs, organizations can enable AI agents to interact with a wide range of cloud-based and on-premise systems and services, from CRM and ERP systems to marketing automation platforms and ecommerce platforms. This integration opens up a world of possibilities for automating tasks, streamlining processes, and improving efficiency.
What IT leaders should do in 2025
- Explore the potential of LAMs: Understand how LAMs differ from LLMs and how they can be used to increase agentic AI initiatives.
- Develop an API strategy: Ensure that your organization has a well-defined API strategy to facilitate the integration of LAMs with existing systems.
- Invest in API management tools: Implement tools that help you to effectively manage and monitor your APIs, ensuring seamless integration with LAMs.
By embracing LAMs and prioritizing their integration with APIs, IT leaders can position their organizations to reap the full benefits of agentic AI, resulting in new levels of automation and efficiency.
3. Small language models: Does size really matter?
While large language models have dominated the AI landscape with their impressive capabilities, a new trend is emerging: small language models (SLMs). These compact and efficient models are gaining traction as organizations recognize their svelte benefits.
SLMs are smaller versions of their LLM counterparts. They have significantly fewer parameters, which translates to several advantages:
- Cost-effectiveness: SLMs require less computational power and memory, making them more affordable to train and deploy.
- Speed and latency: Their smaller size enables faster processing and lower latency, making them ideal for real-time applications.
- Reduced footprint: SLMs can be deployed on smaller devices and edge computing environments, extending their reach to resource-constrained settings.
- Enhanced privacy: SLMs can do all processing on-device, minimizing the need to send sensitive data to a third party or across a network.
- Increased availability: SLMs can be deployed in environments with limited internet connectivity, making them suitable for remote or offline applications.
How SLMs are created
Researchers employ various techniques to create SLMs that maintain performance while reducing size:
- Teacher-student training: A larger “teacher” model guides the training of a smaller “student” model, efficiently transferring knowledge.
- Model pruning: Unnecessary parameters are removed from a larger model, streamlining its architecture.
- LLM quantization: Model parameters are represented with lower precision, reducing memory requirements.
Public sector benefits
SLMs hold particular promise for the public sector, especially for applications that require on-device operation. This is crucial for forward-deployed scenarios or classified applications running on systems without internet access to cloud-based frontier large language models.
What IT leaders should do in 2025
- Evaluate SLMs for specific use cases: Identify tasks and applications where SLMs can provide sufficient performance with greater efficiency.
- Improve SLM quality with integration: Transform and harmonize organizational data in consistent formats that SLMs can readily understand and easily consume.
- Monitor, optimize, and govern SLM adoption: Use API management to manage and govern SLM access and monitor their use to benchmark performance against other SLMs and cloud-based frontier LLMs.
Small is big. By embracing the “thin is in” approach and leveraging the advantages of SLMs, IT leaders can optimize their AI initiatives for cost-effectiveness, speed, privacy, and accessibility.
4. Energy efficiency will be as important as performance
As generative AI enters the mainstream, its environmental impact is coming under increased scrutiny. Organizations realize that the widespread use of large language models (LLMs) can significantly affect their environmental sustainability metrics and goals.
The energy consumption of these powerful models is a growing problem. Training and running LLMs requires vast computational power, translating to a substantial carbon footprint. This has prompted a call for greater transparency and accountability regarding the environmental costs of AI.
Efforts toward transparency
Fortunately, efforts are underway to address this issue. Researchers and industry leaders are working to develop an “Energy Star” equivalent for AI models, providing clear metrics to assess their environmental impact. This will enable organizations to make informed decisions about their chosen models, balancing performance with sustainability considerations.
Optimizing LLM selection
Instead of defaulting to the most advanced and resource-intensive frontier models for every application, organizations will increasingly prioritize the optimal LLM for their specific needs. Just as they consider factors like price and performance, they’ll also factor in the model’s environmental impact.
This shift toward energy efficiency will help to encourage innovation and the development of more sustainable AI models. Researchers will focus on creating models that deliver high performance with a lower carbon footprint, using techniques like model compression, quantization, and efficient training methods.
What IT leaders should do in 2025
- Assess the environmental impact of AI initiatives: Evaluate the energy consumption and carbon footprint of your current AI models.
- Prioritize energy-efficient models: When selecting LLMs, consider their environmental impact alongside performance and cost.
- Support sustainable AI development: Encourage using energy-efficient training methods and model optimization techniques.
- Advocate for transparency: Demand clear and standardized metrics for assessing the environmental impact of AI models.
By prioritizing energy efficiency alongside performance, IT leaders can ensure that their AI initiatives align with their organization’s sustainability goals and contribute to a greener future.
5. Multiple AI model rings will rule them all
As organizations increasingly rely on large language models (LLMs) to power their AI applications, efficient and cost-effective LLM management becomes paramount. This is where LLM routers act as intelligent traffic directors for LLM requests.
LLM routers optimize LLM usage by analyzing incoming prompts and routing them to the most suitable model based on complexity, performance requirements, and cost. This ensures that less expensive models handle simpler tasks while more complex tasks are directed to the most powerful models.
RouteLLM: An example of LLM routing in action
One example of an LLM router is RouteLLM, an open-source framework designed to optimize LLM serving. RouteLLM analyzes incoming queries and dynamically selects the best model to handle each request, considering factors like cost and performance. This approach not only reduces costs by up to 85% while maintaining 95% of the performance of frontier models but can also improve overall efficiency and response times.
Expanding LLM routers’ role
While current LLM routers primarily focus on optimizing for price and performance, their potential extends far beyond these dimensions. As awareness of the environmental impact of LLMs grows, LLM routers can play a crucial role in promoting sustainability.
Imagine an LLM router that considers not only price and performance but also the environmental impact of each model. This would enable organizations to optimize their LLM usage for sustainability, selecting the most energy-efficient model for each prompt.
What IT leaders should do in 2025
- Explore LLM routing solutions: Investigate the benefits of LLM routers for improving LLM usage and reducing costs.
- Consider sustainability in LLM routing: Advocate for developing LLM routers that factor in environmental impact alongside price and performance.
- Integrate LLM routers into AI infrastructure: Implement LLM routers to manage and optimize your organization’s LLM deployments.
- Build in AI agility from the start: Adopt API-led connectivity so developers can use the best AI models and routers today and quickly and easily switch to even better ones tomorrow.
By embracing LLM routers and expanding their capabilities to include sustainability considerations, IT leaders can ensure that their AI initiatives are both efficient and environmentally responsible.
The AI future ain’t what it used to be
The future of AI is filled with exciting possibilities, and these five predictions offer a glimpse into the trends and challenges that await technology leaders in 2025. By staying informed and prepared, organizations can navigate this evolving landscape and harness the power of AI to drive innovation, achieve their business goals, and be the disruptor instead of the disrupted.
Want to explore more? We’ve got you covered.
- Assess your organization’s AI readiness: Take this short assessment to get a clear picture of your AI readiness and receive customized recommendations based on your results. These insights will help you integrate AI into your organization, enabling you to make informed and ethical decisions that boost customer engagement, employee productivity, and your bottom line.
- Learn more about the future of AI: Stay updated on the latest advancements and trends in generative AI and agents, including free hands-on training and certifications.
- Work with your MuleSoft team: Collaborate with your MuleSoft team to develop a comprehensive AI strategy that aligns with your business objectives.
- Embrace the future: Be proactive in exploring and adopting new AI technologies to gain a competitive edge.
Together, let’s navigate the future of AI and unlock its full potential for a brighter tomorrow. To keep on learning, download our 2025 digital trends report to see what seven trends are shaping digital transformation in the upcoming year.
