As we enter 2025, generative AI is continuing to reshape how we communicate, solve problems, and interact with technology and each other. This is highlighted by significant advancements in Large Multimodal Models (LMMs) and the rapid growth of Agentic AI. These and other innovations promise to make AI faster, more capable, and more integrated into our daily lives while also continuing to raise important questions around trust, regulation, and implementation.
Multimodal models - systems that integrate text, vision, speech, and in some cases more - are extending the capabilities of AI workflows far beyond the largely text-only systems that dominated even a year ago. This is already obvious with the evolution of multimodal enterprise services offered by Google, OpenAI, and Anthropic. But Open Source is keeping up, with models like Alibaba’s QVQ-72B Preview and Meta's upcoming Llama 4 release focusing on "speech and reasoning", Open Source AI continues to democratize access, and foster innovation across industries.
Visual AI is also making significant strides. Meta’s Segment Anything Model (SAM) isolates visual elements with minimal input, enabling applications in video editing, research, and healthcare. Meanwhile, Carnegie Mellon and Apple’s ARMOR system, with its distributed depth sensors, has advanced robotic spatial awareness, reducing collisions by 63.7% and processing data 26 times faster than traditional methods.
Speech systems are advancing too. Models like Hertz and Kyutai’s Moshi achieve impressive response times - in some cases under 120 milliseconds - promising ever more natural interactions. Yet challenges persist: voice customization, context retention, and inference costs remain critical challenges.
Agentic AI represents a shift in how LLMs operate, granting them varying degrees of autonomy through controlled access to tools and workflows. Unlike traditional "AI Contact Center Agents," which primarily serve as conversational interfaces, Agentic AI systems exist along a continuum of integrative capabilities. This evolution allows them to solve real-world problems beyond their training data by interacting with external systems.
Agentic AI in the contact center can be broadly understood to operate along a spectrum of autonomy that might look like:
Low Agency (☆☆☆): LLMs generate text responses to user inputs.
Moderate Agency (★☆☆ - ★★☆): LLMs classify and route calls, retrieve customer data, or interact with tools like order lookup and FAQ retrieval.
High Agency (★★★): LLMs autonomously manage conversation flows, initiate or conclude interactions, and make real-time, objectives-based decisions.
This view of Agentic AI is supported by major players like HuggingFace, "Agency evolves on a continuous spectrum, as you give more or less power to the LLM in your workflow.", as well as industry analysts like Gartner, "AI Agency is a Spectrum." This spectrum highlights how Agentic AI can adapt to different business needs. For instance, within the unified Spitch platform, high-agency solutions like the Coaching Simulator, or Voice Assistants and chatbots excel in customizable natural responses and advanced call routing, while tool-integrated systems like Agent Assist and Speech Analytics perform complex behind-the-scenes automation for live interactions and post-call analyses.
In 2025, the rapid expansion of Agentic AI use cases—particularly in customer experience management—is set to redefine workflows, enabling businesses to strike a balance between automation, efficiency, and user experience.
Small Language Models (SLMs) are also trending. In some cases these models match the performance of larger systems like GPT-4 in targeted tasks while running on standard hardware. This shift addresses enterprise needs for scalable, cost-effective AI. On this front, Google Research and DeepMind have recently introduced a new approach to using SLMs as teachers to train much larger models.
Generative AI’s evolving impact is particularly evident in contact centers. According to a recent McKinsey report, this shift reflects three major changes sweeping the industry: contact centers are now seen as business opportunities rather than triage points, organizations are expanding their AI ecosystems to support these aspirations, and there is a growing focus on upskilling staff with better onboarding, training, and coaching.
Early adopters like Klarna have reported a 14% reduction in customer service costs, while AI-driven systems handle up to 80% of routine interactions. AI-powered training platforms like the one currently developed at Spitch are also transforming workforce development, providing real-time coaching and personalized simulations that reduce onboarding time and improve employee performance.
Despite these gains, adoption remains limited. Only 6.1% of U.S. companies implemented AI solutions in 2024. Challenges include integration with legacy systems, high training costs, and significant public distrust. According to Qualtrics, 75% of consumers remain skeptical of corporate AI use, an 11% decline from 2023.
Inference remains a bottleneck for widespread AI adoption. Companies like Groq and Cerebras are developing specialized hardware for real-time applications, achieving up to 70x faster inference speeds. Others, like Acurast, are leveraging distributed computing to run advanced models on commodity devices, offering scalable solutions for non-time-critical tasks. Meanwhile large providers like OpenAI are heavily discounting inference on their standard customer facing models.
With the EU AI Act now in force, Europe provides one of the most advanced regulatory frameworks for AI deployment, emphasizing technical robustness, transparency, and privacy. However, the challenge of evaluating compliance remains unresolved. Recent research from ETH Zurich highlights this gap, and revealed that none of today’s most widely used large language models (LLMs) fully meet the Act’s requirements. In 2025, it will remain an important priority to continue developing compliance benchmarks that align legislative interpretation with practical evaluation methods, ensuring both safety and innovation thrive together.
OpenAI cofounder Ilya Sutskever recently noted, we are approaching “Peak Data.” The availability of high-quality training datasets is dwindling, pushing the industry toward synthetic data generation and inference-time learning. These advancements will likely define AI's growth trajectory in the coming years, moving from pattern matching to true reasoning and adaptive learning.
Meanwhile, businesses need to continue prioritizing data transparency, ethical practices, and measurable customer benefits to overcome public skepticism and drive adoption.
This year looks set to surprise again, and promises accelerated adoption and practical integration of AI into business and consumer applications. Key trends to look out for include:
Multimodal Model Proliferation: Expanding applications across industries like healthcare, logistics, and customer service.
Agentic AI Evolution: Rapidly advancing high-agency model autonomy driving efficiency in workflows and problem-solving.
Enterprise Adoption: Broader implementation, with AI handling an ever increasing percentage of routine tasks in early-adopter organizations.
Cost Efficiency: Inference costs dropping below $10 per million tokens, making AI more accessible, and more applicable to increasingly complex problems.
The challenge for organizations is to balance innovation with trust, automation with customer satisfaction, and compliance with scalability. Success will come to those who navigate these tensions thoughtfully while delivering clear, measurable value - an endeavor where the strategic consulting provided by Spitch and its partner ecosystem plays a crucial role in helping businesses adapt to this rapidly changing landscape.
Generative AI is not merely a tool - it’s a transformative force. In 2025, businesses that embrace its potential responsibly will lead the next wave of human-machine collaboration.
