Artificial intelligence (AI) is not one thing. It now appears in distinct forms: generative artificial intelligence (gen AI), agentic AI, multimodal AI, small language models (SLMs) and self-improving models. Each serves a specific function, from content generation to co-ordination of agents or processing diverse data types. Irish organisations are adopting different forms based on accuracy, control, and practical deployment.
Rory Timlin, data and AI practice lead at KPMG in Ireland, points to early success with gen AI. “In customer operations, gen AI enhances query understanding and response formulation,” he says. “It improves first-time resolution, prompts next-best actions and automates post-call summarisation.”
Beyond customer service, Timlin highlights uses such as “content generation, localisation, translation and document summarisation,” where point solutions are already active and producing measurable gains in speed and consistency.
Alessia Paccagnini, associate professor at UCD’s Smurfit Business School, highlights the benefits of SLMs in settings with limited computing power.
READ MORE
“SLMs offer faster processing speeds and lower computational costs,” she says.
These features make SLMs especially suitable for edge deployment, that is, running locally on a device instead of relying on the cloud. Environments such as wearable health monitors, factory sensors and customer service kiosks are prime candidates for this type of model.
When it comes to adopting artificial intelligence, purpose must come first. For Ivan Jennings, Senior Manager of Solution Architecture at Red Hat, success begins with one simple principle: define what the model is for.
“It’s important to commence AI rollouts by defining a model’s purpose,” he says. “Just because a generative LLM model could do something doesn’t mean it should.”
Jennings points to Small Language Models (SLMs) as a powerful alternative to their larger, general-purpose counterparts. These models are compact, task-focused, and optimised for specific domains such as fraud detection, healthcare diagnostics, and operational efficiency. Critically, they offer strong performance while using significantly fewer computational resources.
SLMs help enterprises reduce power consumption, heat generation and operational costs. More importantly, they support integration with private, secured datasets, lowering exposure to third-party risks and enabling compliance with industry regulations.
“SLMs prioritise what truly matters for enterprises,” says Jennings. Their smaller size makes them easier to run in hybrid cloud environments, whether on-premises, in private clouds or at the edge, bringing AI directly to where it delivers the most value.
Stephen Redmond, head of data analytics and AI at BearingPoint, describes how SLMs function in enterprise systems.
“SLMs are easier to manage, customise and control,” he says, also stressing their role in structured orchestration. “Each one performs a specific step. For example, summarising, classifying, or routing. You can inspect the model weights, set rules and adapt prompts which is essential for enterprise use.” Redmond underlines the need for transparency in environments that demand compliance and operational traceability.
Open source models such as Mistral 7B, LLaMA 3 and IBM’s Granite family are already being fine-tuned for real-world use in finance, healthcare and customer service. Jennings singles out the Granite models as especially useful:
“They’re open source-assured under the Apache 2.0 license. That means developers can experiment with, modify, and distribute Granite models for free,” says Jennings.
Transparency matters. “When we are using AI in our business, we must be able to somewhat understand how the AI model came up with an answer or decision,” he says. Knowing what data a model was trained on helps enterprises build confidence especially when AI agents may act autonomously.
Ultimately, Jennings believes the future of enterprise AI isn’t just about bigger models. It’s about “scaling intelligently, and implementing AI solutions that are efficient, secure, and sustainable within the enterprise ecosystem.”
Multimodal AI brings a different capability: the ability to process more than one type of input simultaneously.
“These systems process and combine multiple types of data, text, images, audio, and video, for more integrated tasks,” says Paccagnini. She identifies high-potential areas such as education, customer support, healthcare and media.
Emmanuel Adeleke, partner in technology and transformation at Deloitte Ireland, agrees: “They integrate and process diverse data types enabling a richer and more nuanced understanding of complex scenarios.”
In academic and business contexts, multimodal AI is being explored for its ability to unlock insights that single-mode models cannot reach.
Adeleke and Paccagnini also focus on self-improving models. These systems refine their performance over time without explicit reprogramming.
“Self-improving AI models continuously learn from data and feedback,” Adeleke says. “They enable personalised experiences and process optimisation, especially in dynamic environments.”
This continual learning loop is valuable in sectors such as logistics, risk management and finance, where conditions change rapidly.
Paccagnini says: “Self-improving AI models can adapt and learn continuously from new data or user feedback, leading to improved performance over time without human intervention”.
She lists practical applications such as recommendation systems, fraud detection, predictive maintenance and dynamic pricing strategies. These use cases rely on the system’s ability to adjust, respond and improve based on patterns it encounters in real-world usage.
Richard Blythman, founder of NapthaAI, an Irish AI start-up, focuses on agentic AI where systems composed of decentralised agents work together to complete tasks.
“There is a lot of confusion over agentic AI, generative AI and multimodal AI,” he says. “Many terms have been introduced very recently and often get bundled together. You can have a generative agent or an agent that performs tasks that are not generative.”
Blythman’s point is that agentic AI is not defined by content output but by autonomous co-ordination and task execution.
At NapthaAI, he is building what he describes as “an internet of agents.” These agents are designed to operate across devices and co-operate through a networked structure.
“Agents run on many devices, often geographically dispersed, yet co-operate with each other across the network.”
Blythman cites education as an example: “The tutor agent might work together with coding agents making it a simple example of a multi-agent system,” he says. Each agent has a clear role and operates as part of a broader structure, rather than relying on a centralised model.
Each AI type plays a different role. Generative AI handles content creation, agentic AI manages co-ordination and multimodal AI brings together diverse inputs. Self-improving models evolve based on use. SLMs bring these capabilities closer to the user with lower cost and greater customisation. The experts agree that modularity and control are emerging as defining principles in this next phase of deployment.
“Each one performs a specific step,” says Jennings. In tightly governed organisations, the ability to inspect, modify and isolate each stage of an AI pipeline is non-negotiable. Timlin notes that even initial deployments are delivering clear operational benefits. Adeleke and Paccagnini also emphasise the long-term adaptability and autonomy, while Blythman points to the growing importance of co-operation between agents.
These are not speculative developments. Across public and private sectors, models are already being embedded, not to replace full workflows but to enhance parts of them. The idea of one model to do everything is being replaced by precise orchestration: each tool performing one function well, integrated into a system that is observable, tunable and evolving.
The strength of this approach lies in clarity. Businesses now choose AI tools based not on novelty but on alignment – does the model fit the task, can it be governed, and will it adapt? AI is not just changing. It is specialising.
