For organisations considering AI agents, the useful questions are practical. What work can the system carry out? What information does it need? When should a person step in? And how will the organisation know whether it is helping?
The Agentic AI Innovation of the Year category at the AI Awards 2026 offers four examples: Future Proof Knowledge, Oac Snacks, SPRYT and yieldHUB. Their shortlisted work spans care administration, business planning, patient appointments and semiconductor testing. Explore the official finalists list.
What is agentic AI?
An AI agent can use tools and information to work through a task, selecting its next steps within the permissions it has been given.
The degree of independence matters. Anthropic distinguishes between workflows, which follow predefined paths, and agents, which can dynamically direct their processes and tool use. It recommends starting with the simplest solution that meets the need, because additional autonomy can bring additional cost and complexity. Read Anthropic’s guide to building effective agents.
For a business evaluating a system, the label “agentic” is only the beginning. The important detail is what the system can actually do.
Meet the four finalists
These summaries reflect the published shortlist descriptions, rather than an independent assessment of the products.
Future Proof Knowledge
Future Proof Knowledge applies agents to regulated care administration, including deadlines, reauthorisation, training and billing. Its description emphasises patient context, verified information and traceable oversight.
Oac Snacks
Oac Snacks uses multiple agents to support commercial, supply and finance planning. The described approach assigns distinct roles, challenges assumptions and retains human approval.
SPRYT
SPRYT’s Asa handles appointment bookings, changes, reminders and patient questions through messaging channels. Its description combines conversational AI with predictive analytics to support attendance.
yieldHUB
yieldHUB Live applies agents to semiconductor test data. Its stated focus includes monitoring performance, supporting operational decisions and identifying opportunities to reduce unnecessary retesting.
Different applications need different tests
Our analysis is that these applications should be assessed against the work they support, rather than a single broad claim about AI productivity.
For a care-administration workflow, a useful evaluation would examine whether tasks are completed accurately and on time, and whether staff can trace an action back to its source. For a planning tool, the test should include whether recommendations expose assumptions and help decision-makers consider alternatives.
For appointment coordination, an evaluation could measure successful bookings, unresolved requests and the work passed back to staff. For a manufacturing application, a buyer would need to examine performance against existing testing practices, including errors, unnecessary repeat work and the cost of operating the system.
These are proposed evaluation questions, not reported results from the finalists. They illustrate why a useful pilot starts with a specific business process and a baseline.
Mark Kelly: start with the task and the boundaries
“The first question for a leadership team is which task an AI agent should be trusted to carry out. Start with a clear business problem, define where human approval is needed, and measure the result before expanding.”
Mark Kelly, Founder of AI Ireland
For teams planning a pilot, we recommend agreeing three things before deployment:
- The task: define the process, the information required and the problem being addressed.
- The permissions: distinguish between reading information, recommending an action and carrying it out. Identify where approval or escalation is required.
- The evidence: record a baseline and measure the whole process, including the time people spend checking and correcting outputs.
Expanding an agent’s responsibilities should follow evidence from that bounded pilot. A clear record of exceptions and failures is as useful as a list of successful tasks.
What the shortlist can tell us
The shortlist identifies applications worth exploring. It does not provide a common performance benchmark or independently verified results. Buyers should request relevant evaluation evidence, understand what human work remains and test suitability in their own setting.
Editorial disclosure: AI Ireland organises the AI Awards. This article draws on the published finalist summaries and adds AI Ireland’s editorial analysis. Shortlisting is not independent verification of performance claims.
Explore the AI Awards 2026 finalists, hear practitioners discuss their work on the AI Ireland Podcast, or explore practical adoption with the AI Leaders Lab.
See practical AI in action
Join us at AI Awards 2026 to celebrate the organisations and people putting AI to work.
Want to bring these lessons to your organisation?
Discover more from AI Ireland
Subscribe to get the latest posts sent to your email.
