Ones to watch
Knowledge Distillation (KD) should be positioned as BUILD CAPABILITY NOW:
A foundational technique required to operationalise modern AI, particularly for LLMs and edge deployment scenarios where cost, latency, and infrastructure constraints prohibit direct use of large models.
It consistently delivers near-teacher performance with significant model compression, enabling production-grade deployment across constrained and regulated environments.
Organisations should consider actively investing in KD pipelines, tooling, and governance integration, as it is becoming a core enabler of scalable, efficient, and explainable AI systems, with potential further gains when combined with pruning and quantisation
Explainable AI (XAI) should be also positioned as urgent, as it is no longer optional but a core requirement for regulatory compliance, auditability, and operational trust in AI systems. With AI increasingly embedded in decisions affecting clients, employees, and financial outcomes, organisations must be able to explain, justify, and defend model outputs to regulators, auditors, and stakeholders. XAI enables transparency through techniques such as feature attribution and post hoc explanation, but requires integration into the model lifecycle, not retrofitted as an afterthought. Building capability now is critical to support governance, fairness validation, and long-term AI adoption, especially in regulated and high-impact decision environments
Synthetic data
Synthetic data is moving from niche to foundational. We’re approaching a point where models can improve without fresh human‑labelled data, changing the economics of training and testing.
Embeddings
Embeddings are getting smaller (and cheaper) without losing capability. Perplexity’s pplx‑ embed models are designed for web‑ scale retrieval and natively produce compact INT8 or binary embeddings, reducing storage by 4×–32×—a meaningful signal for teams trying to scale retrieval and RAG without ballooning cost.
On-device AI
On‑device AI is becoming more practical for real workflows. Google’s AI Edge Gallery now supports on‑device function calling—with FunctionGemma (a compact 270M‑parameter model) translating natural‑language commands into executable app intents, enabling fast, fully offline actions.
Efficiency race
The efficiency race is shifting from models to inference systems. Techniques like DualPath target real serving bottlenecks (KV‑cache storage I/O), reporting throughput gains of up to 1.87× (offline) and ~1.96× (online) without violating SLOs—exactly the kind of “under the hood” improvement that changes production economics.
Agent tooling
Agent tooling is getting more structured and governable. Frameworks such as Helm are moving away from ad‑hoc automation towards typed, permissioned function calls with sandboxed execution—an important signal for organisations that want agent capability without losing control.
Tech Radar
AI is no longer limited by access to models, tools, or capability
Those barriers are coming down. What is emerging in their place is a different constraint—one that is less visible, but more decisive: data.
Across global organisations, the difference between successful AI programmes and those that stall is becoming clearer. It’s less about how quickly teams adopt new technologies, and more about the data decisions they make — structure, governance, and where it resides. These choices determine whether AI can be scaled reliably, operated safely, and trusted in critical workflows.
This isn’t a technical footnote; it’s a leadership responsibility. Decisions about data now shape operational resilience, regulatory exposure, and long-term value creation in ways that extend far beyond individual AI use cases.
"AI readiness starts with foundations: data, governance, and workforce augmentation. Only once those are in place, and organisations can safely automate workflows, does it make sense to reinvent core processes. AI tools that accelerate human work are not the same as AI tools trusted with autonomous delivery. Readiness comes from foundations, not from how fast you move at the peak."
Rosemary J Thomas, PhD
Senior Technical Researcher
The organisations progressing fastest are not chasing capability at the frontier
They are investing in the foundations that make that capability usable—treating data as a strategic asset, not a by-product of systems. The question for leaders is no longer whether to adopt AI, but whether their data is ready and whether they are in control of it. For those looking to strengthen these foundations, the next step is to invest in a robust data strategy and governance capabilities that underpin scalable, trusted AI.