1 The move from generalised AI agents to domain-specific AI agents
General-purpose models trained on public internet data still struggle with the messy reality of enterprise processes because they lack deep organisational context. Moreover, in today’s regulatory and geopolitical climate, enterprises face growing demands for data and AI sovereignty – ensuring data privacy, security, and compliance within their specific jurisdictions and business environments.
Domain-specific agents, grounded in proprietary data with governed lineage, not only interpret internal rules, edge cases, and compliance constraints far more accurately but also uphold critical sovereignty requirements. This control over data and AI models reduces risk, meets legal and ethical obligations, and preserves competitive advantage.
We’ve already seen this play out. Suncorp Australia accelerated their AI experimentation and boosted model accuracy by pairing a flexible agentic AI architecture with robust, enterprise-grade governance and monitoring. The lesson is clear: companies poised to succeed in the next AI wave will invest less in model size and more in data quality, domain depth, secure integration, and strong data and AI governance frameworks aligned with sovereignty demands.
2 The move from single agent to multi-agent orchestration
Enterprise work rarely happens in a single step, and neither will enterprise AI. Real workflows span retrieval, validation, approvals, and decisions across multiple systems and teams – far beyond what a lone agent can reliably handle. The next phase is multi-agent orchestration, where specialised agents handle tasks such as compliance checks, data retrieval, or reasoning, while a supervising agent coordinates them.
Introducing a supervising agent sequences roles, delegates work, and synthesises results in natural language, enabling organisations to scale AI beyond isolated pilots and into governed, auditable, adaptable workflows.
3 The move from one-off checks to continuous evaluation
As AI moves into production, continuous, real-time evaluation becomes non‑negotiable. Models that look strong in training often degrade on live data or drift as inputs change, and reliability erodes quickly without constant evaluation. The coming year will see enterprises adopt evaluation‑centric practices, where agents are continuously measured against real tasks, real feedback, and changing conditions.
Agent Bricks is built around this principle: it streamlines the development of domain-specific agents, lets teams define purpose and quality criteria in natural language, and automatically generates test suites and optimises performance based on enterprise data. By creating an environment where AI evaluates AI, enterprises can reduce uncertainty, accelerate deployment, and ensure agents continually learn from successes and failures to better fit their specific needs.
4 The move from text to multimodality
AI has traditionally been text-first, but both consumers and enterprises now communicate through a mix of voice notes, videos, screenshots, sensor feeds, and chat messages. Multimodal AI matches this reality by understanding and combining these diverse inputs, dramatically expanding what automation can do in real operations.
In practice, multimodal workflows augment human interpretation at scale. A customer service AI agent can read a user’s message, analyse their tone of voice, and interpret screenshots or videos of the issue. In healthcare, models can fuse patient records, medical images, and sensor data to support more precise diagnoses and personalised treatment plans. In retail and e-commerce, multimodal agents can process reviews, product images, and usage videos to better understand customer preferences, improve recommendations, and spot fraud.
5 The move from AI as a feature to invisible integration
The most successful AI systems don’t announce themselves. They disappear into workflows, quietly improving productivity without creating friction for employees or customers.
Invisible AI means that automation is embedded, consistent and intuitive. It becomes the environment teams operate within rather than a feature they must learn how to use. When systems are evaluated continuously, humans and AI can work together seamlessly in partnership and work accelerates.
6 A continued focus on skills
As AI agents become embedded in day-to-day operations, organisations will need to keep investing in their people. This includes teaching them how to manage, guide, and collaborate with these systems, not just build them. You don’t need to be a data professional to benefit: a marketer automating data entry, for example, mainly needs the prompting and workflow skills to direct an AI agent to take that work over.
With over 20 years of experience in the information technology and services industry, Adam Beavis is a seasoned leader and innovator in the cloud computing and open-source software domains. He is currently the Vice President and Country Manager at Databricks, the data and AI company that helps organizations solve the world's toughest problems.
