My focus is not on technology for its own sake; it is on P&L, efficiency, and strategic advantage. Right now, there is a clear paradox in the Enterprise space: every company is scrambling to build an “AI Centre of Excellence,” yet this centralised model is already a sinking ship. It is an overhead, not a lever.
The instinct to centralise expertise is understandable, but it is fundamentally at odds with the pace of innovation we see in 2024. The traditional ‘AI/ML Team,’ a siloed unit of highly specialised model builders, is about to become the single biggest bottleneck to true product velocity. My prediction is simple: by 2025, any centralised AI team will have effectively dissolved, their specialists becoming embedded agents or moving on.
The infrastructure flip and the death of the specialist priesthood
The first reason for this dissolution is the infrastructure. What once required a dedicated, custom-built MLOps pipeline is now commoditised. We are past the era of proprietary, hand-rolled model deployment. Vector databases, automated training pipelines, and simplified deployment APIs are now plug-and-play components.
AI development is no longer a specialised craft; it is a standard engineering task. When the infrastructure becomes trivial, the technical complexity shifts from building the foundation to applying the focus. A central team is designed for foundational, generalised tasks; it lacks the necessary focus and deep domain knowledge required to build narrow, high-impact agents. They become an internal agency, often delaying product squads who wait in a queue for resources or a deployment slot. We cannot sustain that level of inefficiency.
The power of the embedded constraint
The true value of an AI agent is in its constraint; its narrow, precise utility within a product’s workflow. This is a concept I have long advocated for: constraint drives creativity and focus. Just as a self-imposed limitation can sharpen a creative process, a highly constrained, embedded agent delivers disproportionate return.
The product team that owns the customer support journey is the only team that truly understands the latency requirements, the specific nuances of the data, and the precise cost of a faulty hallucination. Therefore, the responsibility for the product’s AI agent must reside with that engineering squad.
We do not need a team of highly-specialised PhDs to manage the fine-tuning and monitoring of a dozen bespoke, narrow LLM agents across the company. We need AI-fluent Full-Stack Developers who can seamlessly integrate an agent, monitor its performance against P&L metrics, and iterate instantly. This is the talent transition we must make. The ML Engineer’s job shifts from building models from scratch to curating, deploying, and maintaining already-developed models within a specific, well-defined product context.
Strategy for the 2025 reality
The path forward requires a fundamental shift in organisational design and training. We must stop hiring “AI rock stars” for a central team and start investing in our existing developers.
Our strategy must focus on:
Decentralised ownership
The embedded agent must report to the Product Owner; its success metrics are the product’s success metrics. This ensures alignment and cuts out the bureaucracy of a central group’s competing priorities.
Upskilling engineering talent
Start training full-stack teams immediately on deployment, monitoring, and basic fine-tuning of commodity models. This is about making AI an everyday tool in the developer’s kit, like a database or a cloud function.
Focus on product flow, not pure science
The goal is to integrate a seamless user experience, not to push the boundaries of machine learning theory. Our focus is on solving the user’s problem, not proving a new model architecture.
The age of the centralised AI lab is ending. It was a necessary bridge, but it is not the destination. The future belongs to the rapid, decentralised execution of narrow, embedded AI agents, owned and operated by the product squads who best understand their value. The companies that recognise this strategic imperative now will be the ones dominating their markets by 2025. We need to organise for speed, not for academic rigour.
Embedded agents is smart move, no? We build the agents to operate on the data at rest inside the platform. This stops the data exiting the secure perimeter. Much better than API calls to outside AI model. Security team is happy with this.
We’re calling this ‘in-situ processing’—it’s less about the agent and more about the compute being right next to the data lake. It reduces latency and massively simplifies GDPR compliance for European data. Are you finding this model accelerates deployment for regulated industries? That’s our bottleneck.