Unlocking Your Unfair Advantage with Context Engineering

A customer walks in your store and asks for a refund. What do you need to make a decision? Transaction details, perhaps the customer's history? These alone are not enough. You will likely need information on refund details, approver authority, payment methods, etc. While transactions live in a database, where does the rest live? SharePoint? PDFs? Your employees' heads? Now imagine you are an AI agent: where do you start to make sense of these everyday processes?

As I have written previously, while AI brains have developed at pace, they are nothing without the right nervous systems – which plumb in the context required for reasoning and decision-making. Two things have been clear over the course of rolling out enterprise AI at scale. Firstly, context is a critical differentiator in agent performance, and therefore, a potential competitive advantage. Secondly, the most valuable context is not always neatly organised and available in databases and other systems of record.

Now the idea of different types of information is nothing new. For years, we have spoken about structured vs. unstructured data, with different roles tackling different aspects of each. Data engineers built pipelines and designed databases to organise the former, whilst the latter was often left to individual employees and teams. Best case scenario, dedicated SharePoint administrators were hired to tame rogue documents through clear ownership, permissions, and version control rules.

What is different now is the valorisation of this unstructured data; or in many cases even uncaptured data. After all, what makes your best sales manager a great sales manager? How much of that is written down or codified in some form? The paradox is that when you look closely, you start to appreciate how much of organisations' secret sauce – their differentiated capability and performance – is down to such latent expertise and experience. If only you could scale it.

Welcome to context engineering.

Let us forget about tasks for a moment and consider outcomes. What is the end we are trying to achieve? For each domain, to make the right context available to any agent, at the right moment, with sufficient clarity, authority, and completeness to make good decisions. For example, when interacting with a customer, to not only deeply understand that individual, but to have access to the policy and best practices that enable the agent to delight them. Creating these context cores requires an ability to aggregate, organise, and enrich information in a comprehensive and customer-centric manner.

In some ways, this is an evolution of analytics engineering – a role created to complement data engineering through capturing business meaning. After all, beyond the raw transaction records: what counts as revenue, how is margin calculated, and how should we classify cancelled orders? However, the opportunity is to expand that remit to look beyond structured data and to opportunistically aggregate, codify, and connect the information required for great outcomes.

Beyond the data and features themselves, cores should therefore also contain rules on how context should be consumed. From consent flags through to confidence scores, a core needs to give an agent everything it needs to successfully navigate a domain – much like a seasoned employee onboarding a new hire would. Insisting on this clarity tends to act as a forcing function internally: unless standardised operating processes are genuinely standardised, how can you build your core?

For all of AI's strengths, agents struggle to navigate internal siloes and office politics. The paradox is that messy input data used to be fatal; now it's the messy processes that are particularly lethal. In a world where agents can autonomously figure out how to achieve an outcome, nothing is more important than being crystal clear about what that outcome looks like – and what the constraints are. Lifting a term from reinforcement learning: what is the reward signal?

The cliche goes that data is the new oil. Maybe that analogy is better applied to neglected SharePoint folders and Drive files. In their raw form, they do little for your organisation. But seek them out, distill the disparate wisdom they contain, refine them using the latent context that lives in colleagues' heads and underpins their email exchanges – and you have the fuel to supercharge your engine. Dusty PowerPoints, PDFs, and policies might just be the ingredients you need to win.

In a world of rapidly commoditising agentic intelligence, every company will have access to increasingly capable AI. Far fewer will have systematically captured the judgement, policies, and relationships that make their organisation tick. Context engineering is how we turn that accumulated intelligence into something agents can leverage to create differentiated value. The next competitive advantage in AI may not come from access to a better model, but from an organisation's ability to make its collective intelligence available as context to those models.

– Ryan

Cover image by ChatGPT.