Beyond centralization: Why APAC retail and CPG teams need a context layer
.jpg)
Ask an AI agent how many days of supply you have on your top promoted SKU across Australia, India, and Indonesia, and you will get an answer. It will be fluent and confident. It may also be wrong.
The agent is not the problem. It has to guess which table holds the answer, which definition of "days of supply" applies, and whether last night's feed actually landed. Nobody told it because that knowledge lives in an analyst's head, a promo spreadsheet, and a wiki page nobody has opened since the last reorg.
Parts one and two of this series covered getting trusted data into the tools where people work (Activation) and making that data trustworthy (dbt). This part is about the missing piece for AI: context, or what your metrics mean, where they come from, who owns them, and whether they can be trusted today.
[CTA_MODULE]
What retail and CPG teams are asking us right now
Recent conversations with retail and CPG teams across APAC reveal a common challenge: adopting AI is relatively easy, but giving it trusted business context is not. Three patterns stand out.
- AI projects are hitting a data wall. One consumer brand traced its agent’s slow responses to the data foundation beneath it, including the lack of a semantic layer with shared business definitions. Other teams want every AI-generated answer to be traceable to its source. For organizations still exploring AI, governance and data readiness are often bigger obstacles than model choice.
- The numbers do not reconcile, and teams cannot fix everything manually. Data remains fragmented across clouds and warehouses, reports produce conflicting figures, and business users lose confidence in the results. Many teams have only 1 or 3 data specialists—or no in-house data engineers at all. They cannot rely on someone to document every table, reconcile every definition, and keep that knowledge current by hand.
- Governance comes before features. Teams want to know which definition takes precedence, who can access what, how answers can be verified, whether existing semantic models can be reused, and what the system will cost to operate. These are not secondary implementation details. They determine whether anyone can trust the agent’s answers.
Together, these conversations point to a clear requirement: AI agents need a governed, scalable way to find trusted business meaning without creating more manual work for already lean data teams.
What a context layer does
An AI agent can only use the context it can find. Without help, it sees table and column names but may not understand what they mean, which assets to trust, or how the business calculates a metric.
That’s where a context layer and Agents Schema come in. Think of the context layer as a library system and the Agents Schema as its card catalog. The context layer gathers, organizes, and serves knowledge from across your business. The Agents Schema stores that context in a standardized, queryable structure inside your warehouse.
Together, they give your team:
- One place to understand the data. The Agents Schema shows agents what metrics mean, which tables to use, how assets relate, and who owns the underlying logic.
- Context from the tools you already use. Metadata connectors and the dbt Catalog collect knowledge from semantic layers, wikis, BI tools, and warehouse activity.
- Documents agents can search. Context layer-ready data connectors parse and index content such as wiki pages and support tickets, making the information inside them retrievable.
- One governed interface. Agent Context MCP, built on the Model Context Protocol, gives compatible AI tools a consistent way to access that context.
The system is also designed to remain open. The Agents Schema is an open-source standard for storing context in your warehouse, so your business knowledge remains in your destination rather than being trapped inside one tool.
For example, an agent asked about monthly recurring revenue does not have to infer its meaning from a column name. It can first consult the Agents Schema for the approved definition, relevant dbt model, lineage, documentation, and owner. Only then does it write the SQL. The same process applies whether the question is about monthly recurring revenue, days of supply, or another business-specific metric.
Context is not a one-time build
Context does not stay accurate on its own. Definitions change, assets are retired, schemas evolve, and owners move on. Effective lifecycle management keeps business context current as those changes occur. When agents can assess the quality and status of that context, they can “qualify an answer, seek alternative sources, or decline to answer” rather than presenting uncertain information as fact.
Automation makes that maintenance even more important. As the warning goes, “automation doesn’t fix ungoverned data. It amplifies it.” An agent working from stale or unverified context can act confidently on the wrong information — and do so at scale.
Launching a context layer should therefore mark the beginning of an ongoing validation process. Before connecting it to production agents, test real questions and confirm answers are accurate. Include local product names, mixed-language phrasing, and the imperfect questions people actually ask. Have metric owners review the results, feed every miss back into the underlying definitions, and establish a regular review cadence.
A focused rollout might look like this:
- Choose one recurring business question, such as days of supply or promotion uplift.
- Select the datasets needed to answer it, along with at least one documentation source, system of record, and analytics source.
- Build the context and test it using real-world phrasing.
- Connect one agent, assign an owner and review cadence, and expand once the answers prove reliable.
Give AI the context to earn trust
APAC retail and CPG teams do not need AI that simply produces answers faster. They need AI that uses the same definitions as the business, works from current data, and can show where every answer came from.
Centralization brings the data together. dbt makes it tested, documented, and trustworthy. Activation puts it into the tools where teams take action. A context layer completes that foundation by making business meaning accessible to AI agents in a governed, reusable format.
The best place to begin is not with every dataset or use case. Start with one important question, define what a trustworthy answer requires, and test it with the people who own the metric. Then expand as confidence grows. That is how AI moves from a promising experiment to a dependable part of retail and CPG operations.
[CTA_MODULE]
Related blog posts
Start for free
Join the thousands of companies using Fivetran to centralize and transform their data.
