A reliable data platform tells you the numbers are accurate. It does not tell an AI system what your business means by them.
Consider a finance team asking an AI assistant which customers are becoming less profitable. The transactions reconcile. The customer records are complete. The data platform is working as intended. Yet the answer changes depending on whether the assistant uses invoiced revenue, recognised revenue or revenue after rebates, and whether it includes the cost of servicing each account.
This is an illustrative situation, but the distinction matters. Every number could be accurate while the conclusion remains unsuitable for the decision. An experienced finance manager knows which definition applies, which exceptions need attention and when a commercial agreement changes the interpretation. An AI system needs access to that knowledge in a form it can use consistently.
Before extending AI into business decisions, leaders need to establish who owns that interpretation and how it will be maintained. Otherwise, the organisation risks automating a conclusion that its own teams have never agreed is correct.
Building a reliable data foundation has been a sound priority for modernisation programmes. Integrating sources, standardising pipelines, improving quality and controlling access addressed real constraints. Those investments remain essential. Better instructions alone do not make missing transactions or unreliable source records trustworthy.
However, making data available does not, by itself, establish what the business means by it. A field labelled "active customer" might refer to a recent purchase, a live contract or an account that has not been closed. Each definition can serve a legitimate purpose. The difficulty begins when an AI system selects one without recognising which purpose governs the question.
This is the role of a context foundation: the business definitions, relationships, source authority and operating rules that allow AI to interpret information and act within agreed boundaries. Here, the term describes the proposed architecture; it does not refer to a formal standard or a single product. Its value lies in making work visible that can otherwise sit between data engineering, business ownership and application delivery.
Much of that context may already exist. It can be embedded in reporting models, planning assumptions, process documents and the judgement of experienced employees. The task is to establish which of it is authoritative, make it usable by AI and keep it current as the business changes. Buying another platform does not settle those questions.
Platform developments are beginning to address the technical effort involved. Snowflake describes Horizon Context as a governed layer for business definitions, metrics, relationships and logic, with capabilities to collect and enrich metadata across systems. Its published offering includes automation intended to reduce the work of building and maintaining context. Some supporting features are labelled as preview capabilities on its product page, so availability needs to be checked for the intended deployment.
Databricks describes Genie Ontology as combining governed context, including metric views and domains, with context inferred automatically from assets and usage. Its documentation lists the feature as a public preview, with access arranged through the Databricks account team. The capability brings explicit business modelling and discovered knowledge into the same context layer, but its availability should not be assumed across every environment.
Taken together, these developments suggest a direction towards more platform assistance with collecting and organising context. They do not establish that inferred context is automatically suitable for a particular business decision. A platform may detect which dashboard is used most frequently. That does not prove the dashboard contains the approved definition for group reporting. Usage provides evidence of relevance; authority requires a business decision.
Suppose, for example, that a sales team uses a forecast which includes opportunities finance excludes from its committed view. Both teams may be working correctly. An AI system that merges those definitions could produce a forecast neither team would approve. A system that preserves the distinction can answer each question appropriately and explain the basis of its response.
Asking the technology team to "improve the context" leaves the substantive decision unresolved.
The organisation therefore needs a named owner who can approve each consequential definition, resolve disagreement and authorise changes. This is specific work. Someone must decide whether a measure applies to a region, product or reporting period, and what happens when a source conflicts with it.
The consequences become more significant when AI can take action. An assistant recommending a change to a customer account gives an employee an opportunity to intervene. An agent applying that change directly requires explicit operating limits. Knowing that an account is unprofitable does not establish permission to change its pricing, suspend service or contact the customer.
Those permissions must be enforced through the workflow and connected systems. The design needs to distinguish actions an agent can complete independently from those requiring approval, define when it must stop, and retain enough evidence to investigate what happened. Where changes can be reversed, the recovery process should be designed before the agent is given access to make them.
An AI gateway can form part of this control architecture, providing a managed point for access to models and, depending on the product, tools and agent interactions. Databricks Unity Gateway, for example, documents controls for AI service access, traffic, guardrails and usage monitoring. The implementation still needs to enforce the business application permissions and approval rules required for each action.
The proposed Data Foundation to Context Foundation approach brings these dependencies together. Semantic modelling defines measures, business terms and their relationships. An ontology provides an explicit representation of concepts and relationships within a domain. Neither needs to become an abstract exercise detached from the work AI is expected to perform.
Integrated sources, standardised pipelines and controlled access. The foundation the rest of the approach depends on.
Gathering the business definitions, relationships and rules already embedded in reporting models, planning assumptions and people's judgement.
Defining measures, business terms and their relationships, and representing concepts and relationships explicitly within a domain.
Applying that context so AI can interpret information and act within agreed boundaries, not just answer questions.
A managed point of control for access to models, tools and agent interactions, enforcing usage and traffic rules.
The permissions, approval rules and evidence trail that determine what an agent can do independently and what requires sign-off.
The six stages provide a way to organise the design, rather than a rigid implementation sequence.
Security, ownership and monitoring must run throughout. Lessons from a live workflow should also feed back into its context: which definition caused confusion, which source was stale and which exception required intervention. Context needs maintenance just as data pipelines do.
For most organisations, the practical starting point should be one decision or workflow with a clear owner and a measurable operational problem. A finance team might begin with explaining margin movements. It can define the approved calculation, identify the supporting sources and test whether AI handles rebates, currency effects and incomplete cost allocations correctly before allowing it to propose downstream changes.
This creates a bounded way to assess readiness. The team can test familiar cases alongside disputed definitions and missing information, then judge whether the system answers correctly, makes its basis clear and escalates appropriately. A fluent answer is insufficient. In some cases, the correct behaviour is to recognise that the evidence does not support a conclusion.
This also changes what leaders should expect from a delivery partner. As platforms take on more discovery and configuration, delivery partners should place greater emphasis on establishing business authority, designing the operating model and connecting AI to real processes. The partner needs to help determine where existing platform capabilities are sufficient, where business definitions remain unresolved and which controls must be engineered into execution.
That work is where a context foundation becomes commercially useful. It gives the organisation a basis for deciding which AI use cases can progress, what must be resolved first and who remains accountable once the system is live. More collected context is not a sufficient measure of progress. The test is whether AI can use the right context for the decision and remain within its authority when acting on it.
If your organisation has invested in a trusted data platform and is now considering AI agents, the next step is to examine one priority workflow through this approach. Start with the data it depends on, establish the meaning and authority of the context it needs, and define the actions the business is prepared to delegate.
Book a session with Decision Inc. to apply the Data Foundation to Context Foundation approach to a priority AI use case. Together, we can map the path from reliable data through context collection and business meaning to agentic AI, gateway controls and governed action, and identify what needs to be in place before that use case moves into operation.
Talk to us about applying this approach to a priority AI use case, from reliable data through to governed action.