What is the real foundation of AI you can trust?

From AI answers to business confidence.

---

The conversation about agentic AI usually centres on the models. That is understandable, and it can hide a more important point. The scalability of agentic AI is set less by how sophisticated the model is than by how mature the data foundations beneath it are.

Models will keep changing. The architecture that gives them something coherent to reason over is what decides whether agentic AI works at scale. Value is moving into the underlying architecture of organisations, and that is the layer Keyrus focuses on when advising its clients of scalable, effective deployment of AI.

Register for this webinar on 12 November, where Keyrus moves the conversation beyond AI tools towards the data architecture that makes AI valuable, scalable and trustworthy https://my.demio.com/ref/sbKsWy2jJbaVEtaj.

A simple test shows why. Ask an AI assistant: "What was our gross margin by product line last quarter?" Where definitions and sources are agreed, everyone gets the same answer. Where they are not, two colleagues can get two completely different answers. We will come back to that test, and to four stages of connecting assistants to your data.

Technical integration solves access. It does not solve meaning.

When terms such as customer, product, contract or risk exposure are defined differently across systems, an agent can retrieve the data and still misread it. That is a semantic issue, not a modelling one. Left alone, it builds up as semantic debt, on top of the data debt most organisations already carry.

Two ideas help.

1. An ontology is a structured business vocabulary that machines can use. It says what things exist, what they mean and how they relate: a customer places an order, an order contains products, a product belongs to a category. A glossary describes. An ontology lets systems reason.
2. A knowledge graph organises data by relationships instead of rows and columns, so an assistant can follow connections, find context and reason across domains. The ontology defines the meaning and the rules. The knowledge graph fills them with real data.

The good news is that this work compounds. Definitions agreed once serve every assistant you use today and the ones that do not exist yet.

Each stage connects an assistant to more of your business. Each is only as good as the data layer it reads through.

Stage 1 is the chat interface

Claude, ChatGPT and Gemini. Most people start by pasting a spreadsheet into the chat, then connect a source such as a database or a file store. It is quick, and it suits one person exploring an idea. The assistant has to decide what the columns mean and which table is the right one, and it decides alone.

A governed data layer adds agreed definitions and one approved source, so the connection reads something worth reading.

Stage 2 is the assistant inside your application

Power BI with Copilot, Tableau Pulse, Snowflake CoWork, Databricks Genie and Qlik Answers. This is the most convenient route, because the assistant lives where the data already is.

Snowflake describes CoCo as the builder's agent that creates governed assets, and CoWork as the business-user agent that uses them. Snowflake says CoWork applies existing row and column security policies and traces each response to its source data. Qlik describes Qlik Answers as grounded in its analytics engine and in curated documents, with citations and explanations.

The limit is that each assistant sees only its own platform, and each needs its own model of what things mean. If margin is defined in three platforms, it can be defined three ways. A well-modelled dataset helps the built-in assistant answer well. A loosely modelled one gives answers that are harder to trust. The data layer lets you build the definitions once and re-use them in every tool.

Stage 3 is your own applications, with MCP

The Model Context Protocol, or MCP, is a standard way for an application to expose its data and actions to an assistant. A system you built yourself, such as an operations platform or a customer portal, can offer an MCP connection, and Claude or ChatGPT can then use it. Vendors are moving the same way. Qlik has launched an MCP server for third-party assistants, and Tableau Knowledge is MCP-enabled.

This is also where assistants begin to do work rather than only answer. Claude Cowork and ChatGPT Work can run tasks on a schedule using those connections: a Monday margin summary, a daily exceptions list, a month-end check. ChatGPT Work can also react to a trigger or watch for a change.

Once work runs unattended, permissions, definitions and an audit trail matter more. An MCP server that sits on governed data means the scheduled output can be trusted and each person sees only what they are allowed to see.

Stage 4 is a shared teammate across your sources

Allmates and Tableau Knowledge. Tableau describes Tableau Knowledge, an early-stage product, as linking published data sources, workbooks, dashboards, the catalogue and unstructured content such as PDFs and wikis into a knowledge graph that agents reach over MCP. Salesforce, which owns Tableau, describes the underlying engine in three layers: semantics, ontology and the knowledge graph. Allmates builds teammates with defined roles, instructions, tools and permissions, connected to your systems of record, with approved shared knowledge, role-based access and traceability.

Here the set-up stops belonging to one person and starts belonging to the organisation, which we return to below.

Take the margin question again. With a governed data layer, there is one approved source and one agreed definition of margin, so two colleagues get the same answer and can trace it. Without one, an assistant picks whichever table looks right and applies its own idea of margin.

The test also shows the difference between personal and shared. In stages 1 to 3, one person connects their own tools and asks their own questions. In stage 4, a domain expert such as the head of finance builds a teammate once, with the right sources, definitions and rules, and anyone in the organisation can use it. The expert decides what it may see and do, so people stay in command.

Many agent projects shine in a demonstration and slow down on the way to production. Demonstrations run on curated, harmonised data. Production brings fragmented architectures, unclear data ownership, incomplete lineage and definitions that disagree. That is systemic friction, and it points to the foundation, not the model.

The organisations that move beyond experimentation tend to settle five things first.

• Strong data governance.
• Clear domain accountability.
• Interoperable architectures that any assistant can reach through open standards.
• Explicit semantic frameworks, meaning agreed definitions people and machines can both use.
• High-quality master data.

Keyrus is hosting a live webinar on these four stages, where expert Craig Andrew will guide you through practical demonstrations of how to eliminate inconsistent AI answers. Join Keyrus on 12 November 2026 by registering at https://my.demio.com/ref/sbKsWy2jJbaVEtaj.

AI does not transform businesses. Architected intelligence does. Keyrus operationalises intelligence, starting with the data foundations. Contact Keyrus at sales@keyrus.co.za.