Blog

Data Steward: Who Makes Sure AI Knows Your Business

Data Steward: Who Makes Sure AI Knows Your Business

Companies are launching AI pilots by the dozen, but only a handful ever reach production. The reason is not that the models are weak — it is that AI does not know your business.
Tech giants have spent billions of dollars and fed their models half the internet, but not your contracts, your processes, or the way decisions are made at your company. So you can plug in the most powerful LLM on the market, ask it "what useful thing could you do specifically for our company?" and get an honest answer:
"I'd be glad to help, but to do that I need to know more about your company."
The vast sums of money and the gigabytes of data poured into a model deliver no value to you on their own. The model is well-read "on the internet average," but it knows nothing about your processes, your clients, or how decisions get made in your organization.

The Obvious Solution: Give the Model Context

The logic is simple: for AI to deliver value, you have to open up your knowledge about the company and its projects. The most popular way to do this is to attach a RAG (Retrieval-Augmented Generation) database to your queries.
Simply put, RAG is a purpose-built store that searches by meaning rather than by literal keyword match. When a user turns to the AI, the query text is run through the RAG database to find the document fragments closest in meaning to the question. Those fragments are appended to the original query, and this "augmented" query is sent to the model. As a result, the question reaches the AI together with "hints" drawn from your data — and it is these hints that help the model answer in the context of your situation.
It sounds as though the problem is solved. Dozens of RAG startups grew on exactly this promise, followed by a wave of MCP (Model Context Protocol) connectors that link models to corporate systems directly. The number of tools has grown. The share of AI pilots that reach production has barely changed.
The reason is that this is not a technology problem.

The Problem Is Not RAG — It Is Knowledge Management

Five sequential questions stand between "let's deploy RAG" and "let's deploy RAG and get value out of it." Leaving any one of them unanswered means risking financial and reputational losses.
  1. How will you collect data from unstructured sources — legacy wikis, Confluence, folders full of PDFs? And, more importantly, how will you extract the knowledge that exists only in individual people's heads?
  2. How will you make sure nothing gets into the model that shouldn't — client names, banking and registration details, your proprietary know-how, information about upcoming deals?
  3. How will you prevent every employee from suddenly gaining access to sensitive information? The model has no way of knowing who is asking — the CEO or a support intern.
  4. How will you tell whether an employee is trying to "break into" the model in order to extract secrets and, say, sell them?
  5. How will you know the model is answering strictly from the data rather than hallucinating?
None of these questions is solved by buying yet another connector. RAG is merely a mechanism. The real question is different: what knowledge are you loading into it, in what form, under what access rules, and how do you keep it current?
Corporate knowledge is almost always unstructured: forgotten Confluence pages, folders of PDF and DOCX files, scraps of decisions buried in chat threads. And the most valuable part often sits in the heads of a few key people and is not written down anywhere at all. So before this knowledge can become useful, it must first be located, extracted, structured, stripped of anything extraneous, and kept up to date. This is continuous work that cannot be closed out with a one-time setup.

So You Need a Person, Not Just Technology

Answering the five questions above requires people with a particular skill set. You need someone who:
  • understands the company's processes and knows which data matters and where to get it;
  • at the same time understands how AI works: how to reduce hallucinations through proper document formatting, how to configure access rights, how to evaluate answer quality;
  • can use AI to speed up their own work rather than drowning in manual annotation.
This person is increasingly called a Data Steward — a custodian of data. Some use the terms "knowledge manager" or "context manager." What matters most is that this is the person accountable for making sure corporate knowledge is collected, cleaned up, tagged by access level, and kept fit for AI to consume.

How This Differs From a Business Analyst or an Instructional Designer

A Data Steward's tasks resemble those of a business analyst or an instructional designer, but there is a difference.
A business analyst documents processes and requirements for a specific project. The deliverable is a document that has already begun to go stale by the time it is handed over. A Data Steward works not with a project but with a living flow of knowledge that must be maintained continuously.
An instructional designer packages knowledge into methodologies, standard operating procedures, and training materials so that a person can study and apply it. The output is aimed at people: courses, instructions, standards. A Data Steward prepares the same knowledge for a machine, and a machine has different requirements: the data must be tagged by access level, cleared of sensitive content, and converted into a format the model will understand and will not distort.
The key difference is that analysts and instructional designers prepare knowledge for people. A Data Steward prepares it for AI, with access control and leak protection, and treats it as an asset that feeds the model.

Without Zealotry: Common Sense Instead of Total Formalization

It is easy to go to the extreme here — to try to document every process down to the atomic level and turn the company into an endless rulebook. That is a dead end, because such work never finishes and knowledge goes out of date faster than it can be documented.
A Data Steward is guided by sound judgment, not formalism. They start with the knowledge that genuinely creates value and carries risk, not with a full inventory of everything in sight. Perfect order for its own sake delivers no benefit. The goal is to set up a working AI you can trust.

The Emergence of These People Is Only a Matter of Time

Within a couple of years, a role like Data Steward will be as natural inside companies as DevOps or a data analyst is today. As soon as AI moves from demo to real processes, someone has to be accountable for the knowledge that AI relies on. It is inevitable.

What We Have Already Built at Gorgona

Rather than waiting for the role to take shape on its own, we have built out tools, dashboards, and AI assistants for the Data Steward in advance:
  • a file upload and processing tool that automatically determines a document's risk class;
  • an AI agent that helps configure the RAG database correctly and efficiently;
  • monitoring of unauthorized access to data and of intrusion attempts;
  • quality assessment tools for both the data and the model's answers.
There is already enough technology for adopting AI — there is arguably too much of it. What is missing is someone to put the knowledge in order and take responsibility for the AI answering accurately, securely, and to the point. This is exactly the person to start with.
But a single Data Steward is not yet a system. For order in your knowledge to rest on processes and rules rather than on one person's enthusiasm, you need a discipline of its own: data governance. We will discuss how it works and where to begin next time.
Spoiler alert!

Related articles

2026-07-22 14:34 Articles