By Ervin Daniels, CISSP, CCSP, CBCP
An employee asks an AI assistant, “Can this customer information be shared with an outside partner?”
The assistant finds a document and produces a confident answer.
But which version of the document did it use? Does the policy apply to that customer? Did the employee have permission to see the underlying information? And did the answer reflect the latest approval status?
These questions expose a challenge that businesses encounter as they begin developing AI. Connecting a model to company information is one step; making that information understandable, reliable, and appropriate to use takes additional work.
For organizations getting started, the data already inside their environment offers a practical foundation. Building on that foundation requires two capabilities working together: the ability to access and process data, and the ability to understand and govern it.
Start with a question the business needs answered.
A useful first AI project begins with a task people already perform. An internal support assistant could help employees find an approved troubleshooting procedure. A procurement assistant could retrieve contract terms and current supplier information. A governance assistant could gather the evidence needed to review a data-sharing request. Each example has a defined user, a recognizable problem, and information the organization already owns. That makes it possible to ask concrete questions before selecting technology: What information does the task require? Where does it live? Who owns it? How current must it be? What would a correct answer look like? Those answers establish the scope of the first project.
Make the information accessible.
Enterprise information often lives across databases, warehouses, cloud storage, document repositories, and internal applications. An AI application needs a controlled way to reach the sources relevant to its task. For some questions, that means querying a database. For others, it means searching approved documents. Some sources can be accessed where they already reside; others need preparation, transformation, or indexing before the application can use them effectively. A shared data platform can support these access and processing needs. Open formats can make prepared data reusable across compatible tools, while different processing engines can handle different jobs efficiently.
For an initial project, the goal is a dependable path to the few sources that matter. A support assistant may need one approved procedure library and access to current ticket status. Establishing that path gives the team something specific to build and test.
Give the information meaning.
Access alone leaves important questions unanswered. A field labeled “active” might describe an open account in one system and an employee’s employment status in another. A document titled “Current Policy” may have been replaced months ago. Data intelligence supplies the context needed to interpret those assets. A catalog helps people discover available information. Business definitions explain what fields and terms mean. Ownership identifies who can resolve questions. Quality checks reveal missing or inconsistent values. Lineage shows where information originated and how it changed.
Together, these capabilities help developers select suitable sources and give reviewers a way to investigate an answer. For example, an assistant answering a question about external sharing may need to distinguish between “approved for internal use” and “approved for third-party distribution.” That distinction must be reflected in the source information and application design.
Build a small foundation for the first application.
A manageable starting point is one business workflow, a limited group of users, and a few approved sources. For each source, establish its owner, business meaning, sensitivity, quality expectations, and update frequency. Define which users may access it and what actions the application may perform. Then build the AI workflow around those requirements. One common approach is retrieval-augmented generation, or RAG. The application retrieves relevant enterprise information and supplies it to a model as context for an answer.
A support assistant, for example, could retrieve an approved procedure and use it to explain the next troubleshooting step. A separate query could check whether the affected service is currently experiencing an outage. The application should identify the evidence supporting its answer. When evidence is missing or conflicting, it needs a defined response, such as requesting clarification or sending the question to a person.
Carry access controls into the AI workflow.
An AI assistant creates another route to enterprise information. Its access design deserves the same attention as any other application.
A user’s question should retrieve only the information that the user is authorized to access. Those restrictions need to apply before sensitive content reaches the model. Search indexes, cached results, conversation histories, and application logs also need appropriate controls because they can retain information from the sources.
If the application performs actions, define those permissions separately. Reading an approval record and granting approval are different responsibilities. A useful first workflow can gather evidence and prepare a request while routing the decision to an authorized reviewer.
Test trust alongside speed.
A fast answer has limited value if employees must repeat the entire investigation to verify it. Evaluate the first application against questions reviewed by people who understand the task. Check whether answers are correct, whether supporting sources are relevant and current, and whether access restrictions work. Include cases where information is missing, outdated, or contradictory. Observe how the application responds and whether it directs the user toward an appropriate next step. Measure the business outcome as well. Track time spent finding information, time needed to complete the task, manual rework, and operating cost. These results reveal whether the application makes the workflow easier and where the data foundation needs improvement.
Let the first project create reusable value.
The first AI project can leave behind more than an application. It can produce documented business definitions, clearer ownership, tested access rules, quality checks, and reusable data connections. Those assets make the next project easier to scope and build. This is where data access and data intelligence reinforce each other. One provides the means to retrieve and process information. The other helps explain whether that information is suitable for the task. A practical AI starter kit begins with a business problem and the organization’s own data. Establish how to reach that data, understand its meaning, control its use, and verify the answers it supports. That gives the team a foundation it can expand as results justify further investment.
Post comments (0)