Enterprise data & analytics

Agentic AI for Leaner Data Solutions

For a large international enterprise, PROVADIT uses agentic AI to develop and enhance Databricks and dbt models. Comparing alternatives, reviewing code and documenting the underlying logic become part of one continuous workflow. This helps us reach well-founded, maintainable solutions faster, with experienced developers directing and verifying the work.

Explore the case

Industry: Enterprise data & analytics
Our role: AI-assisted model development, solution comparison, code review and documentation
Focus & technology: Agentic AI · Databricks SQL · dbt · Confluence · Markdown
Scope: Lakehouse build since 2023 · AI-assisted development and documentation in the current delivery phase

Client anonymized.

The starting point

How can a growing lakehouse become easier to develop, review and understand?

A large international enterprise is consolidating operational data in a central Databricks Lakehouse. Its dbt transformation and semantic layer contains many models with detailed column metadata and business rules. Each change needs to remain understandable to business owners, analysts and the engineers who will maintain it later. Documentation must explain both what the code does and why the chosen approach makes sense.

The inputs include proprietary code, business rules and column semantics. The working environment therefore matters as much as model capability. In this engagement, PROVADIT works inside the enterprise AI platform and integrations already governed by the client.

Our contribution

PROVADIT’s contribution

PROVADIT uses agentic AI as a daily engineering tool within the client’s own enterprise AI platform. The agent works with the existing SQL, metadata and a developer’s explanation of the requirement. Our developers use it to explore alternatives, enhance models, challenge code and prepare documentation. They set the direction, resolve business questions and verify every artifact before it reaches the client.

Compare options and find the leanest suitable approach
Agentic AI makes it easier to explore several ways of meeting a requirement and challenge the first plausible answer. PROVADIT compares the options for fit, complexity, maintainability and performance. We look for the simplest approach that meets the actual constraints, reuses existing models where appropriate and avoids unnecessary dependencies. The engineering decision remains with the developer.
Enhance models in their existing context
The agent reads the existing dbt model and surrounding metadata before proposing changes. It supports new SQL logic, adaptation to changed requirements, refactoring and the review of edge cases. Working from the existing context helps preserve business meaning while making the code clearer and easier to maintain.
Review code automatically and verify the findings
AI-assisted code reviews provide an additional check on logic, consistency, edge cases and opportunities to simplify a change. The agent can challenge assumptions and propose checks that a developer then evaluates. PROVADIT reviews the findings against the requirements and validates the resulting changes. An automated review supports human judgment; it is not an approval or a guarantee that code is correct.
Generate documentation from code and developer intent
The developer provides a functional and technical overview together with SQL and column metadata. From these inputs, the agent drafts a design document covering scope, business logic, the data model, source fields and implementation guidance. Missing definitions and unresolved source questions remain explicit open items for the solution owner. PROVADIT checks the document against the code and the intended behavior.
Publish where the team works
The reviewed content is prepared and published in Confluence Wiki Markup and Markdown. Anchors, info and note macros, and tables follow the client’s conventions. One set of source inputs produces both a readable wiki page and a versionable file, reducing manual reformatting and keeping documentation close to development.

Inside the work

What makes the documentation useful?

Code explains implementation; the developer adds purpose and context. Together, they give the agent a grounded basis for documentation that business users and engineers can actually use.

Purpose and boundaries
Explain the requirement, intended users, scope and what the model does not cover.
Business logic and source fields
Connect calculations and transformation rules to the data model, column meanings and source-field references.
Implementation and maintenance
Set out concrete implementation steps, shared rules and the information another engineer needs to make a change.
Open items that stay visible
Identify missing or ambiguous source definitions for clarification with the solution owner instead of presenting assumptions as facts.

AI that fits the client’s environment

The delivery model follows the confidentiality requirements and the available infrastructure. The two approaches below describe different engagement contexts.

In this engagement: the client’s enterprise AI platform
PROVADIT uses the client’s own agentic AI platform and integration within its established controls. The existing environment determines how access, model use and proprietary information are handled.
In other engagements: local open-weight models
Where a suitable enterprise platform is unavailable, PROVADIT also works with capable local AI hardware and open-weight language models. In a fully local setup, proprietary inputs can remain on the client’s infrastructure. Available hardware and model capacity determine the context size and the long-running or batch tasks that can be handled.

Deliverables

The work delivered. The decisions it supports.

The ongoing work brings implementation, review and documentation closer together. PROVADIT can examine more options with less manual drafting and formatting effort, while the client receives models and explanations that have been checked by people who understand the business logic.

Compared solution options and enhanced Databricks SQL/dbt models, with attention to unnecessary complexity and maintainability.
Automated code-review findings assessed by experienced developers, with changes verified against the requirement.
Code-based functional and technical documentation, with source-field references, implementation guidance and explicit open items.
Consistent Confluence Wiki Markup and Markdown deliverables within the client’s governed AI environment.

Engagement context

The lakehouse build has been ongoing since 2023; agentic-AI-assisted development and documentation describe its current delivery phase. This case focuses on PROVADIT’s model development, review and documentation work within the client’s own enterprise AI platform. Local open-weight deployment is a separate option used in other engagements.

What this illustrates

The benefit is more than writing code faster. Agentic AI gives a senior team more capacity to compare alternatives, question complexity and keep the reasoning behind a model documented as it evolves. PROVADIT combines that speed with business understanding, engineering judgment and human verification, so that the solution is easier to understand and continue developing.