What is our primary use case?
My main use case for dbt was transforming messy, extracted financial data into clean, production-ready datasets. Let me give you a concrete example. We would extract trial balance data from PDFs using OCR and then feed it through our GenFin mapping workflow. Before that data reached our GPT-4 model for classification, we used dbt to normalize account codes, handle currency conversions, work with UAE and multi-currency scenarios for UAE compliance, and create aggregations by cost center and account type. So dbt's job was essentially to take a raw extraction output, validate it against our schema, handle any missing values or duplicate entries, and then materialize clean tables that our LLMs could reliably work with.
There was one more important place beyond just transformation. dbt gave us visibility into data quality throughout the pipeline. We built tests into our models, including checking for null values in critical fields such as account numbers, ensuring amounts were numeric and within expected range, and validating that our transformed data matched the source record counts. This was critical because if there was a gap between extraction and transformation, we would catch it before it hit the LLM. The dependency graph in dbt was invaluable when we had issues downstream in our mapping or disclosure notes workflows, as we could trace back through the DAG.
The most concrete outcome was a significant reduction in data errors reaching our downstream AI models. Before dbt, we were catching bad extractions only after the LLM had already processed them, which meant manual rework. After implementing dbt's testing layer, we caught roughly 70% of those issues at the transformation stage itself, before they ever touched the model. Processing speed also improved because dbt's incremental models only processed the delta into new records. Our nightly reconciliation runs for UAE corporate tax workflows dropped from around 40 minutes to under 10 minutes. The less obvious win was team confidence. Our chartered accountant clients started trusting the outputs more because we could show them the full transformation chain.
The shift in client trust was really tangible. Before dbt, when we delivered a mapped disclosure notes document, clients would ask us to walk them through how we arrived at the specific numbers. We would have to manually trace back through extraction logs, which was time-consuming and sometimes incomplete. After dbt, we could literally show them the DAG, the dependency graph, and say, here is your source PDF, here is where we extracted the account code, here is the normalization rule we applied, and here is the aggregation, and here is your final mapped output. It is all documented and testable. That transparency eliminated a huge amount of back-and-forth. One specific example involved a client who was questioning why a particular cost center total did not match their internal records. With dbt's lineage, we traced it in minutes, found a rounding rule that needed adjustment in one model, fixed it, and reran the pipeline. The client saw the complete audit trail. That kind of visibility is what builds trust in a financial team.
One thing I would add positively was how well dbt's incremental models handled our late-arriving data scenarios. In financial processing, you often get corrections or amendments to documents days later, and incremental models let us efficiently merge those without full refreshes. That was really valuable for our UAE compliance workflows where reconciliation amendments came in batches.
What is most valuable?
dbt's incremental model let us handle late-arriving data efficiently without reprocessing everything. That clean output then fed directly into our GPT-4 structured output parser for financial mapping. The workflow fit really well because dbt gave us clear data lineage. We could trace any number of our final mapping output back to the source extract. That is critical in financial-based reporting, where auditors need to see the full transformation chain. For us, the standout features were the testing framework and the dependency graph. The tests, especially custom tests for financial data like validating that debits equal credits, caught a lot of our data quality issues early.
dbt's incremental model was key here. Rather than representing the entire dataset every run, dbt only touched the new or changed records, which kept our pipeline efficient as document volume grows. That said, dbt's scalability is fundamentally tied to your underlying database. In our setup, we managed performance by being deliberate about materialization strategies, using views for lightweight transformations and tables for incremental models for heavy aggregations.
What needs improvement?
As for something I wish we had, dbt's native support for Python transformations came later, and we did some complex financial classification calculations that felt clunky in pure SQL. We ended up writing Python in our n8n workflows and then fed the results back into dbt, which created a bit of a split-brain situation. If we would have had dbt Python models earlier, we could have kept that logic unified.
Managing multiple reporting standards was our biggest operational pain point with dbt. We were running UAE corporate tax compliance and IFRS disclosure workflows simultaneously for different clients, and dbt does not have a native concept of multi-tenant or multi-standard project organization. Everything lives in one flat structure, so we had to build more conventions: separate schema folders for IFRS models versus UACT models, custom macros to tag models by compliance regime, and environment variables to control which set of transformations run for which client.
For how long have I used the solution?
I have been working with dbt for approximately one and a half years.
What do I think about the stability of the solution?
dbt has been very stable.
What do I think about the scalability of the solution?
For our use case, dbt scaled well. We were processing large volumes of financial documents, hundreds of trial balances, balance sheets, and invoice sets, and dbt handled the transformation layer without issues.
How are customer service and support?
dbt's customer support varies depending on which tier you are using. We ran dbt Core, which is open-source, so there is no direct vendor support. The community support is actually quite good, and the dbt Slack community is also very supportive. We found answers to most of our questions through the docs and community. If we would have been on dbt Cloud, there would be dedicated support, but the pricing did not work for our team size. For a bootstrapped FinTech team in India working on UAE compliance workflows, the cost-benefit was not there.
Which solution did I use previously and why did I switch?
Before dbt, we were handling data transformation in a more ad-hoc way. We had Python scripts and custom SQL queries scattered across our n8n workflows, our automation platform. When we extracted financial data from PDFs using Tesseract OCR and LLM extraction, we would clean and transform it using Python code nodes within n8n. There was no clear lineage. If the number was wrong downstream, tracing back through n8n workflow logic was painful.
Which other solutions did I evaluate?
We did evaluate alternatives before committing to dbt. We looked at a few options. Apache Airflow was one. It is powerful for orchestration but felt like overkill for our use case and would have required significant DevOps overhead to manage. We also considered Talend and Informatica, but those were enterprise tools with enterprise pricing, and we were a lean team at Radiant Services.
What other advice do I have?
The most concrete outcome was a significant reduction in data errors reaching our downstream AI models. Before dbt, we were catching bad extractions only after the LLM had already processed them, which meant manual rework. After implementing dbt's testing layer, we caught roughly 70% of those issues at the transformation stage itself, before they ever touched the model. Processing speed also improved because dbt's incremental models only processed the delta into new records. Our nightly reconciliation runs for UAE corporate tax workflows dropped from around 40 minutes to under 10 minutes. I would rate this product an 8 out of 10.
Which deployment model are you using for this solution?
On-premises
If public cloud, private cloud, or hybrid cloud, which cloud provider do you use?
Other