Zero-basing our finance foundations

AI-native Finance functions need strong data foundations.

Our business at Airwallex is complex. We enable our customers to collect, hold, convert, send and spend money globally. Each of these activities creates a chain of micro-transactions.  A single sale, for example, produces records for what the customer paid, what the card network transferred, what Airwallex charged, and what the business received. Multiply this by millions of transactions across many markets and unique vendor-specific practices and the scale of the data challenge becomes clear.

Our goal as a Finance team is to enable the business to understand what is happening without getting lost in the minutiae. That means we need to be able to unify individual records in a way that allows us both to explain the overall narrative and to drill into the details where needed.

Airwallex has grown significantly in the past few years. In 2018, our annualised revenue was less than $5m; it is now more than $1.3b. During this phase, the priority was data accuracy over data unity. We needed to be utterly confident our numbers were accurate and would stand under scrutiny. Each product owner had the final say on defining their local data schema. But without an overarching vision on how to unify these building blocks, the result was data fragmentation and inconsistent definitions across the business. 

Over the past year, we have undergone a major data migration from BigQuery to Databricks. We saw this as an opportunity to zero-base our data foundations. We rebuilt them with a unified global vision, clear ownership and governance, creating a base that could support the business well into the future.

Putting Finance in the drivers seat of the migration

We took the unconventional approach of putting Finance in the driver’s seat of our data migration. Rather than delegate to Data and Engineering teams, Finance took ownership of identifying the use cases and working backwards from the decisions we needed to make to the data required to support them. From a technical perspective, we pushed to use the Airwallex ledger as the underpinning of our reporting model. We needed a global foundation that we could recognise and trust, and the reconciled ledger was the strongest place to anchor it.

Finance becomes the custodian of the source of truth

Across the functions within Finance we all have different data needs. Tax, Treasury, Controllership, and FP&A each came to this work with a clear view of their requirements and how they needed the data structured. Getting all of those perspectives into one definition was a challenging and impressive piece of cross-functional work, and the part of this I would most want other finance teams to copy.

We created a data dictionary describing every field and metric, settled the nomenclature, and agreed what goes in which table. We also went into the chart of accounts in our ERP and added a lot of columns to enable further data tagging and enrichment. Although nobody had asked for it at the time, we now use all of it downstream, and it anchors everything we are doing on AI.

This dataset is now the one source of truth across the business. And becoming custodians of the source of truth has turned Finance into an influence engine rather than a reporting one, and it makes every cross-functional conversation shorter.

Clean data gives AI the context it needs

Clean, governed data is what makes AI usable in Finance. An agent reading our Context Hub learns the business meaning behind each field so it reasons from our definitions rather than its own.

We’re already seeing the impact of this work in flux analysis. Explaining why operating costs moved in EMEA last month no longer means opening the ledger, chasing tags, and rebuilding the comparison by hand. An analyst can ask an agent running through Cursor, Claude, or Hex and the movements come back attributed. Because the underlying data is clean and the context is defined, I trust the outputs enough to act on them.

Of course, data foundation work is never a one-and-done exercise. It often feels like a game of tetris, fighting to solve existing problems fast enough to make space to handle the new ones. As the business grows, we continue to encounter new and complex data that doesn’t fit our schema. But that is a very different constraint from not having the foundations in place at all.

I’ve heard Finance leaders say that financial data is too sensitive to put in the hands of AI. In many cases, I think the bigger issue is readiness rather than security. If data is fragmented, inconsistently defined, or poorly governed, it is difficult to trust what AI does with it.

This groundwork we’ve done at Airwallex lets us invest in our AI transformation with confidence. It gives us a foundation we can continue to build on as we redesign how Finance operates.