Beyond the averages: How we’re using AI in Strategic Finance and FP&A

A finance team’s view of the business is partly shaped by what it has time to investigate.

Until recently, going deeper into a product, market, or spending line at Airwallex often meant another query, another spreadsheet, and another set of calculations to maintain. Our lean Strategic Finance and FP&A team had to choose where to spend that effort.

AI is easing that constraint in two ways. Our analysts use it to build tools tailored to their work. They also direct agents to investigate business questions, pursuing several lines of inquiry in parallel.

A foundation built by the people doing the work

We rebuilt our corporate data foundations during our Databricks migration. Finance led the design, and Engineering the technical build. We worked backwards from the decisions we needed to make, specifying the data and definitions required to support them.

We built a combined ledger that enriches accounting records with business context. We also strengthened our revenue data foundations and documented definitions in a data dictionary and context layer available to our AI tools. This gives teams across the business a common starting point, so we spend more time discussing what the numbers mean and less resolving whose data is correct.

On this foundation, our analysts use AI-assisted coding to build and maintain budgeting, flux, and forecasting tools. These combine financial calculations, data pipelines, and interactive charts, with AI helping to classify records and summarise movements for investigation.

The analysts use these tools every day. They see which exceptions need handling, which comparisons are missing, and what questions come next. They can then improve the tools themselves as the work evolves.

Keeping the detail in budgeting

Our finance-wide mapping of nearly 600 tasks helped us identify where to focus. Budgeting requires judgement and is difficult to automate end to end, but much of the preparation involves checking submissions, reconciling formats, and assembling comparisons.

That preparation was particularly demanding because we budget granularly. We want to know where each dollar goes and track actual spending against those decisions. Maintaining that view used to mean matching tens of thousands of records, refreshing queries, and rebuilding comparisons as the numbers changed.

Our team built a consolidated FP&A hub to handle this. AI helps match around 3,000 accounting entries to budget items each month and summarises movements for investigation. We can move from the full P&L statement into individual spending lines and historical trends without assembling another set of spreadsheets.

The hub also incorporates each new month’s actuals as they become available. In a budget process spanning several months, this lets us assess requests against the latest results rather than a snapshot taken at the start.

Our budgeting tool validates and consolidates divisional submissions as they arrive, handling differences in format that analysts previously reconciled manually. This gives us an earlier combined view of proposed spending and more time to work through the choices with teams. While the plan is still open to change, we can challenge why the spending is needed, consider alternatives, and agree how we would evaluate success.

A forecast can be right for the wrong reasons

AI also changes the investigation itself. Our analysts can direct agents to examine different cuts of the data in parallel, then follow up on findings without manually producing every comparison.

Much of our business is transaction-driven, spanning countries, products, and channels. Seasonality, the mix of weekdays and weekends, and differences between customer groups affect how a month unfolds. To keep forecasts manageable in spreadsheets, we often relied on broader averages. Our AI-assisted tools now let us examine these patterns more closely for the current month and quarter.

That matters even when the total lands close to plan. Stronger activity from existing customers could offset weaker new-customer growth. We could get the month right while missing a change that affects the quarters ahead. We use AI to investigate those underlying movements, then work with commercial teams to assess whether they are likely to persist and what action to take.

When an investigation produces a useful diagnostic method, we turn it into a reusable workflow: which data to examine, what to compare, and how to check results. Colleagues can rerun the method on fresh data without waiting for Finance to repeat the work.

Deciding what to believe

More analytical capacity gives us more explanations to assess. We need to establish whether a finding is credible and whether it matters for the decision.

Take acquisition payback: the time it takes to recover customer acquisition costs from gross profit. An improvement could reflect stronger economics within comparable customer groups, a shift in customer mix, or returns concentrated in a few unusually large accounts. Our tools make it practical to investigate across markets and channels, compare customers at similar stages after acquisition, and examine how returns are distributed. We can assess where the improvement holds and whether we could reproduce it with further investment.

However, greater detail also creates more opportunities to mistake noise for signal. Several findings may appear to confirm a trend while all being driven by the same few customers or transactions. Yet, an exceptional customer can both skew an average and point to an opportunity worth pursuing. We need to understand what made that customer valuable and whether we could find and serve more like them.

That raises the intellectual bar for our analysts, including junior colleagues. They need to frame questions around a decision, trace the reasoning, and understand what would have to be true for each explanation to hold. They must also seek evidence that could overturn their own conclusions. Checking the calculations remains essential, but it is only part of evaluating a recommendation.

Knowing when to stop investigating is part of that judgement. Before requesting another analysis, we should be clear about how its result could change our recommendation. If we are unsure whether a marketing channel will deliver similar returns at higher spending, we could test a limited budget increase and track acquisition costs and the revenues those customers generate over time. Those results would help us decide whether to scale, using new evidence rather than more scenarios based on the same historical data.

Expanding the strategic envelope

We are bringing our planning tools, forecasts, and diagnostics together to examine how investment choices interact. Shifting spending between markets or channels, for example, could change the customers we acquire, the capacity needed to serve them, and the returns we generate.

We already model these trade-offs. Our focus now is making more combinations practical to explore and revisit without rebuilding each analysis. We will compare actual outcomes with our original expectations, challenge the assumptions, and use what we learn to inform subsequent decisions.

We started by making deeper analysis practical for a lean team. I want to bring that same depth to a much wider range of strategic choices. The ambition is to uncover paths to growth we might otherwise never have considered, and give the business stronger evidence for deciding which to pursue.