Towards a zero-day close: How Controllership scaled AI without losing control

What happens when a strong data foundation meets the freedom to build with AI?

In my four years at Airwallex, our ARR has grown 13-fold, from US$100 million to US$1.3 billion. The most rewarding part for me is telling the story behind these numbers. I'm incredibly proud of my team for managing a much larger volume without significantly scaling headcount, thanks to continuous improvements in our workflows and tech.

We’re an AI-forward company and early adopters of AI, so that naturally created space for Controllership to embrace new tools and introduce AI into our workflows. In Controllership, our foundation is providing reliable and reconciled data that other functions of the business rely on. When finance data is reconciled by nature, we can then create reliable AI processes.

What mapping hundreds of thousands of hours revealed

Before we hand any work to AI, we need to first understand the work, the data, and where the friction actually lies. So we started by mapping it. Our recent baseline mapping exercise across the whole Finance team gave each function the opportunity to document our processes, their frequency, the time required, and the current and target states. This gave us really good visibility on how mature we were in terms of automation and where opportunities lie.

Controllership covers two Shared Service Centres (SSCs) and six accounting regions. The SSC centralises common, recurring finance processes that are standard across the global business, while regional teams focus on building relationships with local teams, regulators, and third parties, and on supporting all matters related to licensing and expansion.

The exercise surfaced three key findings that shape what we tackled next:

1. Recurring work is the clearest automation opportunity

The exercise tracked hundreds of thousands of hours of work each year, with Controllership accounting for the largest share. 86% of Controllership tasks recur frequently, highlighting a significant opportunity to automate and free up more time for strategic work. Even modest automation compounds when a task happens every day.

2. Regional teams are crowded out by routine work

Compared to SSCs, regional teams carry more annual and ad-hoc work, at around 40%. This is often strategic work around regulations and compliance, such as handling regulator requests and making local judgement calls. By automating routine work, we can free up more time for local controllers to focus on strategic work.

3. Technology has a shelf life

Systems that once worked can become bottlenecks. The procurement process was flagged as the most time-consuming task in Controllership because we’re using the Oracle procurement solution, which no longer fits our operating model. Our organisation is moving faster, so we're switching to our own products later this year to drive more automation and tackle some of the pain points we identified in the current process.

The real AI infrastructure is reliable, reconciled data

Once we understood our processes and opportunities, we returned to the data. Good AI processes are only as strong as the data behind them. Building the data foundation meant working backwards from the insights we wanted, identifying the data needed to derive them, and redesigning that foundation from the ground up.

Take our ERP rollout. When we implemented our ERP, we took a step back and asked: what impact do we want to drive, and what critical data points will help us downstream? With that in mind, we added a Chart of Accounts column that tags every expense, accrual, and cost to a specific vendor. That single addition made automated, granular variance analysis possible. Now, AI can autonomously execute sophisticated variance analysis, comparing actuals to budget by vendor and by division, driving efficiency and deeper analysis in our close processes.

We also created a Budget ID that links costs directly to budget line items. Every P&L cost record is now tied to a specific budget line, so each division can see exactly how it is tracking against budget, instead of relying on the typical bottom-up spreadsheet that loses granularity as data is aggregated. We can map down to the most granular level, giving us a reconciled view of actual spending, budgets, and P&L across divisions. That’s only possible because we have structured, reliable data at the core of our foundation.

AI stretches how far we can go

With our data foundation and AI in place, we’re moving from checking the numbers after the fact to continuously understanding, validating, and controlling what's happening beneath them. We’re moving toward a continuous close model, where reconciliation, reporting, and analysis happen in near real time.

The scope within Controllership is now far wider than before. The amount we can take on and self-serve has grown substantially in a short time. We used to take 10 days to close our books, and now we’re moving towards a zero-day close.

None of that compression would have been possible without the data foundation we'd just rebuilt. That foundation also changed who could build: people outside Data and Engineering can now build on it directly, and they did.

AI solutions built by finance, for finance

AI has greatly raised what a finance team can build on its own, and the biggest impact comes from the people closest to the work. With a strong data foundation and the freedom to build, team members began creating their own solutions and AI tools to reduce manual effort and improve processes. This sparked a wave of "citizen-developed projects."

For example, one of our senior accountants, with no engineering background, self-taught coding and cut a manual payroll consolidation process (which previously required pulling and anonymising 15 files) from two days to ten minutes. That’s ROI paid back within a month.

In another instance, a member of my team built AirQuire, which started as a General Ledger (GL) posting tool that allowed Mac users to post journal entries directly into Oracle, something Oracle did not support. Because someone who understood the finance workflow built it, the tool also auto-enriches transaction data as it's entered, checking that line descriptions and required fields are properly filled out and improving data quality at the point of entry. It did not stop there: my team member realised he could remove friction from the entire process and went beyond solving the initial problem. AirQuire now also supports quick entries, batch uploads, and pasting directly into a web table.

A centralised model with scalable wins

In finance, one of the hardest things is keeping processes simple and globally unified while still accommodating local compliance, tax, and reporting nuances. We were seeing strong results from citizen developer projects. But that same freedom meant three regions could spend weeks solving the same problem independently, unaware of one another.

That’s when I launched an AI innovation programme to bring everything back together through a central steering committee connecting all regions. We set standards for the documentation required for every AI program we launch. People need to write PRDs, which help structure our builds, roadmaps, and ultimately lead to higher-quality products.

Because of its success, we’ve scaled the programme across the Finance team to encourage more citizen developers. Our CFO, Pranav, also hired an engineer to support the transformation across the function, with stronger governance in place. But governance only manages what people already choose to build. It doesn't create the instinct to look for problems, and that instinct is harder to build.

The right mindset will raise the ceiling for AI

After the baseline exercise, we realised that the people closest to the work were starting to think build that instinct. Their domain expertise and visibility into their own processes put them in the best position to spot where the biggest opportunities lie within our function. That requires them to think more deeply and structurally about their own processes. That means questioning why certain processes exist at all and understanding the capabilities of AI. This is still a muscle that needs training and requires being open to conceptually redesigning things.

Even I sometimes think too linearly and need to stretch my thinking further. One challenge I’m tackling now is bringing everyone on the team along on this journey of transformation, to build a team and foundation capable of discovering what's possible. Building doesn't have to be everyone's job, but I encourage everyone to think like one. That's the mindset behind everything on our list.

The road ahead: 4 priorities we’re focusing on

There’s still much more that we want to achieve. To give you an idea of that, here are the 4 areas we’re focused on next on our roadmap:

1. Adopting AgentOS

Airwallex just released AgentOS, which connects any LLM to the Airwallex infrastructure. This unlocks a whole new wave of automation opportunities within the finance domain. For example, an agent now reads our HR data for new hires in our commercial division, automatically creates a user, and issues an Airwallex Corporate Card, all aligned with internal policies. We are also building agents that prepare intercompany settlements when balances (in our ERP) exceed our preferred maximum. Every settlement still requires human approval within Airwallex before the money actually moves. Next, we want to broaden our use of this connectivity by updating and auditing user access rights so they’re always aligned to our policies.

2. Closing and consolidating faster

We're building AI-driven flux analysis that can explain balance sheet movements between months, quarters, and years on its own. We’re also developing native reconciliation capabilities that automatically flag outliers against source data. Alongside that, we're moving towards one-click consolidation, automating the consolidation and elimination entries a team currently has to prepare by hand.

3. Keeping our reporting consistent everywhere we operate

We’re building a financial statement generator that will standardise the same accounting policy language, account mapping, disclosures, and layout across every local entity. On top of the efficiency benefits, this tool will also give our Group team control over all externally shared financial statements

4. Taking the manual work out of recurring compliance

We're building an AI tool to automate our global transfer pricing calculations across three pricing models, pulling directly from HR and ERP data instead of rebuilding the logic by hand each cycle. We’re also creating AI workflows that let users raise receipts, close purchase orders, and manage OPEX accruals in plain language through Slack.