Bold, not perfect: How Tax rebuilt itself around AI

Tax is uniquely complex. Does AI help or complicate things further? 

Tax is one of the most complex areas of finance. It requires interpreting the principles behind regulations, and then assessing how to align them with the business while creating efficiencies. Airwallex operates in 26 countries now, which means we need to comply with 26 different sets of rules and adapt to the changes that happen each year.

While tax is traditionally seen as a manual, compliance-heavy function, the reality is that it is leading the field in autonomous AI. In a recent study we commissioned from Forrester, we found that 15% of tax teams are already executing AI workflows autonomously, more than most other finance functions. This means that tax automation is the new baseline, and it requires a thoughtful approach.

Lessons from our Finance baseline mapping exercise

As part of our AI transformation, we mapped every task across the Finance department. We conducted a full task-level mapping exercise across Tax, scrutinising the data sources, actions performed, employees involved, and outputs generated. The exercise gave us clarity on our processes and opportunities, and here’s what we learned:

1. Automating tax compliance offers the biggest opportunity

Up to 75% of our tax compliance tasks can be automated, and there will always be a degree of human judgement and decision-making required. That 75% matters more every year, as the volume of compliance work will continue to grow with the business. The more we automate now, the more headroom we create for later.

2. Regions need to standardise before automating

We realised that each region has its own way of doing things, and that we need to standardise. Tax depends on financial information from the controllers, and since the controllers in each region are on different teams, each data point comes from a different process. We needed to agree with the controllers on a single process and a single dataset for the final state before we could start automating. This is not an easy thing to achieve but is well worth the effort.

3. The underlying data work had to come first

The regional fragmentation above was one symptom of a bigger pattern: across the board, automation maturity usually came down to the quality of the underlying data. Before we could automate anything, we had to do a deep dive into that data and get it right. That meant a lot of effort spent cleaning up the entire tax data foundation.

AI changes how tax actually works

The learnings from the baseline exercise gave us a clear direction on where to focus our efforts. We did the groundwork that gave AI a clean, consistent data foundation to work with. With that in place, the Tax team has removed inefficiencies in our processes. So far, our work has mainly changed in two ways:

1. Automate tax compliance

We’re in the process of automating the full cycle of tax compliance, including tax return preparation and review, as well as controls over them. Removing manual processes frees us up for the value-added part of the tax function: tax planning, tax strategies, transfer pricing policies, and solving larger technical problems. These are parts of the job we couldn't get to when 50–70% of our time was spent on recurring tasks.

Take VAT/GST returns, for instance. We built pipelines to pull data from AP/AR sources, standardised the process across our large jurisdictions, and have now successfully executed our first automated parallel runs alongside the old way of doing it. VAT and GST returns are now built almost end-to-end, where the system calculates VAT liability and recovery ratio, then auto-fills a draft of the tax return template. We give that to our external tax advisors, whose role has shifted from preparing drafts to reviewing what we’ve produced.

2. Advise the business as a partner, much earlier

AI has also changed the way we advise the business as a function, allowing us to be a partner at the forefront. We used to rely more heavily on external advisors for their privileged access to technical knowledge, market practice, and cross-industry experience, which was difficult for in-house teams to obtain independently. Now, the information asymmetry has been reduced, thanks to AI.

AI has dramatically cut the time required for tax research, information gathering, and initial technical analysis. We use it to better understand tax laws across jurisdictions and to create first drafts that we previously relied on external advisors for. It's the same shift we made with our external tax agent on VAT and GST: advisors review our work now, rather than producing it for us.

We’ve freed up time for the team to be leading the advisory work, and now we're called at the beginning rather than just to sign off on a deal at the end. We need to be able to act on that trust fast, so we started building our own systems and solutions. That meant rethinking the type of tax professionals we want in our team.

Tax experts who code and make this a reality

Tax sits at the apex of legal knowledge, accounting, and finance. The type of professional we need is a unique, imaginative person with a drive for data and process.

The reality is that the "tax technology" experts that exist out there aren’t usually of this profile. They usually fall into two groups: engineers or data people who’ve learned a bit of tax but lack the core tax technical knowledge to build solutions themselves, or hybrid compliance professionals from Big 4s who now focus on automation but don’t actually know how to build anything.

The tax experts here at Airwallex are different. Rather than bringing in an engineer, we hired a tax professional with more than 12 years of experience to focus only on automation and AI. This person is guided by another Tax Senior Manager and me, since we both have extensive experience automating processes. Even though he has no full-time experience in automation or technology, his curiosity and drive to change things were key reasons we hired him.

Having domain expertise is non-negotiable. If you don't understand what you're trying to build, you'll build something that doesn’t work. AI hallucinates, and automation breaks. That’s why you need people with technical knowledge and experience to make the calls, to challenge AI when needed, and to solve problems when AI workflows break.

The case for moving fast: ship it, break it, rebuild it

Before AI, we wouldn't have built anything ourselves. If there was a problem, we’d have to outsource it, buy software, or bring in an advisor. Now, we figure it out ourselves. We started building our own apps and creating our own workflows. For example, one of my team members built a transfer pricing documentation review agent that reviews documentation against OECD and local principles, and flags gaps before human review.

Not everyone should be vibecoding, but we need people to shift their mindset and be open to building their own solutions and AI tools, rather than waiting for someone else to do it. As tax professionals, we tend to be very process-driven and perfectionist, which makes that shift harder than it sounds.

We need to build fast and just start shipping and testing. We're never going to reach a perfect version of any product, but we'll never get anywhere if we don’t start trying. And we always ensure there's a safety net. For example, in data extraction, we made significant enhancements by keeping the historic column in the SQL and adding a new column. If something breaks, we still have the historic state to fall back on.

Our priorities as we move forward

With a strong data foundation, the right tax professionals, and a set of core principles that lets us move fast, we’ve identified our next priorities:

• Go global with our indirect tax solution

• Automate how we create and review transfer pricing documentation

• Automate core recurring tax advisory projects (e.g. our review of every FP and Enterprise deal that goes live in Airwallex)