Part of the SYNAPSE 2026 series | Panel: Parmi (Amazon), J Gu (Meta), Josh Reardon (Zoom)

Getting AI into production in a tax function is not primarily a technology challenge. It is a data challenge. The models exist. The use cases are clear enough. What stops most organizations is the state of the data sitting underneath the workflows they want to automate—and the realization that no amount of AI investment fixes a data quality problem.

That tension was at the center of one of the most practical sessions at SYNAPSE 2026, Fonoa's annual conference for indirect tax professionals. Parmi Makh from Amazon, J Gu from Meta, and Josh Reardon from Zoom each brought a different stage of the AI journey to the conversation, and between them painted a realistic picture of what it actually takes to get there.

1. You have to start before everything is ready

At Amazon, a tax rate was wrong on one product. The same classification logic applied across an entire catalogue, which meant a single incorrect call replicated at scale became a significant operational problem. The team faced a straightforward question: is the right solution for someone to work through this manually in a spreadsheet?

It was not. But the path from that realization to a working ML classification model with 99.6% accuracy was not a clean one. There were no data scientists on the team at the start.

"We just had to start. We didn't have data scientists. We got in a room, borrowed resources, joined office hours, found a groove and a direction."
— Parmi, Amazon

Building capability while solving the problem

The team eventually hired data scientists and built business intelligence capability alongside the model. Accuracy reached 99.6%, and they were still not satisfied. Customer questions and complaints provided the ongoing signal for where the model was still getting things wrong. The standard kept moving.

The lesson is not that Amazon had a clear roadmap. It is that they committed to the direction before they had all the pieces in place and built incrementally from there.

2. AI is intelligence, not just tooling

J Gu from Meta reframed what AI represents in a tax context:

"If you think AI is a tool, that's old news. It's intelligence."
— J Gu, Meta

Execution speed and scaled output

The practical proof of that at Meta is in execution speed. Feature requests that previously took weeks now take days. The capacity of the tax function has grown without growing the headcount, which is what the business actually needed.

Human oversight remains non-negotiable

But Parmi's response to that framing provided the necessary counterweight: "You can leverage it, but it can be confidently wrong. So you need to check." Amazon still layers human review on top of ML classification. Meta still manually reviews AI audit outputs. Zoom used AI to identify PO numbers on supplier invoices, freeing up meaningful time, but within a monitored workflow.

The throughline across all three companies is that AI has expanded what their tax teams can do. It has not replaced the judgment those teams apply to the output.

3. The job is shifting from doing to managing

The clearest articulation of where this is all heading came from Parmi's description of the transition underway at Amazon: from tax doers to tax managers. The role stops being about processing transactions and starts being about overseeing the systems that process them, applying judgment where it is genuinely needed, and making decisions that no model can make on its own.

Josh Reardon from Zoom described a similar evolution. Automation at Zoom handles the work that used to consume significant manual effort, like:

• Finding the right code on a supplier invoice
• Validating a tax ID
• Matching records across systems

The tax team's attention goes to the work that actually requires it.

Data quality as the limiting factor

This shift only holds up if the data underneath it is reliable. AI applied to poor data does not produce poor results slowly. It produces them at scale, and with confidence. The data foundation is not a prerequisite for starting. It is the thing you are building toward, continuously, as the program matures.

Bottom line: The AI strategy and the data strategy are the same thing

The companies making real progress with AI in tax, like Amazon, Meta, Zoom, are not the ones that found the best models. They are the ones that treated data quality as the core of the programme from the beginning, started before they had perfect conditions, and maintained human oversight as the work scaled. That combination is available to any organization willing to commit to it.

‍