Tax in the compute era

India’s goods and services tax portal processes around 100 million invoices a day. Brazil's NF-e electronic invoicing system has cleared more than 100 billion documents cumulatively. By 2030, the EU’s ViDA digital reporting framework will move tens of billions of cross-border transactions under real-time tax authority observation. Global commerce is becoming visible to tax authorities in real time by default.

The humans responsible for that data don’t have the same real-time visibility. Tax teams reading, researching, reconciling, and signing off on every global transaction are the same size or smaller than they were ten years ago. They are doing the same work, from the same spreadsheets, under pressure to deliver more with less.

This is structurally impossible. On one side, real-time reporting to tax authorities everywhere. On the other, after-the-fact human reconciliation in small teams. Compliance in today’s world of e-invoicing demands transaction-level accurate tax data, continuously, everywhere a company sells. Most organisations cannot produce it, and the gap widens every quarter. No amount of extra headcount will close it, because the work has crossed a volume and velocity threshold that human processes cannot meet. You can’t fix a machine-scale problem with a human-scale process.

The fix is automation. And AI unlocks new possibilities. The interesting question is how. The answer will reshape tax work, the service provider landscape, and the tax function itself. Within five years, every in-house tax team, every tax technology provider, every professional services practice will be fundamentally repositioned by this shift.

If that transition plays out, what happens to the people doing tax work today?

The intuitive answer is fewer of them, doing less. That answer is wrong. The opposite is true: more of them, doing more, with higher stakes per head.

Nvidia’s founder, president, and CEO, Jensen Huang, makes this argument well about radiology. A decade ago, radiology was supposed to be the first profession AI replaced. That prediction was wrong on structure, not just timing. More AI did not mean less demand for radiologists. It produced more scans, more referrals, earlier interventions, and more radiologists employed at higher productivity. The job expanded because the unit cost of the capability dropped.

Satya Nadella, the CEO of Microsoft, tells the same story from the economic side. Every prior compute shift expanded the total pie, from mainframes to PCs and from PCs to cloud. When the unit cost of a capability drops, consumption does not stay flat. It rises.

Economists call this Jevons’ Paradox, and it is the most under-discussed force in the AI transition. It is also the one most likely to surprise the tax industry over the next five years. Here is why it matters. Today, huge amounts of tax work go undone because they cost too much to do. Recoverable VAT goes unresearched because the analysis is too expensive, internally or externally. Nexus exposures go untracked because the exercise takes time nobody has. A regulatory change in Malaysia that should have triggered a configuration review last quarter never did, because nobody had the bandwidth.

Remove the cost constraint and demand does not shrink. It expands into every question that was not worth asking before. The research-and-lookup layer becomes a commodity, and everything above it grows more valuable per hour: judgement, controversy, sign-off, relationships.

There is a second reason the tax professional's role grows in importance rather than shrinking. As long as humans are legally responsible for a business, humans sign off on its decisions. The in-house tax director signs the return. The tax partner signs the opinion. The board signs the statements. The law does not accept a machine as the accountable party, and the accountable human wants assurance that the automated work is sound.

AI amplifies that accountability rather than removing it. If AI creates ten times the leverage, the same human is now responsible for ten times the output. Each process carries more transactions, more decisions, and more exposure. The stake behind every signature goes up, not down. That is why the human role concentrates upward instead of disappearing, and why the review and control layer around AI becomes the most important part of the stack.

Enterprise AI has rolled out more slowly than the headlines promised. Replacing human work is hard in every industry, because most of that work runs on rules and institutional knowledge nobody ever wrote down. A senior indirect tax manager making a recoverability call draws on 20 years of pattern recognition across jurisdictions, customer types, and audit outcomes. None of that sits in a training set.

So AI has not replaced (as) much of the tax workflow yet. Some people take that as proof no transition is happening. They are wrong, in the same way people were wrong in 1996 when they pointed to all the cash still changing hands and concluded digital payments would never take over. The direction of travel was obvious then, and it is obvious now. The transition is real; it has just been slower than the loudest voices claimed.

The real work ahead is embedding AI natively across every enterprise tax and finance stack. Not as a feature or a chatbot, but as an operational layer sitting on top of deterministic tax infrastructure, underpinning every decision and every document the tax function produces. If humans are going to take personal liability for AI-produced work, the systems producing it need traceability back to source data, deterministic computation on structured transactions, and testable behaviour against ground truth. Human oversight does not go away. It moves upstream, to the review of machine-produced work and the judgement calls at the top of the workflow. The person with AI leverage pulls ahead. The person without it falls behind.

Tax infrastructure is not a single product. It is a layered ecosystem.

Source systems come first: ERPs, billing platforms, commerce engines, marketplaces, and payment rails. This is where transactions are born. Above that sits tax infrastructure: data validation, certified e-invoicing, and indirect tax engines. This layer turns a transaction into a tax-aware transaction. Above that is workflow execution: return preparation and review, anomaly detection, reconciliation, filing, and remittance. At the top sits regulatory intelligence, which monitors and researches rule changes and keeps everything beneath it current as the law shifts.

Today each layer has its own vendors, its own workflows, and its own data. None of them talk to each other natively. That matters because AI only works well when they do. A language model cannot reason usefully across a landscape where ‘customer’ means one thing in the ERP, another in the determination engine, and a third in the e-invoicing tool.

Fragmentation is the real ceiling on the value AI can deliver in tax. Great foundation models will not compensate for a fragmented data layer. They will expose it. This is reinforced in many of my conversations with customers; lately, it’s been a common topic with Ankeet Mehta, the global tax technology lead at Nvidia – check out Fonoa’s webinar, The New Indirect Tax Technologist, where it will be discussed.

Ask an AI to reason across fragmented data and one of two things happens. It hallucinates, filling the gaps with plausible fiction, which is catastrophic in compliance. Or it hedges and refuses to commit, which makes it useless. Garbage ontology in, garbage AI out.

So the tax technology that wins the next decade is not the one with the most advanced AI on top. It is the one with the cleanest and most comprehensive data model underneath: a single shared language across the full indirect tax life cycle from validation, determination, e-invoicing, returns to audit, with one integration layer into the source systems where transactions originate.

‘AI for tax’ is being sold today by three types of vendor under the same label. Each carries a different structural risk that it is helpful to be aware of.

Model wrappers are applications built on top of frontier LLMs and trained on a tax corpus. They demo well and they cite guidance. But they own neither the reasoning layer nor the data layer, and both get compressed as foundation models improve. HarveyAI is the most instructive example from adjacent legal tech: it started as a custom trained model, but more recently publicly abandoned custom model fine-tuning in favour of using general frontier models, because the labs’ capability curve outpaces anything a vertical application can build on its own. Instead, they focus on workflows. That was a rational call, and the same fate awaits tax model wrappers.

Legacy incumbents struggle on data architecture. Built over decades through on-prem legacies, cloud retrofits, and acquisition roll-ups, their data is fragmented because their products were separately acquired. You cannot bolt a unified data model on to a stack that was never designed for one. The fix is a full rebuild, and no listed incumbent will undertake a multi-year, customer-disrupting rebuild of its own platform in front of public-market investors or private equity owners. The result is simple AI features that never go beyond the feature level.

Point specialists are excellent at a single step or geography – such as e-invoicing only, or determination only – but absent from the rest of the life cycle. Indirect tax is a connected chain. And it is getting even more connected through government initiatives. In Brazil, for example, there is government visibility on reconciliation from ledger to e-invoice to statutory accounts already. A determination affects the invoice, which affects the filing, which affects the recovery, which affects the audit defence. Break the chain and every downstream step reverts to manual reconciliation. In a world of after-the-fact processing, this was barely doable. In a real-time reporting world, that is impossible.

All three share one outcome: the AI layer is gated by the infrastructure beneath it.

Even perfectly built AI cannot outrun the structure underneath it. In 1967, Gene Amdahl, the renowned computer architect, showed that you can only speed up a system as much as its slowest part allows. If 10% of your workflow still needs a human to reconcile disconnected systems by hand, that 10% caps your end-to-end speed-up no matter how much AI you throw at the other 90%.

This is why ‘best-of-breed’ stops being a compliment. Five best-of-breed tools stitched together by a human are slower than one coherent platform that is merely ‘good’ at each step, because the human doing the stitching is the bottleneck, and so are the seams between the tools.

Four properties, together, define the next tax platform:

• A single transactional data model across the life cycle, because AI is only as good as its data, and fragmented data produces fragmented reasoning.

• A connected infrastructure with preserved lineage, because tax outputs must be audit-defensible. Every output traces back to the input transaction and the rule applied.

• Foundation models embedded as a function rather than sold as a product, because the capability curve belongs to the frontier labs and the sensible move is to ride it, not race it.

• Regulatory intelligence that writes into structure rather than into PDFs, because the value of a rule change is not a research memo. It is the configuration update sitting in the determination engine on Monday morning.

These four complement each other, and no three of them are enough. A system with structured data and lineage but no regulatory intelligence lags the law. A system with regulatory intelligence but no data model cannot act on it. A system without embedded foundation models falls behind every time the labs ship a new generation. The next tax platform has all four, or it is not the next tax platform.

Can a tax function build this in-house? The classic build versus buy principle still holds. As Nvidia’s Jensen Huang puts it: “do as much as necessary, and as little as possible”. Build where you have a specific and lasting edge. Buy where the market already offers a good answer.

The indirect tax technology market still lags, but we know what the right answer looks like: connected data, embedded AI, and regulatory intelligence that updates the engine rather than the inbox. Trying to rebuild that in-house on top of an existing ERP is the same mistake as building your own payment rails because you do not trust an external provider. You can do it. In the past you sometimes had to, because the market offered nothing better. That is changing fast.

The in-house build that still makes sense is the orchestration layer on top: the workflows, controls, and business rules unique to a specific company, plus the integrations into proprietary upstream data. Everything below that is infrastructure, and infrastructure is a buy. The build-versus-buy conversation most tax functions need to have in the next six months is about the data layer, the orchestration across their stack, and how to give AI something to work with. Too many companies are focused on the tooling and not the data.

The deepest shift underneath all of this is that tax is becoming a data-driven risk-management discipline.

For most of its history, tax has been a legal discipline built on hypothetical scenarios, reasoned opinions, and qualitative probability weightings rendered in prose. That has real value and it is not going away. But it is shrinking relative to a new world in which positions are supported by structured transactional evidence, weighted by probability, monitored continuously against regulatory change, and defended with lineage an authority can read directly.

Tax authorities are already there. They are deploying transaction-level monitoring through real-time e-invoicing, continuous transaction controls, pre-filled returns, and data-matched audits. Every major administration is heading this way. A qualitative approach will not hold, and a legal-only defence against a data-native authority is a losing posture.

That is the transition under way: slower than the headlines promised, and faster than most are ready for.

That is why Fonoa has been building this architecture for more than seven years: a single transactional data model, a single integration layer, and the full indirect tax life cycle on top. The company did not build it because it foresaw the AI moment. It built it because running global tax for the world’s largest transactional businesses required it, AI or no AI. That this is also the one architecture on which real AI value can be delivered is not a coincidence. It is the structure of the problem.

Humans in this ecosystem will do more work, not less, because demand for tax certainty compounds as the unit cost of producing it falls. The technology layer converges on whoever has the cleanest data, the richest ontology, and the right posture towards the frontier labs.

We are past the question of whether AI changes tax. The interesting question is what the new equilibrium looks like: a connected, intelligent, continuously current tax operating layer beneath the enterprise, with a smaller, higher-leverage set of skilled humans on top, doing the work only humans can do.

That is the ecosystem Fonoa is building.