The promise of AI in finance is automation. Remove the human, reduce the cost, increase the speed. In many areas of financial services, this promise is being delivered. Fraud detection runs on machine learning. Credit scoring is algorithmic. Payment processing is fully automated.
Tax filing is different. And the reasons it is different are not temporary limitations of current AI technology. They are structural features of how tax systems work, and they are unlikely to change.
The Liability Question
When a tax return is submitted to a government authority, someone is legally responsible for its accuracy. In most jurisdictions, that responsibility sits with the taxpayer. But when a tax agent, accountant, or preparer files on behalf of a taxpayer, they take on a portion of that liability, and in many cases, they must be professionally licensed to do so.
In Malta, only a warranted Certified Public Accountant can file tax returns for clients. The warrant is issued under the Accountancy Profession Act and requires passing professional examinations, completing supervised practice, and maintaining annual continuing professional development. In Italy, the commercialista holds a similar role, regulated by the Ordine dei Dottori Commercialisti. In Germany, the Steuerberater is licensed by the Steuerberaterkammer. In the UK, while there is no legal requirement to be qualified to prepare a return, HMRC's agent authorisation process and professional indemnity requirements create a de facto licensing regime.
No AI system can hold a professional warrant. No machine learning model can be sued for negligence. No algorithm carries professional indemnity insurance. These are not edge cases. They are the foundational regulatory framework of tax filing in every developed country.
What AI Does Well
This is not an argument against AI in tax. AI is extraordinarily useful in the tax preparation workflow. It just cannot replace the licensed professional at the end of the chain.
Transaction categorisation is an area where AI excels. Given a bank feed of thousands of transactions, a well-trained model can categorise the vast majority correctly. It can distinguish between a business expense and a personal purchase. It can identify recurring patterns, like a monthly software subscription, and apply consistent categorisation. It can flag unusual transactions for human review.
Tax computation, once transactions are correctly categorised, is a rules-based process. It does not require AI at all. It requires a well-maintained rules engine that encodes the current year's tax rates, brackets, deductions, allowances, and thresholds. This is software engineering, not artificial intelligence. The computation should produce exactly the same result every time given the same inputs.
Document extraction is another strong AI use case. Scanning receipts, extracting amounts from invoices, reading PDF bank statements. These are pattern recognition problems where AI has reached near-human accuracy.
Where AI falls short is in the judgment calls that tax law frequently requires. Is this expense ordinary and necessary for the business? Does this arrangement constitute a trade or a hobby? Is the taxpayer entitled to this relief? These questions involve interpretation, context, and professional judgment. They are the reason tax professionals exist, and they are the reason tax professional licensing exists.
The Supervised AI Model
The solution is not to choose between AI and accountants. It is to combine them in a workflow where each handles what it does best.
AI handles data processing. It categorises transactions, extracts data from documents, computes tax using rules-based engines, and prepares a draft return. This step is fast, cheap, and scales to millions of users.
A licensed professional handles review. They examine the draft return, verify that the categorisations are correct, make judgment calls on ambiguous items, confirm that the return complies with current law, and sign off. This step is slower and more expensive per unit, but it is also where legal compliance happens.
This is the accountant-in-the-loop model. The AI does 90% of the work by volume. The accountant does 10% of the work by volume, but 100% of the work by legal necessity.
The economics work because AI reduces the accountant's time per return from hours to minutes. Without AI, an accountant processing a freelancer's tax return might spend two to three hours on data entry, reconciliation, and computation. With AI handling those steps, the accountant spends 15 to 30 minutes reviewing the output, verifying a handful of flagged items, and signing off.
This is how the model scales. The AI handles the volume. The accountant handles the liability.
Why "Fully Automated" Fails
Several startups have attempted to build fully automated tax filing. The pitch is appealing: connect your bank account, and we file your taxes automatically, no accountant needed.
The problems emerge quickly.
The first problem is errors. Tax categorisation is not a binary problem. A dinner receipt could be a business entertainment expense, a client meeting, a personal meal, or a team social event. The correct tax treatment differs for each. An AI model will make a probabilistic guess. In many cases, it will guess correctly. In some cases, it will not. The user has no way to know which is which, and the user is liable for the result.
The second problem is regulatory. Government tax authorities are increasingly scrutinising automated filings for accuracy. When error rates are higher than filings prepared by professionals, the automated channel attracts audit attention, which causes problems not just for the software provider but for every user who filed through it.
The third problem is trust. Freelancers filing taxes are dealing with their most consequential financial obligation. An error can result in penalties, interest, or investigation. Telling a user "our AI filed your taxes, trust us" is a different value proposition than "our AI prepared your taxes and a licensed accountant reviewed and signed off." The second version is more expensive to deliver but dramatically more valuable to the user.
The fourth problem is the one that cannot be solved by better technology: in many jurisdictions, it is simply not legal to file a tax return for a third party without professional credentials. No amount of AI accuracy changes the regulatory requirement.
How This Works in Practice
In a well-designed accountant-in-the-loop system, the workflow proceeds in stages.
Stage one is data ingestion. The system receives the taxpayer's transaction data, either from a connected bank account, an accounting software integration, or a direct API connection from a neobank or platform.
Stage two is automated categorisation. The AI model categorises each transaction according to the relevant jurisdiction's tax categories. A restaurant payment on a Tuesday with a client name in the memo becomes "business entertainment." A software subscription paid monthly becomes "office expenses." The model assigns confidence scores to each categorisation.
Stage three is computation. A rules-based engine takes the categorised transactions and computes the tax return. This is not AI. This is arithmetic with jurisdiction-specific parameters. The output is a draft return with exact figures.
Stage four is review routing. The draft return is sent to a licensed professional in the relevant jurisdiction. The professional sees the return, the underlying transaction data, the AI's categorisations with confidence scores, and any items flagged for manual review (low confidence scores, unusual amounts, potential personal expenses).
Stage five is professional review. The accountant examines the flagged items, verifies the high-confidence categorisations on a sample basis, makes judgment calls on ambiguous items, and either approves the return or sends it back for correction.
Stage six is submission. Once approved, the system submits the return to the government tax authority through the appropriate electronic channel and records the confirmation.
Stage seven is notification. The taxpayer receives confirmation that their return has been reviewed by a licensed professional and submitted to the tax authority.
The Economic Model
The accountant-in-the-loop model changes the economics of tax filing fundamentally.
Traditional accounting firms charge between 300 and 2,000 euros per return for a freelancer, depending on complexity and jurisdiction. The majority of that cost is the accountant's time spent on data entry, reconciliation, and computation, tasks that AI can automate entirely.
In a supervised AI model, the accountant's time per return drops to 15 to 30 minutes. At professional billing rates of 50 to 150 euros per hour, the review cost per return is roughly 15 to 75 euros. Add the infrastructure cost of the AI system and the per-return API cost, and the total cost per return is well under 100 euros.
This creates margin at every level. The platform offering embedded tax filing can charge the end user a fee that is significantly lower than a traditional accountant, while still generating meaningful margin. The accountant reviewing returns earns more per hour than they would doing manual preparation because the AI has eliminated the low-value work. The end user pays less and gets a faster, more convenient service.
Everyone is better off. That is the sign of a well-designed system.
Why This Matters for Platforms
For neobanks, gig platforms, invoicing tools, and other platforms considering how to offer tax filing to their users, the accountant-in-the-loop model is the only approach that is simultaneously scalable, legally compliant, and trustworthy.
Fully manual (traditional accountant) does not scale to hundreds of thousands of platform users. Fully automated (AI only) does not meet legal requirements in most jurisdictions and creates unacceptable error risk. The supervised model, AI preparation with professional review, scales with the AI component while meeting legal requirements through the human component.
The platform does not need to build any of this. It needs an API that accepts transaction data and returns filed tax returns. The AI, the rules engines, the professional reviewer network, and the government filing infrastructure are the API provider's responsibility.
The platform's job is to hold the user's data and present a seamless experience. Everything else is embedded.
Michael Cutajar, CPA — Founder of Accora.