How large language model Artificial Intelligence can help get VAT right
Generative Artificial Intelligence based on large language models offer exciting opportunities to enhance the accuracy and efficiency of global VAT determination. And harnessed with carefully curated VAT or GST legislative databases, it offers the opportunity to simplify some of the most complex components of indirect tax calculations. This is very exciting since the nature of VAT – millions, even billions of transactions – present the colossal levels of data that AI thrives on to learn from and refine its responses towards better accuracy. This can be seen as tax authorities adopt AI for the detection of fraud.
At VATCalc, we are early adopters of AI to enhance our already unique deep content model for a single tax engine, VAT returns and VAT e-invoicing application. You can see how we already deploy it to produce AI VAT advice and for item classification to get VAT rules and rates correct.
How do generative AI chatbots work?
OpenAI’s ChatGPT and similar chatbots follow the same basic steps, using natural language processing, to create humanlike conversational dialogue:
- Read the entire internet and other data supplied with some basic guiderail rules
- Gives each word a 4-character ‘token’ (there are about 500 billion tokens in GPT-3) – tokenisation – and records its relative ‘location’ to other tokens within the text
- Uses probability to determine the likely position of each word it’s relation to other words based on these tokens and locations
- Builds a relationship map using neural networks – a series of algorithms that tries to recognize relationships in a set of data through a process that mimics the way the human brain operates. These are also used in chatbot’s AI cousin, machine learning.
- When supplied with a user question, ‘prompt’, the chatbot will predict the answer based on its neural network, word tokens and locations.
- This can then be improved on based on user corrections or further prompts. This is the big leap, that it remembers what you have already asked, and use that to inform its conversation with you.
So the chatbot has no inherent knowledge or expertise. It is using probabilities based on the known internet. So it picks-up on missing or erroneous information. It can then exacerbate the gaps by attempting to make-up or ‘hallucinate’ the answer with a conviction that tricks confidence in the use.
What problems does this approach create for VAT?
Chatbots cannot police the web content they depend, and, where they find gaps, use and gaps they fill. ChatGPT is a vehicle for transmitting the wisdom of crowds, but it has no wisdom itself. It has no semantic component, being able to understand the context the topic. It is all rhetoric and no logic. It is delivering the optimal rhetorical answer to your prompt, not the most expert one.
Add to this the web is not particularly up-to-date. And ChatGPT often have used frozen versions of the internet from up to two years previously. This means any analysis or answer is almost certainly based on old, out of date base information.
That is fatal for the outcome and confidence of the technique for something as financially sensitive as VAT. Wrong tax calculations means cash flow losses, audits and penalties.
What risks does AI present for VAT and tax?
As we all come to grips with generative AI, some of its limits are already becoming apparent. These include:
- Errors – Who is responsible for any error or ommissions? God luck attempting to sue a chatbot for not understanding the domestic reverse charge ‘may’ provisions.
- AI skills – Is AI able to bring in client issues and context. This depends on the questions, or ‘prompts’, the user address to the chatbot. There’s a real skill to be learned that may take some more years to master on how to interrogate chatbots to ensure the tool knows what to look for. At VATCalc, it’s been our experience so far that this is were the important skill is needs – typically someone with a strong VAT background who understand the answers or holes that come back from the prompts.
- Completeness – Is all of the VAT legislation, implementing regulations, guidance, tax cases etc available online for the chatbot to interrogate? What about actually practise where the tax authorities are informally accepting certain treatments as workarounds?
- Up-to-date? Related to the above, most Chatbots being offered today are based on the information available on the internet at set times in the past. They aren’t refreshed – the great strength of the internet. So you won’t get results based on the latest VAT rules, interpretations or common practise. Fatal!
- Client confidentially – What information can you give without fear of divulging your clients identify and private financial affairs. While European data laws protect a person’s personal data, whether they post that info publicly or privately.
- Taxpayer trust – it may take years for taxpayers to come to know and trust chatbots for financially sensitive tax issues. The tax profession – like law – is known for (fairly!) well-regulated, educated and trained providers. Who oversees the bots?
AI and VAT to become besties!
But what if, instead of the soup of the global internet, generative AI is used in conjunction with carefully created and maintained databases – so no gaps, out-of-date information and hallucinations.
This is the approach we have taken at VATCalc, using AI chatbots as a plug-in to our tax legislative driven application. It can help make better selections of the components of the calculation, or write-up clear explanations of the outputs of calculations to share with colleagues and counter-parties.
AI and VAT will be besties. Sign-up to our VAT news to learn about how we’re making that happen at VATCalc.