How do I start using AI in Tax? A brief journey through AI experimentation using a real example

Let me start with a confession: I’m neither a pure tax technical expert nor a technology wizard. I sit somewhere in between, which, in practice, means I spend a lot of time thinking about what technology can do to help tax professionals.

After nearly 15 years in tax, and over a decade in technology-focused roles - I’ve enjoyed the benefits and felt the limitations of end-user, low/no code tools while in the weeds of bad data and complicated analytics. Most tools have provided real value over the period. The newest area of focus is AI - particularly the use of Large Language Models (LLMs). I’ve been exploring this area for a little while and I’m excited!

So, how do you get started?  I'm writing this deliberately as if I'm exploring using an LLM for the first time alongside you. I often find new technology goes through a period of gatekeeping and, with the rate of change in AI at the moment, we are all still figuring this out. Think of this as a first-time walkthrough designed to help you break those initial knowledge barriers that make AI feel intimidating or overly complex. I have tried to use some specific examples in this blog that highlight AIs ability to deal with the ‘grey areas’ of tax. 

Note: This blog isn't following a scientific methodology, and I'm not claiming that what we discover will work in your organization, nor am I offering you too much by way of scalability. This blog is for illustrative purposes only and does not constitute tax advice. Always validate outputs with appropriate human review.  

Starting simple: One-shot prompting

Most of us ask a simple, typically single, question such as ‘Help me come up with a recipe using chicken, rice and vegetables’. This is known as "one-shot" prompting when using tools like ChatGPT. It works fine for many tasks and can be genuinely helpful.

But how does this basic approach work with a tax categorization challenge? Let's start with a single transaction and see what happens. Here's my first attempt with a straightforward prompt to provide the UK employment tax treatment of an employee expense. I will use this example throughout for consistency, but it could easily be a corporate tax analysis of the legal and professional ledger, or the VAT analysis of a purchase invoice. 

Why have I chosen this transaction? Because it was always one that classical rules - even those with keyword analysis - found difficult because it contains contradicting features (“team” and “client”) that needs interpretation. 

My prompt:

I am unsure whether this transaction should be included in the UK PSA return. Please analyse it and give me an answer: 

Vendor: Gourmet Burger Kitchen

Amount: £175.00 

Description: Team meal after client demo 

Category: Meals 

Location: London

AI response:

This £175 team meal at Gourmet Burger Kitchen, held after a client demo, can be included in the UK PSA return if it meets HMRC’s criteria: the expense must be irregular, non-contractual, not part of the employees’ salary, and not reimbursed as a business expense. As the meal was for staff (not client entertainment) and appears to be a one-off celebratory event, it qualifies as a minor, non-routine staff benefit. Provided it wasn’t processed through payroll or previously taxed, it is appropriate for PSA inclusion.

Not bad! Maybe a bit verbose. Maybe missed some possible exemptions. How do we make it better?

How could an AI read tax information and respond? 

Before we get into engineering techniques, it's worth understanding a bit about how these AI tools actually process our requests. This will be a basic overview to help you understand why some approaches work better than others.

In very simple terms, when you give an AI system your request (your "prompt"), it breaks down your text into key components called "tokens". This is mostly, but not always, a token per word. It maps each token to a number in its knowledge base. It then tries to map these tokens to patterns it learned during training to understand the prompt and predict the best possible response. What does this mean in practice? 

For our tax analysis, this means the AI is looking for patterns like “restaurant” + “employee” + “celebration” = “potential staff entertainment” or "hotel” + “training course” = “allowable business expense" to provide what it thinks is the most correct answer. You will often hear the phrase, ‘it’s not intelligent, it just guesses the next character’ which is a simplistic criticism of how AI uses patterns, but not wholly incorrect.

Why is it good to understand this? Because you can then write your prompts in a way that maximise the chances of picking up the right patterns.

Making your requests more effective

The process of improving and refining your prompts is called ‘prompt engineering’, and it is the easiest way to make a big difference to your results. There are many different best-practices on how to engineer your prompt - OpenAI’s summary and guide is here. Often, these are things you can copy and paste into each prompt for ease.

For this blog, I will focus on a couple:

Personas: Telling an AI to ‘Think-Like-a-Tax-Professional’

This doesn’t quite mean your AI will sometimes act ‘hungover’ on a Friday morning - that is still the preserve of the first year graduates (I think? It’s been a while since I was a first year graduate). Using a ‘persona’ here means getting the AI to change how they approach a task based on how you ask them to behave. This can impact the answer you get.

Continuing with our example, you may ask your AI to take the following “persona”: ‘You are an experienced tax advisor reviewing expenses for PSA analysis. Maximise the use of assumptions that lead to non taxable results’:

Result:

“Assuming the meal was a one-off, informal team gathering not part of any contractual obligation or salary sacrifice arrangement, and no clients attended, this expense likely qualifies as a trivial benefit or staff welfare. If not reimbursed through payroll and not part of a routine reward scheme, it is non-taxable and need not be included in the PSA return.”

How about a more cautious persona? ‘You are an experienced HMRC inspector reviewing expenses for PSA analysis using HMRC guidance.’

Result:

Based on HMRC guidance, this £175 team meal following a client demo is likely a taxable benefit. As it was not part of business duties, involved only staff, and exceeds the trivial benefit threshold, it should be included in the PSA return—provided it is irregular, non-contractual, and not processed through payroll. If regular or a reward, it may not qualify.

As you can see, the difference is significant. This suggests we can tune AI behavior to match an organisation's risk appetite just like we do with human reviewers. Alternatively, and perhaps more interestingly, we could use multiple personas to analyse a single transaction to identify where the analysis is open to interpretation. This would mean we understand where analysis is open to more risk which can then be quantified. 

Hallucination: Preventing AI from making stuff up

One of the biggest concerns with AI is "hallucination". This is when it produces fabricated information with complete confidence. In tax work, this could be disastrous. Fortunately, there are techniques to reduce this risk.

There are a few ways you can do this:

• Reference material: You can point AI towards specific guidance documents or websites to ground its analysis in actual tax law rather than general knowledge or build up the knowledge base by giving reference material. This helps ensure the AI is working from the same reference materials you would use.

• Hard rules: You could also use very explicit rules minimising the amount the AI needs to contextualise a transaction. Using the above example, you may want the AI to always ask for the number of staff who attended before giving an answer. This could also apply to other tax questions e.g. always clarify the “ship to” location for international shipments. You can even combine your classical rules / code into the prompt, if you have it. 

• Think slowly: Analysis shows that LLMs can think better when asked to slow down. 

• Simplify: Instead of asking your LLM to give you the final tax answer, give it a smaller task such as summarising, contextualising or grouping the transaction. This then allows you to leave the tax analysis to other people, rulesets or software.

• Format: You can clarify the output format you want to achieve. You can also clarify the format of any inputs, which reduces the risk of misinterpretation and/ or the wrong information being referenced.

• Reasoning requirements: Instead of just asking for an answer, you can ask the AI to show its working. This is even more helpful when considering audit trails, etc. How could this look in practice:
  

Analyze this expense for PSA purposes and provide: 

1. Summarise your understanding of the expense

2. Your recommendation (Taxable/Non Taxable/Uncertain) 

3. Your reasoning based on UK tax law 

4. What additional information would help confirm your analysis 

5. Your confidence level (High/Medium/Low)

Few-shot learning: Teaching with examples

The basic one-shot prompt can deliver improved analysis using some of the techniques above. Another way is called "few-shot" learning which means giving the AI a “few” examples of good analysis to guide its future responses. 

If you are facing a large data task today, you could start adding labelled batches of good (and bad) results to your prompt when you have an analysis you’re happy with. 

This could also be helpful if, for example, you have specific tax positions or if you have a library of scenarios you want an AI model to leverage.

Conclusion

There’s an enormous amount of information around AI at the moment  and it can become complicated quite quickly. For those of you who haven’t yet looked at AI or wanted to but didn’t know where to start, I hope this has been useful. I encourage those who haven’t experimented with AI to do so (safely) to see what it can do to help you in your day to day. 

If I was sitting next to you exploring this for the first time then this would probably be the point we get a coffee and start brainstorming what other things AI could help with. 

There are a number of areas to continue exploring how AI could integrate and support your tax workflows, processes - including Agents, APIs and GPTs. Something for a later blog…