What’s in a name? The real questions to ask of AI

In summary

• Much of the talk of AI in tax remains hype 
• While third party AI tax tools exist, outside tax tools tax research there is still limited use 
• But AI is being used in end-user models similar to the specialised Excel spreadsheets widely used to manage tax processes 
• Organisations should learn the lessons from these and act early to introduce policies and structure to control its use and manage the risks 

We are often asked to speak with tax leaders to help with an AI strategy.  

The first step is to deconstruct the question.  In these cases, “AI" very rarely means artificial intelligence. It generally means "anything interesting in the tech world”.  And "strategy" is typically shorthand for another question: "Should we be doing something?” 

Behind the question is excitement for what AI can bring – and a fear that the organisation is being left behind.   

AI in tax is an answer looking for a question. There is a lot to unpack before even discussing technology.   

Even the best bits of kit can only ever offer a partial solution.  Technology needs to be embedded in the process; it needs to reflect the requirements of the team; and it needs to be fed by data. This wider view is essential for creating effective solutions.   

The first step is often a whiteboard to map out processes, and this is, in fact, an area where AI can add significant value.  There are some excellent online whiteboarding tools, such as Miro, and tools such as CoPilot are adept at summarising discussions. Crucially, however, these are only ever an aid to human expertise and experience. 

Outputs often focus on blind spots, which are areas where the team doesn’t know what it doesn’t know, and time sinks, where they spend too much time on routine or repetitive tasks.  

The important point in this process is that the discussion on technology becomes problem-led. 

It is still easy to get fixated on solutions, however.  At first glance, AI is everywhere in tax.  There is a lot of hype in the industry.  Tax authorities have gone on the record about using AI to select audit targets.  Many tax software vendors, meanwhile, are unleashing AI-powered tools on the market, and many consultancies talk up custom-built solutions they say will transform how tax teams function.  The choice can seem bewildering. 

As with the AI market in general, there is a lot of froth around these solutions.   

The most successful AI tax tools are currently focused on tax research.  Tolley+ AI, Thomson Reuters CoCounsel and others that have launched in recent months make it much easier to bring together primary and secondary legislation, case law and other sources. This may change the nature of an advisory relationship, from answering questions to verifying answers.  Those teams using these tools report an incremental, rather than a transformative, impact, however.   

AI is present in other areas, of course.  Aibidia secured $28 million funding for its new AI-powered tax documentation solution, for instance, and there are effective AI and machine learning models in the market from several other providers. S&W has worked with many tax teams to help them get behind the marketing glitz and the tech hype.  We cover tax technology vendors across the spectrum to help clients make the right choice. The difficulty for many buyers is in discerning what is real – and helpful – from that which is merely effective marketing. 

At the moment, however, using third-party AI-powered solutions remains something other people are doing.  But that is only a small part of the story. 

A survey last year found that around two-thirds of in-house tax staff were using AI as part of their job.  Most of these are not using third-party software, however.   

Instead, the finding reflects the tax function’s long-term association with end-user computing models.  Every business has its share of highly complex, specialised spreadsheets to manage tax processes. Many are now adding end-user created AI approaches.   

Much of this focuses on automating general administration – summarising meetings, drafting emails or creating internal reports.  However, we also recently spoke to an infrastructure-heavy business that was using an LLM (large language model) powered approach to reading contractor documentation to distinguish between enhancement and maintenance expenditure. This approach was also interesting in that it was well documented and had built-in safeguards.

While this level of sophistication is unusual, there are some very interesting models being developed. 

That many are already harnessing the power of AI in this way is welcome, but businesses should beware of the lessons from Excel in tax: It provides great flexibility, but it can also introduce key person dependency, is error prone, and can end up creating hugely complex internal processes that expand well beyond their original purpose.   

Without a policy and structure to control it, end-user-led AI risks going down a similar path – and with additional concerns around the replicability and auditability of its outputs.   

Defining robust safeguards is a hugely important first step. Getting them right could help clear up a lot of the risks inherent in the existing spreadsheet models.