error
Subscriptions are not available for this site while you are logged into your current account.
close
Skip to main content
Submit RFP

      Stuart Tait examines the growing importance of AI prompt engineering skills for tax professionals – and how to learn them.

      “Fine.”

      That’s the less than informative response that every parent gets when they ask their child about their day at school.

      The trick, of course, is to ask a more specific question. ‘What did you learn today?’ ‘Who did you play with at lunchtime?’ Then the answers you’re after will be forthcoming.

      No, you haven’t stumbled on a parenting blog…it’s just that the same goes for getting the answers you need from artificial intelligence (AI) solutions. And ‘prompt engineering’, as it’s known, is becoming an essential skill for tax professionals.

      Frames of reference

      Stuart Tait
      Stuart Tait

      Partner, Chief Technology Officer, Tax & Legal

      KPMG in the UK

      I’ve explored how generative AI is set to transform tax operations elsewhere. Tax teams will increasingly find themselves using AI tools, and must understand how to get the most out of them.

      Prompt engineering is the key to doing that. Exactly how you instruct an AI solution will determine the information you get back. There’s a skill to framing your question in the right way, to ensure the response is relevant, accurate and comprehensive.

      Take, for example, an AI software that evaluates whether or not individuals should be paid under IR35. Simply asking for John or Jane Smith’s employment status is too broad a query. But feeding in the specifics of their working relationship with the organisation will allow the tool to make a precise assessment.

      Unsuitable prompts can have serious consequences, partly due to AI programmes’ tendency to ‘hallucinate’ when they can’t provide correct or complete data. Parties to tax tribunals have been known to present case law that doesn’t actually exist. The cases cited were generated by an AI system.

      The practicalities of prompting

      So what should tax practitioners know about AI prompt engineering? What’s the secret to getting it right?

      I’d say there are four essential ‘ingredients’ to an effective prompt:

      • The objective

        What you want the solution to achieve. E.g. write an email to a particular client.

      • The context

        Why you need the outputs you’re requesting. E.g. to tell the client you’ve prepared their tax return, and need them to review it by a certain deadline.

      • The constraints

        Any limitations on how you want the response to be delivered. E.g. the email should be no longer than 50 words.

      • The tone

        The writing style. E.g. professional but friendly, jargon-free and with a clear call to action to review the tax return.

      Much of this may sound like common sense; it’s how you might delegate a task to a member of your team. But prompt engineering isn’t as intuitive as it seems, because it’s not how we’re used to searching for information digitally.

      For two decades now, we’ve used search engines to discover what we want to know. Searching requires broad prompts; we then trawl the results for the right answer or source. AI prompts are the opposite: they need to be as precise as possible. Thanks to twenty-odd years of googling, we’ve been conditioned to use AI systems in entirely the wrong way.

      With that in mind, following these ground rules will make sure your prompts contain the necessary ingredients:

      • Be specific

        Clearly defined parameters will reduce the need for additional requests to get you to the desired result.

      • Clarify

        What not to do. It can be helpful to instruct the solution what to avoid doing, as well as what you need it to perform.

      • Provide examples

        Like people, AI systems work better when given examples of the required outputs – especially if these closely mirror existing online content.

      • Request imitation

        If appropriate, get the solution to emulate an individual, or a type of person, who is ideally equipped to answer your query.

      • Experiment and iterate

        Try different prompts to see which work, and adapt them if you’re not getting the required results.

      Getting familiar with AI

      This last point about experimentation is especially important. AI models can be unpredictable, so trial, error and practice will be necessary. That will mean giving your staff access to solutions as early as possible, so they can become comfortable using them to make their jobs easier.

      One way to do this is to create personas – typical AI user ‘types’ within your team – then pre-load each type’s most common prompts into your AI systems. This can be done on Digital Gateway, KPMG’s powerful client tax and legal platform.

      We can also help you train a core group of AI ‘superusers’ within your tax function. You’ll need one or two experts who know how to make more complex requests using advanced prompting techniques. These include creating user personas; breaking down prompts to improve the model’s understanding and processing efficiency (‘chunking’); and guiding it through a reasoning process for more accurate and coherent responses (‘chain of thought’).

      Please get in touch to find out more about how our experts can prepare your tax team for the AI era. We can offer a one-hour overview of AI in the tax function; a half-day session to identify your pain points; or a full Ignition workshop, where we’ll recommend the right solutions for your business.

      Our tax Insights

      Something went wrong

      Oops!! Something went wrong, please try again

      Get in touch

      Read enough? Get in touch with our team and find out why organisations across the UK trust us to make the difference.

      Contact us
      Person smiling whilst using a mobile phone