Prompt engineering is fast becoming a critical skill for accountants who want to make the best use of emerging generative artificial intelligence (GenAI) tools. Knowing how to create prompts for such tools will boost productivity, accuracy and speed.

It is also becoming a job in its own right, with a variety of industries beginning to advertise for these roles. Just Google ‘AI prompt engineer’ and discover the salary range for this nascent job type, which currently ranges between £30k and £400k depending on experience and industry.

Don’t be put off by the term ‘engineering’, though. It has little to do with actual engineering, but rather more to do with communication and providing the most accurate detailed instructions for the AI tool. Critically, it is an iterative process.

Different large language models (LLMs), for example, may interpret the same prompt differently, which can affect outputs. Some keywords might also have particular meanings for certain models. 

Dudley Gould, VP Business Development at Circit, says: “Prompt engineering is the way that we interact with the large language models. It’s the way you phrase the question. It’s about how you input the right words and data into the large language model to get the desired output.”

Ineffective prompts lack complete information for an AI tool, which could result in bias or inaccurate information because often the LLM will try to find an answer for the user and could result in ‘hallucinating’ one based on parameters that are too wide.

Effective prompts are specific, detailed and provide clear parameters. Users can even ask their AI model to suggest improvements to their existing prompt. The following are the types of detailed commands worth including in your prompts: goal, tone, word count, audience and format.

Gould says: “The best bit of advice I can give anyone is: imagine you’re speaking to another human. If you’re going to delegate a task and you want someone to do a good job, don’t just say, ‘write an article’, for example. Provide context, goals, word count, intended audience, with really clear instructions. Ideally, even include examples.” 

Types of prompts

There are two types of prompts, according to Mark Wickersham, chartered accountant and bestselling author: quick prompts and template prompts. Quick prompts are fairly basic and for simple tasks that do not require much thought – book recommendations, for example. 

“A really quick prompt will get a reasonable result and for many things that may be sufficient, but it’s still worth understanding the components of a prompt because you could spend an extra couple of minutes on your prompt and get a much better result,” he says.

Template prompts work well for tasks you might use frequently or are more complex. For example, it could be a report summary for every client meeting. For these tasks, it’s worth taking the time to craft an effective prompt that has been tested and proven to give a good output, which can then be used every time that task needs to be performed.

“I look at prompt templates as basically being a system,” Wickersham says. “So, we might create prompt templates for doing reports or analysing financial data. Another thing might be to carry out research before we meet a brand-new client, or something we could spend ages doing, but ChatGPT does it really, really well. We can create a research report for every client we’re about to meet. It’s about getting a consistent result in the way that you’re expecting.”

On the plus side, there are some patterns to aid prompt engineering that can be used as part of your input, such as providing existing examples for the AI model.

“A big problem is access to the right data. Organisations have still got a lot of systems that don’t speak to each other. So if I type ‘give me all my invoices’ into ChatGPT, it won’t necessarily have access to all the invoices so it can’t solve that problem,” Gould says.

It is still very early days for these AI tools, but where prompt engineering is proving valuable is in commands such as summarising lengthy, complex reports and contracts and extracting key information. This helps to speed up the process, directing the user to a specific page and avoiding having to trawl through swathes of time-consuming text.

Earning your stripes

There is a plethora of online courses popping up for prompt engineering and although these courses can give you a basic understanding, it seems that the best way to hone your skills is by trial and error and practising on a dummy model.

Many accounting firms are beginning to create best practice libraries of prompts for their teams to use for specific tasks. These libraries will undoubtedly grow as more and more accountants use the tools and refine the best prompts.

It’s about inputting data and if users don’t get the result they wanted, refine the prompt with more instructions and try again. For many, the fear of starting is the biggest obstacle. 

Gould says: “The main pitfall to watch out for is not to put client data into an unsafe environment. A lot of people just open ChatGPT and input client data, and that’s the biggest no-no. Just make sure you’ve got a secure enterprise set up, or use dummy data in your mock scenarios.”

Some advanced prompt engineering techniques

Chain of thought (CoT): This is about providing a series of prompts to an LLM to get the desired output. Instead of using one prompt, you develop a step-by-step process.

Self-consistency: This is used to improve the performance of an LLM. You generate multiple outputs from the same prompt, sometimes with slight variations. This also helps to identify patterns and inconsistencies within the model. 

Tree of thought: This takes CoT a step further, deconstructing a problem into a series of smaller steps that are then used to create multiple chains of thought. It helps LLMs to consider multiple paths of reasoning and evaluate decisions attached to them. 

For more information about prompt engineering, visit the ICAEW Generative AI Guide, which includes a link to a prompt engineering webinar.