Amid the rapid surge of interest in generative artificial intelligence (GenAI), prospective software buyers are facing a scrum of vendors claiming that their tools are AI-enabled for pretty much everything. That’s the assessment of Konrad Bukowski-Kruszyna, Audit Data Analytics Director at RSM UK.

It is a confusing and often rowdy picture that accountants in business and practice need to make sense of, especially at a time when many will be under pressure to adopt AI tools and ‘partner’ with them to gain an edge in the productivity stakes. But before that partnering can happen, how can accountants overcome the initial pressure of deciding which AI tool is most compatible with their needs?

“The first thing you need to work out is what the tool will actually be doing,” Bukowski-Kruszyna says. “There are lots of different ‘streams’ of AI out there right now. And they all have their place.”

In what he calls the “foothills” of AI – given that there is some debate as to whether it even counts – is robotic process automation (RPA). “Lots of people start with RPA before they move into ‘full-fat’ AI,” he says. “There are some great tools available – Automation Anywhere springs to mind. There will be many rules-based tasks your accountants, auditors and tax professionals are dealing with on a daily basis that require only limited human involvement. So RPA will help you hit that low-hanging fruit and quickly speed up the completion of those processes.”

For Bukowski-Kruszyna, compatibility is generally not an issue with RPA tools as they tend to be very flexible and adaptable: “There’s an awful lot you can do without getting too technical.”

Human-led approach

Moving up the ladder, the next rung is optical character recognition (OCR) software. This can absorb data from numerous document types at a glance and contextually analyse it. Leading examples include Dext, which is already in use among many accounting firms, and DataSnipper – beloved by auditors – which recently received a much-coveted ‘unicorn’ valuation of $1bn. 

In terms of OCR’s use case, Bukowski-Kruszyna points out that professionals can scan in large volumes of invoices, or feed them in as PDFs, and then ask the tool to break down the data into various categories.

“An OCR tool will typically require some form of setup,” he explains, “so that it will understand what sort of trends you as a specific user want it to look for. GenAI is starting to bleed into the market and, in some cases, remove the need for this training. But the downside is that the outputs will need a robust review to ensure they’re not based on incorrect assumptions.”

Next up, moving into the field of ‘traditional AI’, is machine learning. “This is where you start to harness complex algorithms, including decision trees,” Bukowski-Kruszyna says. “The main use case would be for very high-volume, Big Data analysis, especially of a predictive nature. For instance: ‘If I change this input or variable, what could the various outputs be?’ In the world of audit, fraud identification is a commonly touted use case.”

At the top of the ladder is GenAI; what Bukowski-Kruszyna calls “the new kid on the block that everyone’s either very excited or concerned about”. He acknowledges that ChatGPT has quickly become synonymous with the term, but cites Claude and Llama as other examples of intuitive tools that, if used appropriately, can hold literal conversations with their users.

This is a sphere with a lot of scope for partnering with AI, as the capabilities of the tools are incredibly broad, Bukowski-Kruszyna explains. But you need professional judgement and a human-led approach to identify when the AI is “going off on a bit of a tangent” and needs a course correction.

He notes: “It’s not as if you’ll suddenly have a new starter who joined the firm two weeks ago creating fully compliant IFRS accounts – at least, not yet. But that’s what people use professional advisers for: the comfort of knowing that someone who knows what they’re doing has combed through the outputs and deemed them accurate and worthy of presentation to clients or auditors.”

Pain points

In Bukowski-Kruszyna’s assessment, when it comes to choosing their AI partner, accountants should be looking to cover off two different types of ‘fit’. Firstly, technological compatibility: which task(s) you want the tool in question to execute. For example, do you need it to plug into your audit work paper solution, your accounts prep solution or your tax solution, or can it stand alone? 

Secondly, there is human compatibility: do your colleagues and employees have a clear, operational sense of what to do with the tool and how to get the best out of it? “A Rolls-Royce or Ferrari is a very accomplished, fancy machine,” he says. “But if your aim is only to get from A to B, a Fiat Punto or Ford Escort would do the trick. So it’s important for accountants to focus very clearly on the question: “What are your main pain points?’ 

“From there, you can conduct some root-cause analysis of those points to determine the best partner. If you have a very stodgy, manual process that takes ages to complete, an RPA tool could shorten it from, say, two days to five minutes. Yes, a GenAI tool could do that too, but would be massive overkill for that particular use case.”

Crucially, Bukowski-Kruszyna warns: “Don’t get sucked in by any hype, because there’s lots of it flying around these tools. Keep at the forefront of your mind: ‘What is the end goal I’m trying to achieve?’”

He stresses that, at key points in the chain, you will need a human to review the outputs before they are shared externally. Indeed, he says, it would be “incredibly bold and risky” to simply send material off to clients without appropriate vetting, and as the newest and most powerful tool, GenAI needs particularly careful scrutiny.

“I always think of GenAI as a hyper-enthusiastic intern,” he adds. “It may not fully understand the context you’re working in but it will try to get you a result at lightning speed. Some of its outputs won’t be on target. But that’s where it falls to you as an experienced professional to spot the gold nuggets you can iterate from. You need to avoid a sort of Dunning-Kruger effect of: ‘Computer says X, so it must be right.’”

• Read part 2 of this mini-guide here: Partnering effectively with AI, part 2: evaluation in use