Data volumes and quality 'crucial' to AI success in accounting firms

Accountancy firms must better organise their data to maximise the potential of artificial intelligence (AI), according to Becky Shields, head of digital transformation at Moore Kingston Smith.

“A lot of businesses need to work on their data strategies as they’re too siloed at the moment,” she says.

“The data is there but it’s not organised or catalogued effectively.”

Shields believes firms shouldn’t embrace AI until their data has been properly curated as the AI models could be learning from inaccurate information.

It’s a situation Moore Kingston Smith tackled around seven years ago when it started deploying analytics in the audit part of the business.

One of the challenges, she points out, was that not everyone in the firm possessed the skills needed to transform the data and put it on all the necessary platforms.

“We decided to build a data repository so that we could standardise clients’ data and store it in one place,” Shields explains.

“While there was a training burden if we didn’t standardise, we also now have a very powerful large machine dataset that can be easily deployed.”

Data silos are standalone databases managed by individual departments that are not easily accessible by other parts of the same business.

While commonplace in many professional services firms, their restrictive nature means they need to be removed to ensure data can be shared, according to Shields.

“Where possible, don’t have your data stored in programmes or on platforms that don’t have API’s [application programming interface] allowing you to extract that data,” she says.

The importance of data

The influence of data on AI is affirmed in a 2018 report published by the Institute for Chartered Accountants in England and Wales (ICAEW), entitled Artifical intelligence and the future of accountancy.

“Data volumes and quality are crucial to the success of AI systems,” it stated. “Without enough good data, models will simply not be able to learn.”

The paper warned that smaller organisations may not have enough data to enable accurate results. It also highlighted a potential lack of information about very specific problems to support good models.

“Powerful models may need external sources of data, which may not always be possible to access at an appropriate cost,” it added.

However, Rob Talson, chief digital officer at Evelyn Partners, argues that the amount of data available isn’t necessarily a problem as firms have been collecting vast amounts over recent years.

“The quality has improved greatly but to train the AI platforms well enough, the core challenge will be around how data is structured, organised and governed,” he said.

Although Talson agreed that such concerns were delaying the wider implementation of AI models, he insisted these weren’t the only factors, noting that “other specific issues preventing take up” include data security, privacy, ethics, accuracy, and compliance.

But according to Professor Monomita Nandy, accounting and finance professor at Brunel University London, not every piece of data will be good enough for modelling.

“It can only be considered as quality data when there is a rigorous independent auditing practice in place,” she said. However, she added this will be very much on a case-by-case basis.

“For training and testing machine learning algorithms there is a need of high frequency data. Most of the companies don’t have the expertise, skills or resources to produce this data.”

Nandy also pointed out that ethical quality data could be sourced from big businesses audited by some of the bigger accounting firms .

“These companies use technology like blockchain to record high volume of transactions and the audit firms have enough in-house expertise to assess the quality information needed for AI modelling,” she said.

Assessing data quality

But Moore Kingston Smith’s Shields argues that that the quality and quantity of data required is dependent on the use case for the AI technology. For instance, simple machine learning models don’t actually need that many data lines to learn from, she says.

If you’re building a really complex model to accurately forecast the potential performance of a business over the next three years, you’ll need lots of historic, contextual data. It depends on what you’re trying to build.”

Looking ahead, Professor Bonnie Buchanan, head of finance and accounting at the University of Surrey, believes the benefits of AI are worth the time spent organising the data.

“It’s already helping immensely with natural language processing, such as textual analysis, regulatory filings and financial statement filings,” she said. “It’s also used in compliance and auditing.”

Buchanan believes AI will become increasingly important in fraud detection over the coming years, but warned against firms trying to adopt a one-size-fits-all policy.

“As AI becomes more pervasive in the accounting industry there will need to be a shift towards appropriately educating workers.

“Many of the accounting professional societies are making great strides as graduates with tech, accounting and finance skills will be in high demand.”