In accountancy, artificial intelligence (AI) is often thought of as a tool that is being adopted at pace by major firms, while smaller ones race to catch up. But solo Insolvency Practitioner (IP) Paul Brindley has kept track of AI for more than a decade and has used it enough to know that its effectiveness depends on scrupulous data management.

Brindley set up his practice Midlands Business Recovery around 15 years ago. In the run up, he got together with a group of ‘white-hat’ hacker friends from various professions to talk about how they could use AI in their everyday work. 

All being tech obsessives, the friends hatched some compelling concepts, but agreed that the state of software at the time was not up to fulfilling them. Nonetheless, Brindley knew that a technological revolution was on the way. So, he dedicated himself to hoarding as much data as possible on the specific subject of insolvency, from legal texts to case law and procedural materials on how companies were wound up or saved.

In November 2022, with the advent of ChatGPT, the revolution he had foreseen arrived. In short order, Brindley found a piece of software that would sit on top of that groundbreaking platform and serve as a ‘bot’ for answering queries. He describes the tool as “like a light and nimble jockey that wrangles a thoroughbred racehorse”. 

That Christmas, he began to load the bot with the mass of data he had harvested. Several months of testing later, he made the bot live on his practice’s website. Its main purpose: to provide company directors with a quick and easy means of obtaining accurate steers on topics related to corporate insolvency.

Sensitive terrain

Alongside that primary role, the bot also helps the practice by carrying out process-driven admin tasks, providing Brindley with more time to focus on strategic and/or client-based work. “For example, one client had a potential liquidation coming up. So, we needed to get in touch with customers who’d paid the company up front by debit card, credit card or PayPal and let them know how to claim their money back. The bot drafted the letter in seconds.”

By any standards, this is sensitive terrain to navigate, often involving, as Brindley points out, “directors who are in trouble, struggling and very vulnerable”. With that in mind, he notes, it is of vital importance to ensure, on a constant, rolling basis, that the bot is working from a body of high-quality data.

Good data hygiene

As such, one key point to address is data hygiene. “I don’t want the tool to have access to any old waffle on the internet,” Brindley says. “So I force it to focus only on the foundation of data I’ve fed it.” 

But that’s not the only thing you’ve got to be aware of, he explains. You’ve got to monitor the results it provides and use that secondary data to fine-tune the system. “You can’t monitor the results in real time, because they fly out so quickly. It has to be done after the fact, so at 4pm every day I get a report that tells me which questions users have asked the bot, and what sort of answers it has given, and I use that information to tweak the dataset.”

Some of the ‘catches’ that Brindley has had to step in with have stemmed from users asking the bot about matters related to personal insolvency, which is outside the practice’s remit. But he stresses that handling sophisticated questions is something that the software is cut out to do. Indeed, he welcomes users posing them because they actually help to train the tool. “We’ve had 20-odd years of using Google and people still tend to ask fairly simple, ‘Google-type’ questions, which are posed with no expectation of interaction.”

Regular housekeeping

Typically, sound data management for any AI tool involves keeping a firm grasp on three key areas: availability, quality and governance.

In terms of availability, Brindley says you can never really have enough data to train an AI on. The plugin he uses to run his bot has an input limit of 11m characters. 

“I’ve been brushing up against that almost from the word go,” he says. “I have to do a regular housekeeping routine of taking down older data and uploading more recent material just to save space. In fact, I’m negotiating with the software provider behind the plugin to see if they will allow me to double my limit. AI tools need huge amounts of data to run properly.”

Keeping data as up to date as possible is also essential for maintaining quality, Brindley points out. He notes that AI is able to pull data from links to external websites – a useful facility when it comes to ensuring that the bot is basing its answers on the most current laws or rulings in the insolvency field, or other relevant policy announcements. 

But it does put the onus on the bot manager to delete any outdated links from the system and add fresh ones as they emerge. In addition, Brindley tightly monitors the quality of the bot’s outputs. If any answers have strayed from factual accuracy, it is vital to find out which data input(s) have caused the error.

In parallel with those quality measures, Brindley says, are some strict governance rules to limit the scope for biases. “I purposely say to users: ‘Don’t provide the bot with any personal details, such as phone numbers or email addresses – if you want to contact me, send me a separate message.’ I also provide the disclaimer that the bot’s answers do not constitute formal legal advice. If that’s what users are after, they should contact a specialist IP.”