Joe Burns had a problem with his balance sheet, and he couldn’t figure it out. His company, Reformed IT, offers cyber security and IT services for businesses in the East Midlands and London. 

“I’m in a quarterly peer group with other IT business owners,” Burns explains. “Members have to share numbers from their accounts, with full transparency. Given the number of years I’ve spent in business, I’m pretty confident I can read – and spot problems in – profit and loss reports quite easily. But with the balance sheet, it’s a different story.”

As Burns compiled balance sheet figures for the next meeting, negatives were appearing in unexpected places, for no apparent reason.

“It didn’t look right,” he says. “And I was determined not to report false numbers. So, for me, it wasn’t just about getting the issue fixed. I also wanted to understand why the problem had occurred.”

It was well into the evening before the deadline for the peer group submission and the COO – who also handles day-to-day accounts management – was on holiday. It was a reconciliation issue, but he couldn’t figure it out.

“I looked at the time,” Burns says, “and it was late enough for me to not want to bother our external accountants. So, I had to come up with a different solution.”

Building an agent

Reformed IT creates bespoke Copilot agents for its clients, so Burns knew the processes required to build one. “The way I saw it was: ‘Here’s another test. Let’s see what I can do with Copilot if I direct it to this particular purpose.’”

Burns planned to build a very simple agent, which meant he could use the basic agent builder, Create Agent (more complicated agents can be built in Copilot Studio). “When you’re putting Copilot agents together, you start by telling them who they are, what they do and what expertise they have.”

Burns turned to Copilot’s chat function and asked it: “If I were to hire an accountant to tackle this particular type of balance sheet problem, what sort of knowledge would I be looking for?” 

He then used the tool’s response to shape the reference materials that were into the agent’s knowledge base. This included online guidance from HMRC and the Financial Reporting Council, key information from ICAEW’s website, and documents from Xero.

He restricted the agent’s internet access to those resources, reducing the risk of misinformation and hallucinations. “I didn't want it to learn from things Bob said on Reddit.”

This can be done by accessing the agent’s ‘knowledge’ settings and switching off ‘search all websites’.

Talking to the agent

With the agent ready to go, the problem-solving process unfolded as a conversation. Following some initial, context-setting queries from Burns, the agent asked to import his full balance sheet in its current form, plus his trial balance and account codes. The agent identified the issue. “Transactions we should have logged as Income in Advance were entered in Accrued Income. Essentially, they were in the wrong bucket.”

Burns asked what he should do to fix the problem, and the agent told him that he would have to journal out the reversal of the error, along with the correct re-entry of the transactions.

Burns asked for a list of steps to do this, but it was too complicated to implement, so he asked for an easier solution. “The agent replied with, ‘Sure – here’s a CSV file. If you import this into the journals tab on Excel, it will do it all for you.’ Sure enough, after that, the journal rolled out automatically.” 

The agent also flagged another minor but well-concealed problem: a discrepancy with a value of around £300.

This solved the issue for the time being, though the work needed to be checked by his accountants.

Accountants check the work

Burns spoke to his company’s external accountants the next day. He explained that he’d been reluctant to disturb anyone so late in the evening, and asked them to check the agent’s handiwork. On review, the accountants confirmed that the virtual CFO had actually correctly diagnosed the problem, and that the solution was also on target.

“Our accountants are very forthcoming about this sort of thing,” Burns notes. “They’re not one of these firms that tries to bury its head in the sand and ignore how AI is helping to make accounting more efficient.”

This use of Copilot – where a client solves reconciliation issues using the technology and the accountant then checks it – could become more common as more businesses start testing uses of AI. 

“Accountants will be increasingly reviewing and checking AI’s output when it comes to low level functions such as reconciliation issues, freeing them up to focus on the high value work for their clients,” says Ian Pay, ICAEW’s Head of Data Analytics.

As for Burns, he has built another Copilot agent and connected it to his Xero platform via a model context protocol – this did involve the use of Copilot Studio. This enables the agent to monitor the company’s balance sheet for discrepancies on a monthly and quarterly basis. “Along the way,” Burns says, “I’ve learned a lot more about how the balance sheet works.”

ICAEW is working with Microsoft to inform members on how AI can be applied in practice. A full-day course, highlighting use cases such as this one, will take place on 14 September 2026.