AI In Accounting: How To Work Smart While Staying Safe
Filip Borcov, Incredible from Site.pro.
gettyBecause they are easy to interact with, AI chatbots can make complex systems more accessible to non-experts. Not surprisingly, AI is also effective at automating repetitive tasks. Financial institutions already use AI for operational efficiency, risk management and fraud detection.
Why not just start using AI in accounting? There are nuances to it. The main concerns I’ve seen in this space are data security, compliance and accuracy.
In 2025, researchers found an exposed DeepSeek database containing chat histories, prompts and more. Even Meta AI had a bug that could read private prompts and generated content.
The obvious question is whether it is worth using AI when working with sensitive data, or is it still too risky. I believe AI is still worth adopting in accounting and ERP systems. Let’s explore how to implement it without compromising your company.
Accounting systems are complicated and are often used by more than just accountants. Managers and directors who do not work with these systems every day can easily get lost in them.
Accounting has its own language and a lot of nuance. A regular user cannot simply look at the data and quickly find answers, whether it’s identifying how much money was spent on a certain project or finding information from one particular document.
AI could simplify this a lot. Most AI chatbots can quickly scan through data and give users simple answers. This type of automation can also save time. JPMorgan reported that its contract-analysis system could process thousands of commercial credit agreements in seconds, replacing large amounts of manual work.
At my company, we’re putting effort into implementing AI in every part of our workflows. We’ve been experimenting with all accessible and public AI models. But every model has its flaws and consistency issues, so we started looking for a more reliable solution to implement AI.
Technically, you can upload your data directly into commonly used AI chatbots, like ChatGPT and Grok. It is relatively easy to configure API access, so any AI model could access your data and analyze it.
But do you want to let a chatbot into your financial database with sensitive documents and invoices? Probably not. OpenAI confirmed that a ChatGPT bug exposed some users’ conversation titles and payment information.
So the chance of leakage is real. A service can get compromised, accounts can be hacked and links to queries can accidentally be made public.
Most modern ERP systems provide APIs. You can give a chatbot your login and password and ask it to fetch the data you want. It’s safer than uploading your files directly into the chat, but it’s still not the most secure option.
To simplify things, think of an API as a worker with one specific job. An accounting system may have several different workers: One worker can retrieve invoices, another can check payments, another can return customer information and another can show sales data. Each of these workers has a specific access point and knows how to perform a limited task. That access point is what you give an API.
If you ask the invoice API for invoices, it can return invoices. It does not need access to everything else in the accounting system.
Model context protocol (MCP) is a more structured way to connect AI with APIs and other tools. It works like a manager for API “workers,” each of which accesses its specific point.
Instead of the user deciding which worker to contact, the user asks, “Show me sales for last month.” Through MCP, the AI can select the relevant tool, call the required API and return the result.
It is possible to set up an MCP in regular chatbots, but the problems include data safety, memory and server locations. If your company is based in Europe and the AI provider processes data outside the EU, you may face additional GDPR and data-transfer requirements.
It’s worth noting that not all AI models are equally suitable for financial workflows. Some companies may warn users not to rely on their chatbots for financial advice.
Another problem involves how business terms are defined. For example, in your ERP, “sales” may mean sales including VAT or excluding VAT. An MCP tool may use a different definition or apply additional rules—for example, returning only sales converted into a specific reporting currency.
As a result, two MCP implementations connected to the same ERP can return different figures for the same request.
MCP can connect AI to dozens of operations inside a system and direct it to the right tools. Employees do not need to know how the API works or write code to use it, which makes MCP relatively easy to introduce into daily work.
• Start with a narrow task. Begin with one clear use case, such as finding invoices, checking payments or preparing a sales report.
• Define business terms inside the tools. Make sure terms such as sales, revenue, VAT or reporting currency have a clear definition. Otherwise, the AI may return technically correct but misleading results.
• Make the source visible. For financial data, the AI should show which records, documents or reports were used to produce the answer.
• Separate read and write actions. Searching and analyzing data can be automated more freely than editing records, deleting documents or making payments.
• Define what needs confirmation. Employees should know which actions AI can complete automatically and which ones require human approval before they are executed.
At first, teams may be hesitant to start using AI in their regular work. Change is scary. But new tools can unlock new possibilities for growth. AI agents have already improved the development speed for our development team. Other teams are starting to use them for their tasks—we can all make progress and learn new things with AI every day.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
