AI is rapidly becoming part of finance teams’ most critical processes. A global KPMG study of more than 1,000 senior finance executives shows that active use of AI in finance has more than doubled in two years, from 30 to 75 per cent. AI is already being used in financial planning, reporting and commercial analysis.
At the same time, a large share of AI use is taking place outside companies’ control. As early as 2024, 86 per cent of Swedish white-collar employees who used AI at work said they brought their own AI tools into the workplace.
Put those two figures together and we have a problem that I believe many management teams underestimate. AI has already entered the finance function. But governance has not always kept pace. When employees’ own AI experiments start being used for forecasts, analysis and decision support, a far more uncomfortable question arises:
Who takes responsibility when AI gets it wrong? The CFO, the employee who built the solution, or the AI provider? The answer may not be what everyone wants to hear. The company.
For many years, Excel has been one of the finance function’s most important, yet least controlled, tools. It can be used to build sophisticated forecasting models, reporting solutions and business-critical processes in a spreadsheet. The problem begins when something originally created to solve a small issue suddenly starts driving a critical process. Formulas are changed. Files are copied. Data is moved. Eventually, the company is left with a model that influences important decisions but is only truly understood by one or two people.
With generative AI, we risk repeating the Excel mistake at hyper-speed. A financial controller can build an AI tool to analyse variances. A CFO can create an AI agent that prepares material for the board. An accountant can automate a process that previously would have required a development team.
This is not an argument for slowing down AI adoption. Quite the opposite. But we need to stop pretending that an AI experiment is still just an experiment once the business starts relying on its output. Somewhere along the way, the experiment becomes infrastructure. And at that point, it is no longer enough to say: “The AI got it wrong.”
Direction beyond the numbers. Try Oxceed free of charge.
Try Oxceed completely free for 10 days, with no commitment. View your company’s financial data in real time, carry out analyses and create reports. Get started in just a few minutes.
Take the recent Air Canada case, for example, where the company tried to shift responsibility to its chatbot. Air Canada learned this lesson the hard way. As you may remember, a customer used the airline’s chatbot and received incorrect information about the rules for a specific discount. The customer followed the instructions. When Air Canada later refused to honour the information, the company argued, among other things, that the chatbot was a separate source of information. That argument did not hold. A Canadian tribunal ruled that Air Canada was responsible for the information presented on its website.
Of course, this was a customer service case, not a finance case. But the principle should make every CFO think carefully. You can outsource the task to AI. You cannot outsource the responsibility.
If an AI-generated analysis is used as the basis for a forecast, investment decision or board report, it matters little how impressive the underlying technology is. The organisation must be able to stand behind the result. And that means answering far more questions than whether the AI “works”. As an organisation, do you know what data was used? Where was it stored? How did the model reach its conclusion? Can the result be reproduced and reviewed? Who approved it? And who identifies when it is wrong? The board will not accept “ChatGPT said so”
For the finance function, this is particularly important because output quickly becomes input for the next decision. An analysis becomes a forecast. The forecast becomes a budget decision. The budget becomes material for the board. When a board member or auditor asks why a figure looks the way it does, it is not enough to say that the AI model produced it. The reasoning must be demonstrable. There must be a clear link to the underlying data, a process that can be traced and a person who is prepared to take responsibility for the result.
That is why there is a fundamental difference between using AI to write an email and using AI as part of forecasting, financial analysis or board reporting. The greater the consequences of an error, the less acceptable it is to have a black box between input and output. More AI does not necessarily mean better AI.
Faster decision-making for Hagabadet
Running a spa, hotel and experience business means managing multiple revenue streams, numerous profit centres and rapid fluctuations in demand.
There is also an interesting paradox in KPMG’s figures. AI adoption in finance has more than doubled since 2024, and organisations report improvements in areas such as decision-making speed, decision quality and forecasting accuracy. Yet despite this rapid development, only 23 per cent say that AI is exceeding their expectations.
That should prompt us to stop measuring AI success by the number of features, agents and pilot projects. The next phase is not about who has the most AI. It is about who has AI that can be trusted. And this is where I believe many organisations have underestimated the challenge. Buying an AI tool is easy. Building an agent is relatively easy. The difficult part begins when it is put into real-world use. That is when questions arise around information security, data quality, access permissions, traceability, regulatory requirements and human oversight.
In my view, this is the CFO’s responsibility. Preventing employees from experimenting with AI would be the wrong approach. The potential for innovation is simply too great. But unrestricted innovation without clear accountability is not an AI strategy. It is the absence of a strategy. CFOs and management teams therefore need to define the point at which an AI experiment becomes part of the company’s infrastructure. Which AI tools are allowed to process financial information? Which processes must be documented? When is human oversight required? Who owns a solution built by an individual employee? What data is allowed to leave the company’s own systems? And which processes are simply too critical to automate without clear control mechanisms?
This is not primarily a technology issue. It is a management issue. And in the next phase of AI development, I believe trust will become a competitive advantage. In recent years, the AI race has been about who launches first, automates the most and builds the fastest. That race is becoming less interesting. What matters now is which companies we are prepared to trust. For customers, boards, investors and auditors, it ultimately matters less how advanced a company’s AI is. They want to know that information is secure, that the figures can be explained and that someone actually takes responsibility when something goes wrong. AI has made it cheaper and faster than ever to build software. But it has not made accountability cheaper.
We will allow AI to analyse more, recommend more and play an increasingly important role in the information used to make decisions. That is both inevitable and fundamentally positive. But the more responsibility we give the technology, the clearer we must be about who ultimately stands behind the result.
AI can take over the process. But never the responsibility.
Lin Eriksson Nordström, CMO at Oxceed, a Swedish SaaS company specialising in financial reporting, budgeting and analysis.
This text has been edited with the support of generative AI. The author remains responsible for the content, arguments and conclusions.
