AI inherits your mess – but it can also help you clean it up

If it is difficult for employees to find their way through the finance function’s systems, data and integrations, it will also be difficult for a digital employee – that is, an AI agent capable of performing tasks across systems. This does not mean that you should wait to implement AI until the foundation is perfect. On the contrary, the first relevant application may be to use AI to uncover the mess, improve the structure and rethink the processes that no longer fit the business.

For one client, we were asked to develop a digital employee that could combine three accounts receivable lists from three countries into a single report. At first glance, the task seemed straightforward: read the files, harmonise their contents and create a consolidated overview.

When we opened the files, reality turned out to be rather different.

They were written in three languages, the headings were not on the same row, and the amounts were stated in two currencies. The accounts receivable ageing structure varied from country to country, meaning that some categories had to be combined before the figures could be compared. The data included subtotals and blank separator rows, while one of the files was missing the customer number that should have linked the information together.

Even the amounts required special handling. Some cells contained numbers, while others contained text that merely looked like numbers. Decimal points, thousands separators and parentheses around negative amounts varied across the files.

Each individual discrepancy was trivial, but together they accounted for most of the work.

This is an experience that is repeated in many AI projects: the technology itself is rarely the only or the greatest challenge. The critical work often consists of understanding the environment in which the technology must operate and providing the digital employee with the right context.

When employees have to navigate between several systems, local spreadsheets, manual transfers and different definitions of the same data, they continuously compensate for the complexity. They may know that a particular field cannot be used, that a specific company applies a different structure, or that an amount is only accurate once a manual correction has been made.

This knowledge is rarely documented in one place. It is tacit knowledge embedded in employees, their workflows and the controls they have developed over the years.

A digital employee does not automatically have access to this implicit understanding. It sees the data, instructions and system access made available to it. If the connections are unclear, its basis for making or suggesting decisions is equally uncertain.

The question is therefore not only whether your data is of high quality. It is also whether the processes in your finance function are sufficiently clear and consistent for someone other than the most experienced employees to follow them.

A practical rule of thumb is this:

If a new employee struggles to understand the process without help from three experienced colleagues, a digital employee will also need more than access to the system

Microsoft’s Payables Agent for Dynamics 365 Business Central illustrates the issue. Among other things, the agent can read invoices, match suppliers and suggest account coding. According to Microsoft, these suggestions are based partly on the chart of accounts, the available transaction history and the company’s accounting practices to date.

This makes the solution useful because it can recognise and apply established patterns. At the same time, however, historical data does not necessarily represent the practices the company wants to follow in the future. It reflects the practices the company has followed so far.

If a particular cost has been coded incorrectly for an extended period, or if master data contains local variations of the same supplier, AI may carry the pattern forward. Not because the technology is failing, but because it is working from the foundation it has been given.

Microsoft therefore also emphasises that suggestions must be reviewed by a human, and that the agent does not automatically post invoices without approval. This is an important safeguard, but it does not resolve the entire challenge. If employees place a high degree of trust in the system’s suggestions, an inappropriate historical practice may still be approved and continued.

In other words, AI does not merely make work faster. It can also make existing practices more consistent, regardless of whether those practices are appropriate.

The conclusion should not be that the finance function must undertake a multi-year clean-up programme before it is allowed to use AI. In many companies, that would mean never getting started.

A better approach is to select a limited area where AI can both create value and make the finance function more aware of its foundation.

In the example involving the accounts receivable lists, working with the digital employee could reveal precisely where the companies’ data definitions and workflows differed. This did not merely provide a basis for building the solution. It also made it possible to ask more fundamental questions:

• Do we need different ageing structures in each country?

• Why is the customer number missing from one file?

• Which manual corrections do employees make every month?

• Are the differences based on genuine local needs, or are they simply the result of historical circumstances?

• Can data be retrieved directly from the source systems instead of being exchanged through local spreadsheets?

In this way, the AI project becomes an opportunity to understand and improve the process, rather than merely automate the existing version.

This represents an important shift. If the sole ambition is to digitise the current process, you risk making an inappropriate workflow faster. If the ambition also includes investigating why the process looks the way it does, AI can help challenge the status quo.

The data foundation is only one part of the equation. The other part concerns the actions the digital employee is allowed to perform.

When an employee is given access to a finance system, their permissions are normally assessed in relation to their role. When a digital employee is configured, there is a risk that its access will instead be determined by everything the solution technically needs to be able to do.

As a result, several functions may be brought together in the same digital actor. It may be able to create or amend data, suggest account coding, prepare a journal entry and send it on in the process. If the permissions are not assessed collectively, an otherwise well-established segregation of duties may gradually be eroded.

The Danish Agency for Public Finance and Management’s guidance on public payments requires personal segregation between registration and payment. The rules are aimed at the public sector, but the underlying control principle is relevant much more broadly: the actor that records a transaction should not automatically be able to complete the entire process alone.

For the CFO, it is therefore useful to treat a digital employee like any other actor in the process. Map what it can initiate, approve, post and pay. Then assess which approvals, restrictions and subsequent controls are necessary. The key issue is which actions the digital employee can perform and how those actions are controlled.

An initial AI solution does not require a perfect system landscape. It should, however, be selected in a way that both creates a tangible benefit and improves your understanding of the foundation.

Start with one clearly defined process and ask three questions.

1. Where should the digital employee obtain its knowledge?

Map the systems, files, data sources and instructions involved in the process. Investigate where employees make manual adjustments and where they rely on experience to interpret data that is not unambiguous.

The aim is not to develop an overall data strategy. The aim is to understand the specific information that the digital employee must be able to rely on.

2. What should the digital employee be allowed to do?

Describe the digital employee’s role in the same way you would describe the role of a human employee. Which information may it read and amend? Which suggestions may it prepare? When must a human approve an action, and which actions must the solution never perform on its own?

Permissions and controls should be part of the design from the outset, rather than something added when the solution is ready to go live.

3. What should the process look like in the future?

Avoid treating the current workflow as an inviolable requirements specification. Instead, use the project to investigate which steps create value, which compensate for poor system integrations and which exist solely for historical reasons.

Sometimes the right outcome is a digital employee. At other times, it is a simpler process, better master data or an integration that removes the need for anyone to transfer the information at all.

AI models and vendors are evolving rapidly, and the solution you choose today will likely change significantly over the coming years. This makes it tempting to wait for the next model or the next mature standard product.

However, the improvements that make the finance function understandable to a digital employee also benefit the people working in it. A more unambiguous chart of accounts, better master data, clear integrations and well-designed access rights reduce complexity, regardless of which technology is subsequently built on top of them.

The choice therefore does not have to be between cleaning up and getting started with AI. The first well-chosen application can do both: create a tangible improvement here and now while showing where the foundation needs to be strengthened.

AI inherits the system landscape, data and processes you already have. But when used correctly, it can also help you see them more clearly and build them better.

A digital employee does not operate independently of the rest of the finance function. It must be able to retrieve reliable data, understand the rules of the process, work with the existing systems and operate within an organisation with clear lines of responsibility.

The digitalisation circle helps put the solution into context. AI is not an additional layer that automatically removes the complexity underneath. However, a specific AI application can make the complexity visible and show where data, processes, systems or capabilities need to be improved.