Simplifying expense claims for employees: What a modern system should be able to do

Expectations for receipt scanning should be high

Almost every system lets you photograph receipts and then reads them for you. This functionality has been around for a long time, but the quality varies greatly and expectations should be set high.

A modern system should extract the country, date, amount and currency with very high accuracy: at least 95%. It should work just as well for Swedish receipts as it does for Chinese ones. Beyond that, the system should understand what was actually purchased and be able to classify the purchase in the right expense category. Where a receipt contains several different types of purchase, such as a hotel invoice with a room, breakfast and parking, the system should automatically split it into the correct categories if your company requires this. On top of that, you want the system to learn from its mistakes, so if the finance team makes a correction, the system should remember it.

In other words, there is a big difference between this and half-heartedly reading a few fields and leaving the employee to do the rest manually.

The system should guide the employee

One major source of friction for employees is having to know what is required for correct reporting: which supporting documents are needed, what complies with policy, what needs an explanatory comment, or when something breaks a tax rule. Many employees don’t take the time to learn this and instead fill in as little information as possible before submitting their report. The result is returned reports, delayed reimbursements and frustration for everyone involved.

A good system turns this around. Instead of requiring the employee to know all the rules, the system should use the minimal information the employee has provided to infer how something should be booked, and flag anything that needs correcting before submission. It is not enough for errors to be caught afterwards.

For example, the system might:

• Warn that the document is an order confirmation rather than a receipt, and ask for the correct document
• Flag that a purchase goes against policy and explain why
• Identify duplicates and point them out
• Remind the employee to add a comment when required, e.g. if alcohol was included at a dinner
• Warn that the wellness allowance being claimed does not qualify
• Note that a per diem should include a meal deduction because there is a restaurant receipt for the same day

The more the system understands about the purchase, such as the supplier, document type, contents and context, the smarter this guidance can become. In other words, it’s about much more than reading a few simple fields.

Selected features that make a real difference

A few concrete things to evaluate:

Depth of receipt scanning. Can the system only extract basic fields, or does it have a deeper understanding of the receipt?

Learning. Can the system learn from the corrections that are made?

Receipt splitting. A receipt with several categories should be split automatically if the company requires it. The employee shouldn’t have to do this manually.

Card matching. If the employee pays with a company card, the system should automatically match the transaction to uploaded receipts, so the employee doesn’t have to link them manually.

Attendee management. The system should remember attendees from one occasion to the next and infer whether they were internal or external based on the available information.

Mileage allowance. It should be easy to register a journey you’ve made several times without logging each trip separately, and the system should be able to calculate distances automatically based on the start and end addresses entered.

The future of expense claims is collaboration between AI agents

It is reasonable to ask what the expense claim process will look like going forward.

The B2B software of the future is not a single tool that does everything, but a system of specialised agents that communicate with each other. Each agent does what it does best, and a human is brought in when a decision needs to be made.

Applied to expense claims, it could look like this: the company’s AI agent is notified that a card purchase has been made (for example, by the expense management system). It then retrieves the receipt from the email inbox or asks the employee to take a photo, attendees are filled in from the calendar invitation, and the agent writes a purpose based on the available context. It then communicates with the expense management system’s agent to ensure that the documentation complies with policy. The employee receives a finished proposal to review and approve.

In this scenario, it is the company’s agent that owns the context: who was there, what the meeting was about, and why something was purchased. A company may not want to give the expense management system full access to tools such as email and calendar. Instead, the expense management system acts as the specialist that knows the policy and tax rules and what the company has approved in the past. The two agents communicate with each other so that the expense management system receives complete documentation for further processing.

The technical building blocks for this already exist. AI agents can already communicate with each other, and standards are being developed. What is holding things back are unresolved questions around access, privacy and risk control, and not least how quickly companies are willing to adopt new ways of working. For natural reasons, companies move more slowly than consumers in this area.

The important thing is to choose an expense management system that is built to communicate with the outside world: a system with open integrations that is intelligent enough to answer what applies, not just smart enough to take instructions.