Use cases by function

AI agents for SME finance and accounting

Roles involved
accountants, finance managers, financial controllers, payroll officers, accounts receivable and payable staff, fiduciaries
Common software
Abacus, Bexio, Banana Accounting, Sage 50, Crésus, WinBiz, Proffix, Excel, e-banking (ISO 20022, QR-bill), the FTA ePortal, Swissdec-certified payroll
09

Monthly results commentary agent

Today

Every month, the finance manager compares the figures with the budget and the previous year, then writes a commentary for management. It takes time and the text arrives late.

With AI

Once the month is closed, the agent reads the profit and loss export and the budget stored in SharePoint, picks out the variances above the thresholds you set, then opens the general ledger to find the entries that explain them. Finally, it drafts a first commentary in Word, with the source of each explanation. The finance manager checks, corrects and signs off the text before passing it to management.

What to measureThe time between month-end and delivery of the commentary, and the share of explanations corrected during review.

Tools Microsoft 365 Copilot and its agents in Excel and Word, Claude or ChatGPT Business with SharePoint connector, export from Abacus, Bexio or Banana

10

Payroll check agent before each payout

Today

Variable hours, night and Sunday supplements, changes in working percentage: before each payout, the payroll officer compares payslips with the previous month by hand.

With AI

As soon as the month’s payroll journal is exported, the agent compares it with the previous month, the time records and the supplement rules of your collective agreement, then lists the deviations to explain, citing the relevant rule for each. It saves the list in the payroll folder, in an environment excluded from model training. The payroll officer checks each case in the payroll software; management approves the payout as it does today.

What to measureThe number of payroll corrections made after payout and the checking time before each pay run.

Tools Claude Team or ChatGPT Business with SharePoint connector, Microsoft 365 Copilot, export from the payroll software (Abacus, Bexio, Sage 50…), time clock system·Sectors Hospitality, Healthcare

11

Supplier invoice pre-booking agent

Today

Invoices arrive as PDFs, on paper or by e-mail and are keyed in one by one: supplier, amount, VAT, account, cost centre. Differences from the order and coding errors only surface at year-end.

With AI

When an invoice reaches the accounts payable mailbox or arrives by eBill, the agent reads it, finds the supplier in the software, compares it with the purchase order and delivery note, then prepares a pending entry with account, VAT and cost centre according to your rules. It flags price or quantity differences and drafts the question to the buyer. The accountant checks and approves each entry; payment is still authorised by the person entitled to do so.

What to measureThe share of prepared entries changed by the accountant and the time between receipt and booking.

Tools Abacus, Bexio, Banana, Crésus or WinBiz, Outlook, QR-bill, eBill, Power Automate or n8n, AI model via API

12

Agent for unpaid invoice reminders

Today

Reminders go out late, all with the same wording, regardless of the customer’s history or an ongoing dispute. Some unpaid invoices drag on for lack of time to follow them up.

With AI

Every week, the agent reads the overdue open items in the software, searches the mailbox and CRM for a payment promise or a dispute, then sets the reminder level according to your rules and prepares a suitable text for each customer as a draft in Outlook. It sets disputed cases aside with a note summarising the exchange. The accounts receivable person reviews, sets aside sensitive cases and sends the reminders themselves.

What to measureCustomers’ average payment time and the age of open items, tracked month by month.

Tools Bexio, Abacus, WinBiz or Proffix (receivables), Outlook, CRM, Copilot Studio or ChatGPT Business with connectors

13

Agent answering service charge questions

Today

After heating and service charge statements are sent, the property manager receives dozens of tenant questions about items and allocation keys that have to be explained again each time.

With AI

When a tenant writes about their statement, the agent identifies the building and lease in the property management software, retrieves the statement, the allocation key and supplier invoices, checks the item in question and prepares a clear draft reply. Any formal objection becomes a task for the property manager. The property manager or building accountant approves the reply before it is sent.

What to measureResponse time to tenant questions after statements are sent and the number of formal objections.

Tools Property management software (Rimo R5, ImmoTop, GARAIO REM or Abacus), Outlook, Copilot Studio or agent connected by API·Sectors Real estate

14

Agent for matching unrecognised payments

Today

QR payments with a reference are matched automatically, but transfers without a reference, combined payments and partial amounts stay in suspense and are handled by hand at month-end.

With AI

Every morning, the agent retrieves the camt notifications from the bank, isolates the payments the software has not matched and looks for the corresponding invoices in the open items, payment advices received by e-mail and the debtor’s name. It prepares the pending match with the reason for it and, for a partial or unknown payment, drafts the question to the customer. The accountant accepts or corrects each suggestion; nothing is booked without their approval.

What to measureThe number of payments in suspense at month-end and the share of suggestions accepted without correction.

Tools E-banking (camt.053, camt.054), Abacus, Bexio, Proffix or your ERP via API, Outlook, agent built and connected by API

15

Year-end closing agent for client files

Today

During closing season, every client file needs the same checks: accruals, provisions, depreciation, unusual balances. They are done by hand, client after client, with Excel lists that differ from one staff member to another.

With AI

When a client file enters closing, the agent retrieves the trial balance, applies your checklist, compares it with the previous year and looks for supporting documents for the balances in the client folder of the document management system. It then assembles a file of points to verify and drafts the list of questions for the client. The client manager handles each point and decides on closing entries; the agent writes nothing into the books.

What to measureReview time per client file and the number of corrections requested after the manager’s review.

Tools Abacus, Bexio or Banana (trial balance via API or export), firm’s document management, Outlook, Excel, agent built and connected by API, AI model hosted to your requirements·Sectors Fiduciaries

16

Agent tracking costs per site or project

Today

Subcontractor invoices, hours, materials: the costs of a building site or client project are spread between the accounts and field reports. The real margin only becomes clear once the project is finished.

With AI

Every week, the agent takes the new subcontractor invoices and the hours recorded, allocates them to sites or projects using the order number or project code, and asks the project manager in Teams to confirm doubtful allocations. It then updates the cost tracking against the quote or NPK/CAN items and flags overruns. The project manager confirms, management decides what to do.

What to measureThe gap between the margin estimated during the project and the final margin, and how long it takes to allocate invoices to projects.

Tools Messerli or Sorba, project and time tracking tool, Abacus or Bexio, Teams, Excel, agent built and connected by API·Sectors Construction, Agencies

In short

In finance, an AI agent does not keep the books for you: Abacus, Bexio, Banana or your ERP remain the reference. It reacts to a trigger (an invoice arriving, a bank statement, month-end) and works through the steps on its own: it reads the document, compares it with the order, prepares a pending entry in the software, matches a payment or leaves a reminder as a draft. It approves no entry, releases no payment and files no return: someone with authority checks and decides, and every action the agent takes is logged.

Points to watch

Accounts are confidential data and salaries are personal data under the nFADP. Accounting records must be kept for ten years and remain legible and intact (art. 958f Code of Obligations and the Accounting Records Ordinance). The tools we select make it possible to follow these rules if your data is not used to train models, hosting is documented and the agent only has the rights it needs: reading and preparing, without approving or paying. For a fiduciary firm, professional secrecy also applies.

Bexio and Abacus already read invoices. Why add an agent?

Often it is not needed: we start by checking what your software already does and whether it is set up properly. An agent becomes useful for the steps around those features that are still manual: comparing the invoice with the order, finding the explanation for an unrecognised payment, preparing a reminder that takes the exchanges with the customer into account. Adding a tool only makes sense if a specific task is still done by hand.

Can we put our accounts into ChatGPT?

Not in a free or personal version. A business plan whose contract excludes training on your data, with documented hosting and named accounts, can be suitable for targeted exports. We define with you which data may go in, in what form, and what stays in your accounting software. Payroll data calls for stricter rules.

Can an agent pay an invoice on its own?

No. The agent can prepare the payment proposal in your software or the payment file for e-banking, but it has no signing rights: release stays with the authorised people, with dual signature if you use it. Its rights are limited to reading and preparing, every step is logged and uncertain cases are flagged rather than booked. As with any entry, the company remains responsible for what it approves.

Choose your first use cases.

In half a day or a day, the AI Workshop sorts your ideas and keeps three, tested on your own documents.