Automate quarterly reports without manual work

Most people who sit down with the quarterly numbers believe they are doing something the law requires. Usually they are not. An unlisted Swedish limited company has no obligation to produce a quarterly report, and since 2016 not even listed companies have that obligation. Yet thousands of business owners spend three or four weeks a year assembling the same summary by hand. This guide covers what the law actually requires before you automate quarterly reports, what can be automated, and exactly where the automation stops.

The order matters. Obligations first, because they decide which dates you are locked into. Then the build. Limits last, because there are things in a Swedish finance process that no machine is allowed to do for you.

Do Swedish companies have to file a quarterly report?

No. An unlisted limited company has to produce an annual report, nothing more. The obligation to produce an interim report exists in the Swedish Annual Accounts Act but in practice only covers financial groups such as credit institutions, securities companies and insurance undertakings. Sole traders are not covered at all, so for the vast majority of Swedish businesses there is no quarterly report to file anywhere.

The counterintuitive part is that listed companies no longer have to report quarterly either. The requirement for interim reports covering the first and third quarter was removed through government bill 2015/16:26, following changes to the EU Transparency Directive. Nasdaq Stockholm then dropped its own requirement, arguing that it would reduce administrative burden and encourage long-term thinking. What remains in law is the annual report and a half-year report.

Several Swedish sources still claim the opposite, including the Swedish Wikipedia article on quarterly reports. Always check against the Swedish Accounting Standards Board summary of what applies to limited companies before anyone tells you that you are obliged to do something.

The absence of an obligation does not make the report pointless. It makes it voluntary, and a voluntary report has to earn the time it takes.

What forces a quarterly rhythm anyway?

Tax filing deadlines. VAT and employer declarations have to be filed at a fixed frequency regardless of what you think, and they require the books to be in order on specific dates. The rhythm already exists in the business. The quarterly report is simply a way to get paid for work you are doing anyway.

VAT sets the pace

Which period you report VAT for is decided by turnover. The Swedish Tax Agency rules on reporting periods set three levels:

Taxable base	Period	Any choice?
Up to 1 million SEK	Tax year	Yes, monthly or quarterly can be chosen
Up to 40 million SEK	Calendar quarter	Yes, monthly can be chosen
Above 40 million SEK	Calendar month	No

If you sit in the middle band you report VAT quarterly by default. The declaration is due by the 12th of the second month after the period, with August as an exception at the 17th. If the date falls on a weekend it moves to the next working day.

Employer declarations come every month

If you have employees, a different pace applies. Employer contributions and withheld tax must be declared the month after, every month, even when there is nothing to report. In that case you file a zero declaration.

That means the payroll side of the books has to be closed twelve times a year while the VAT side may only close four times. An automated flow has to handle both rhythms, not just the one you happen to think about.

What should the report contain to be worth producing?

It depends on who reads it. A bank wants liquidity and debt levels. A board wants variances against plan. You yourself usually want to know whether margins are holding and whether the cash will last. A report trying to answer all three gets long and ends up read by nobody.

Start with the reader and work backwards. In practice most smaller companies land on four blocks: the result for the period compared to the same quarter last year, liquidity right now, the largest variances explained in plain language, and a forward view.

The fourth block is the one usually missing, and it is the only one that actually changes decisions. A summary of what has already happened is history. A forecast is a basis for action.

What has to be finished before the quarter can close?

Reconciliation. Before the numbers can be compiled, the books have to match reality: bank against booked balance, payouts against orders, receivables against what has actually been paid. If that step is not automated it does not matter how polished the report looks, because it rests on numbers nobody checked.

That is a separate job with its own method, and we have written about it separately. If you sell through several channels, how to match orders against payouts walks through the protocol step by step. If your orders are scattered across systems it starts even earlier, with getting the order flow into one place.

Frequency matters more than many assume. The accounting profession recommends ongoing reconciliation under fixed routines, meaning monthly or more often. If you save up three months of variances until quarter end, you get a pile to dig through in exactly the week you have the least time. Automated monthly reconciliation turns quarter close into a compilation rather than an investigation.

If AI agents are new to you, our overview of AI agents for smaller businesses gives you the foundation before going further here.

Which details must every accounting record contain?

Seven of them, under the Swedish Bookkeeping Act, chapter 5 section 7. It is the single most important list in the whole automation effort, because each point corresponds to a field that machine extraction has to get right for the documentation to hold up legally. The law requires the record to contain:

1. When the record was compiled
2. When the business transaction took place
3. What the transaction refers to
4. The amount involved
5. Which counterparty it concerns
6. A record number or other identifier
7. Any other details needed to establish the connection between the record and the booked transaction without difficulty

Modern document extraction handles the first six well. Dates, amounts, suppliers and invoice numbers are read from an ordinary PDF with high accuracy.

Point seven is where it fails. The connection between the document and the booked entry is not something written in the document. It is something the system has to create and store. A receipt sitting in a folder without a link to the accounting record does not meet the requirement, however well it was read. That is why the link back to the books, not the reading itself, is the hard part to build.

The documents also have to survive

Accounting information must be kept through the seventh year after the end of the calendar year in which the financial year ended. The Swedish Accounting Standards Board is clear that electronic form is fully equivalent to paper, and that a company may choose the form when the document arrives in both around the same time. Scanning the paper much later counts as a transfer with stricter requirements.

One rule change few know about makes this easier than it used to be. Through SFS 2024:342, in force on 1 July 2024, the requirement to keep the original until the fourth year was removed. A company may now destroy the paper document as soon as the information has been transferred, provided the transfer carries no risk of anything being changed or lost. That rule was what used to force a binder to sit next to the digital archive.

In practice it means an automated flow has to know where the documentation ends up and be able to produce it seven years later, not just read it today.

How do you automate the compilation?

By pulling accounting data straight from the accounting software API instead of exporting files by hand. Both Fortnox and Visma expose interfaces for records, accounts, financial years, customer and supplier invoices. The report is then built by code, not by copied cells.

Technically, the work to automate quarterly reports comes down to three layers, and most companies do not need more.

Retrieval. A scheduled job that reads the period from the books and stores a normalised copy. Expect limits: Fortnox allows 300 calls per minute per client on a five second sliding window, and responds with error code 429 when the ceiling is hit. A build fetching details for hundreds of records has to queue and retry rather than firing everything at once.

Calculation. Key figures are derived deterministically from the chart of accounts, not by a language model. Revenue, gross margin, operating result and liquidity are sums over account ranges. If you trade in foreign currency, rates come from the Riksbank open API instead of being looked up by hand.

Presentation. This is where a language model belongs: explaining in plain language why an item deviates from last quarter, and writing the summary. The number comes from the code. The wording comes from the model.

Where does the automation stop?

At filing. A VAT return can be prepared by machine but not submitted without a human. The Tax Agency documentation for filing by file states that contact details cannot be included in the file and have to be entered manually, and that whoever files must identify themselves with an electronic ID.

The same applies inside the accounting software. Filing a VAT return directly from Fortnox requires an agent authorisation with the Tax Agency called Momsdeklaration, ombud, and the return still goes out as a draft that is signed with an electronic ID.

This is not a shortcoming in the technology. It is designed that way, because responsibility sits with the company. An automated chain promising to handle everything up to the tax authority without human signing either describes something other than what you think, or something you do not want.

The practical consequence: build to shorten the work up to the approval, not to remove the approval. If the system touches company data it is also worth reading what GDPR requires when AI handles your information before connecting the sources.

How do you build a cash forecast that updates itself?

By combining three known quantities: the actual balance today, receivables with due dates, and payables with payment dates. Add recurring items such as salaries, rent and taxes on top and you get a rolling forecast without anyone guessing. Most builds project twelve or thirteen weeks ahead.

Two things decide whether the forecast is useful or misleading.

The first is keeping the booked and actual balance apart. The books show what has been recorded, the bank shows what is actually there. The difference between them is not an error but information, and a system that merges them hides exactly what you need to see.

The second is labelling every number with its status. Booked, preliminary or forecast. A forecast figure that looks like a booked figure is more dangerous than no forecast at all, because it invites decisions it cannot carry.

What did we learn from building it for real?

That the hard part is never the reading. In a build for an online retailer, the system reads incoming supplier invoices from email, interprets them and links them to the right entry in the books. The model that reads the document was working within days. The rules for when the link may happen automatically took several times longer to get right.

Blocks always beat permissions

Automatic linking happens only when several conditions hold at once: high confidence in the match, an amount difference under one krona, a date inside a set window, and a sender on an approved list. If a single condition fails, the item goes into a queue for human review instead. The check against double linking is built so that on failure it answers that the item is already linked, which is the cautious answer. A system that guesses in the wrong direction creates more work than it saves.

Not every entry needs documentation

The insight with the biggest effect was to stop chasing everything. Far from every accounting entry needs a document to hunt for. Salary payments, depreciation, tax transfers and internal reclassifications either create their own documentation or already have it. By reading the entry lines and deciding based on which accounts are involved, rather than looking at the description text, most of what first looked like missing documentation disappeared. What remained were the entries that genuinely needed a human.

The system improves from corrections

When a person confirms a link the machine was unsure about, the connection between the cryptic bank statement text and the real supplier is stored. After two confirmations it starts influencing future matching. That requires no retraining of any model, only making use of decisions that are being made anyway.

Our own internal finance system follows the same principle throughout: it retrieves, compares and compiles, but books nothing and sends nothing. Every figure carries a reference back to its source. That is the only form of automation that holds up when someone asks where a number came from.

What does it cost, and when is it not worth it?

A bounded system automating the report compilation starts at 45,000 SEK in implementation, with operations between 5,000 and 15,000 SEK a month and no lock-in period. Connecting several systems into one shared flow lands in the 90,000 to 180,000 SEK range. What a build costs in detail is covered in the cost guide for AI agents.

Base the calculation on time released, not technology purchased. At NordicRank, a Swedish SEO agency, 18 automated processes replaced manual lists for orders, project tracking, invoices and client reports. The value is 13.4 hours saved per week, with payback after four months.

But it is not always worth it. Three situations where you should wait:

• The books are not in order. Automating a broken process delivers the same errors faster. Clean up first.
• The volume is too low. With twenty accounting entries a month it is cheaper to do it by hand, however inefficient that feels.
• Nobody reads the report. If you build a summary no one uses, you have automated the wrong thing. Start by asking who will read it and what they will decide.

What argues for doing the calculation now is that the time spent and the reporting burden are documented problems, not a feeling. In the Swedish Agency for Economic and Regional Growth survey of business conditions 2026, answered by 6,342 companies, 62 percent of the businesses that want to grow report at least one major growth obstacle, with laws and regulations and the time it takes to comply among the largest. The Swedish Federation of Business Owners report on VAT complexity points the same way: three in ten small business owners say VAT rules cause administrative hassle to a fairly or very high degree.

Choosing to automate quarterly reports does not remove the rules. It removes the time it takes to comply with them, and moves your effort from hunting for numbers to deciding what the numbers should lead to.

Frequently asked questions