Edition history

Every company knows them and nobody likes them: incoming invoices. A supplier sends one, someone reads it, types it up and enters it into the books. That is called posting. Day after day, from the painter's workshop to the large corporation.

Today an AI can take this over. A program similar to ChatGPT reads an invoice in about three seconds. A person needs about three minutes. That is 60 times faster. I think that is real progress.

But speed is not everything. You know the satnav that works out the route in a second, is almost always right, and then sends you into a one-way street with complete conviction? From the wrong end?

Imagine 100 invoices on the desk. The AI posts each one in three seconds. After five minutes the pile is gone. But something is wrong with invoice 37. The amount and the supplier match. Yet the service dates from last year and belongs in the previous financial year, and it is unclear which VAT treatment applies. The AI posted it anyway. Plausible, but wrong.

Who notices? And what does it cost until someone does?

This is where it gets interesting. Not the question of how fast an AI can post, but what happens when it is wrong. You do not need any accounting knowledge for this. A simple worked example shows what matters.

## The routine disappears. The exceptions remain.

An amount can be read. An account can be suggested. The hard part begins where documents are incomplete, a service belongs in another period or a document arrives twice. These are the cases that show whether automation saves work, or only creates new corrections.

So a system must not only know how to post. It must recognise when it does not know enough.

A duplicate invoice, unclear VAT, a document without a service date: the standard case is often simple. A single exception can eat up the entire time saved.

This is not an argument against automation. It is an argument against the wrong yardstick. A low processing time says little about the benefit if rework, checking and later corrections are not measured too.

<figure> <img src="/api/media/field-notes/until-invoice-37-f01-en-32dc3fbe93c35f07b5291d4c2a459ce9616bdb18aeef75fc5d5269107fdf6876.webp" alt="Flow diagram: an incoming invoice is read by the AI, which proposes a posting. Clear cases are posted; unclear cases are escalated to human review and posted afterwards." /> <figcaption>Simplified control flow. Unclear cases must be escalated for review.</figcaption> </figure>

## The 5% trap: where does the work go?

Take 100 incoming invoices. A person needs about three minutes per invoice, the AI three seconds. For an exception we assume, as an example, eight minutes of human work. Standard cases go through without a second check.

With five exceptions, the 95 standard cases together take 4 minutes and 45 seconds. The five exceptions, by contrast, cause 40 minutes of work. So about 89.4% of the processing time falls on just 5% of the cases.

The gain remains: by hand it would be about 300 minutes, with automation about 45. But the decisive lever is the exception rate, the share of cases a person has to check. Not the speed per invoice.

This is a hypothetical calculation, not a measurement. It shows why the exception rate and the cost per correctly completed posting matter more than the duration of a single AI call.

<figure> <img src="/api/media/field-notes/until-invoice-37-f02-en-c083f62610c47d595285755b06af4feadc075d4ca33264d2c77f90ae8bc767a2.webp" alt="Bar chart: 100 invoices take 300 minutes by hand. With AI they take 4.75 minutes for 95 standard cases plus 40 minutes for five exceptions; 89.4% of the time falls on 5% of the cases." /> <figcaption>The 5% trap. Hypothetical example, not measured project results.</figcaption> </figure>

## Two market signals. No forecast.

Digital invoices are everyday life. In 2025 around 93 million eBill transactions were processed in Switzerland (eBill is the digital invoice in e-banking), and at the end of June 2026 more than four million users were registered (SIX, 27 August 2026).[^1] AI has also arrived at financial institutions: according to a survey of around 400 supervised institutions published by the Swiss Financial Market Supervisory Authority FINMA on 24 April 2025, about half use AI or have first applications in development.[^2] Both figures show how widespread these tools are, not the quality of the controls. Nor do they say anything about AI use at fiduciary firms, the companies that keep the books for others.

## In 2026 it is one step of the work. By 2035 it could change a profession.

<figure> <img src="/api/media/field-notes/until-invoice-37-f03-en-33d5717710a9f480ec047046afdb8e16bef7e98d86615e29742b3f40e8c2629b.webp" alt="Timeline with three points: 2026 observation, one accounting step becomes faster; 2030 hypothesis, documents, payments and rules become connected; 2035 scenario, a digital fiduciary answers business questions." /> <figcaption>2026: reported experience. 2030 and 2035: scenarios, not empirical forecasts.</figcaption> </figure>

**2026: observation.** A single accounting process becomes faster. Whether the whole workflow becomes correspondingly more productive has to be measured separately.

**2030: hypothesis.** Documents, payments, posting rules and follow-up questions become more closely connected. End-to-end workflows emerge instead of isolated automations.

**2035: scenario.** A business owner could ask their digital fiduciary whether they can finance a new machine. Behind it, the accounts, liquidity and open receivables are already linked. Whether this fiduciary appears as a chat, a voice or an avatar is secondary. It would only be the surface.

To turn this into a reliable service takes more than a clever AI: documented posting rules, identity checks, governed data access, clear limits for automatic decisions, human review where things are unclear, and decisions that can be traced.

The Swiss Code of Obligations' rules on bookkeeping and record retention also apply to records kept electronically.[^3]

## Who checks the books when AI keeps them?

The future of accounting does not belong to the system that posts fastest. It belongs to the one whose results can demonstrably be checked.

Anyone using AI in accounting should therefore not look at "seconds per invoice". What counts is this: what does one correctly completed, verifiable transaction cost, exceptions included?

Your next step: for a task that repeats all the time, incoming invoices for example, measure the exception rate and the total effort per professionally accepted transaction. Only then will you know whether the automation really takes work off your hands.

### Method and transparency

The figures of 180 and three seconds come from the author's reported experience with one bounded processing step. A publicly documented sample, error rate, follow-up check and operating-cost calculation are still missing. The exception calculation is hypothetical. The years 2030 and 2035 stand for scenarios, not evidence-based forecasts. AI was used for concept and editing; professional sign-off lies with the author.

## Sources

[^1]: SIX, 27 August 2026. [eBill Surpasses Four Million Users and Further Expands Direct Debit Functions](https://www.six-group.com/en/newsroom/media-releases/2026/20260826-ebill-surpasses-four-million-user-mark.html). Media release: over four million registered users at the end of June 2026; around 93 million eBill transactions in 2025.

[^2]: FINMA, 24 April 2025. [FINMA survey: artificial intelligence gaining traction at Swiss financial institutions](https://www.finma.ch/en/news/2025/04/20250424-mm-umfrage-ki/). Survey of around 400 authorised institutions, conducted from late November 2024 to mid-January 2025: about 50% use AI or have first applications in development.

[^3]: SECO SME Portal. [Electronic retention of business records](https://www.kmu.admin.ch/de/elektronische-aufbewahrung-der-geschaeftsbuecher), summarising Art. 958f of the Code of Obligations and the Ordinance on Business Records (GeBüV): books and vouchers are kept for ten years and may be kept electronically if they stay readable and match the underlying transactions.