Business Automation: What Changed When AI Arrived

Business automation changed its boundary, not its nature. Until 2022, automating a process required that process to be describable as rules: if this field holds this value, do that. Anything that depended on interpreting a loosely written email, a badly scanned PDF or a request each customer phrases differently stayed outside.

Artificial Intelligence did not make automation smarter. It removed one specific prerequisite: the information entering the process no longer has to be structured.

That sounds like a technical distinction. In practice it is what separates the processes that became automatable from the ones sitting exactly where they were three years ago.

The wall classic automation kept hitting

Traditional process automation is deterministic. It takes a predictable input, applies a rule, writes the result into another system. It works very well, provided somebody has translated the process into conditions.

The bottleneck was always in the same place: the input.

Take the most common case in a European SME, booking supplier invoices. A 2019 workflow could detect the email, store the attachment, create the right folder and notify accounts. It could not read the invoice.

The answer back then was template-based OCR: you defined, supplier by supplier, where the number, the date and the total sat on the PDF. That meant one configuration per supplier, and it broke every time a supplier changed its layout. At 40 suppliers, maintenance cancelled out the gain.

That was the wall. The capacity to execute the action was never missing. The capacity to interpret the document was.

First shift: reading stopped being the hard part

Language models solve precisely that layer. They pull fields out of a document they have never seen, in a layout nobody configured, and return them structured.

Adoption data shows this is what companies are actually using AI for, rather than mainly for generating content. In 2025, 20.0% of EU enterprises with 10 or more employees used AI technologies, up from 13.5% in 2024. The leading application across the EU is analysing written language, at 11.8% of all enterprises, ahead of generating images, video or audio (9.5%) and generating text or code (8.8%).

Reading ahead of writing. AI entered companies through the interpretation door, which happens to be the piece automation was missing.

Second shift: the price per document

Technical capability without a viable price changes no decisions. Cost is what moved first.

Stanford University's 2025 AI Index documents the drop: inference cost for a system performing at GPT-3.5 level fell from 20 dollars per million tokens in November 2022 to 0.07 dollars by October 2024. More than a 280-fold reduction in under two years.

In day-to-day terms, a two-page invoice amounts to roughly 1,500 to 2,500 tokens of text. At that price level, extracting the data from one document costs fractions of a cent.

The point for a decision-maker is not the exact figure. It is that model cost is no longer the variable that stalls the project. What weighs now is the time of whoever designs the workflow and the time of whoever validates the output, and that calculation is the same one that already made sense before AI. If you have not run it yet, the real cost of keeping the process manual is the number to start from.

A worked example, with the arithmetic visible

Triage of a shared mailbox, the kind sitting behind info@ or orders@.

Starting point: 60 emails per working day, around 2 minutes each to read, classify and route. That is 2 hours a day, roughly 44 hours a month.

With an AI step classifying each email and deterministic rules routing it, the arithmetic runs like this:

• Assume 70% are classified with enough confidence to proceed without review, a conservative figure that only measurement confirms. That leaves 18 emails handled manually.
• 18 emails at 2 minutes each is 36 minutes a day.
• Add 15 minutes a day to spot-check what went through automatically. Leaving this line out is the most common mistake in these calculations.
• Total: about 51 minutes a day against 120. A saving of roughly 1.1 hours a day, or 25 hours a month.

These figures are an estimate built from the assumptions above, not a measured field result. The method matters more than the number: state an explicit automation rate, subtract the cost of supervision, then measure once the workflow is live.

What AI did not change about business automation

Three things are unchanged, and this is where most projects come apart.

The action still has to be deterministic. The model interprets and proposes; the rule is what writes to the ERP. A model with direct write access to the system of record, with no validation in between, is an architecture decision you pay for later.

Some tasks still require human oversight. There are back-office steps where human control remains essential for reasons model accuracy does not address: legal liability, contractual exceptions, decisions with direct customer impact.

Data limits are a governance problem, not a tooling problem. Anyone processing invoices, contracts or personal data has to decide where the model runs and what leaves the infrastructure. That is the security, limits and oversight side of AI agents, usually left until last when it belongs at the start.

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The same analysis warns about agent washing: RPA products and chatbots rebranded as agents without real agentic capability. Treat it as a purchasing warning.

Where SMEs actually stand

The headline figure hides the important part. Adoption scales with company size, and the gap is wide. In Portugal, where our own clients sit, national statistics office INE puts 2025 AI adoption at 9.4% among enterprises with 10 to 49 employees, 18.2% for 50 to 249, and 49.1% for 250 or more.

A company with 30 people is therefore in a minority of roughly one in ten.

And the distance between 9.4% and 49.1% is not access to technology. Models cost the same for everyone and require no upfront investment. The difference is that the large company had already mapped its process and had its data inside a system. That work, not AI, is still the prerequisite.

How this is built today

The architecture that works is hybrid: a deterministic workflow with an AI step only where interpretation is needed.

That is what the n8n AI Agent node does, implementing the tool calling interface: the model decides which tool to call and the workflow executes the call. The actions stay in ordinary, auditable nodes with an execution history. And n8n can run on your own server, which keeps the documents inside the company infrastructure.

One piece then gets skipped almost every time: logging what the model returned and what the person corrected. Without that log there is no way to know whether accuracy is 95% or 70%, and the decision to widen the workflow ends up being made on instinct.

The decision in front of you

Before picking a tool or a vendor, answer one question: in which process does somebody on your team read something in one place and type the same thing somewhere else?

That is where the arrival of AI changed what is possible. For everything else, classic business automation still applies, under the same rules as always.

If you do not know how many times a day that happens in your company, that is the first job. And it does not need AI.