AI automation in mid-sized companies: when rules are enough

Most companies now use AI, almost none get much out of it

Adoption has crossed the halfway mark. Bitkom reported on 14 September 2026 that 57 percent of German companies use AI, up from 36 percent a year earlier and 20 percent two years before, from phone interviews with 603 companies of 20 staff or more.

The second number from that survey matters more. Asked how much of the potential they actually tap, 59 percent answered none at all, 3 percent answered quite a lot, and nobody answered fully. Among the firms still on the sidelines, 85 percent blame a lack of implementation know-how and 63 percent unpredictable costs.

That gap is a sorting problem, not a technology problem. The workflow that starts life as an AI project is often plain rule automation: fixed input, fixed condition, fixed output. No model required, which is the part few vendors volunteer.

Three questions sort any process in two minutes

Three questions settle whether a model is needed. Answer them for the workflow you want to automate first, yes or no, with no room for mostly.

• Input: does the information always arrive in the same place and the same shape, such as a form field or a spreadsheet column? Yes means rules.
• Decision: can the routing be written down as a finite list of if-then lines that one person finishes in an hour? Yes means rules.
• Output: must identical input always produce identical output because a deadline, an amount or a booking depends on it? Yes means rules.
• Three times no means AI: free text coming in, no closed list of rules, and a draft going out that a person signs off.
• Two yes and one no means rules plus an AI check: the workflow runs on rules and the model only sees the cases that fall through.

A decision table for your first process

Six workflows that show up in almost every company, run through that test. Each line reads: workflow, verdict, reason.

• Filing an invoice from the shared mailbox into the accounting folder: rules, because sender and subject are enough and nobody reads the amount here.
• Pulling amount, date and tax rate out of the invoice PDF: AI, because the layout changes with every supplier. The workflow is in preparing incoming invoices with AI.
• Writing an appointment from a web form into the calendar: rules, because every field is known and a lookup catches double bookings.
• Routing a free-text damage report to the right tradesperson: rules plus an AI check, because a keyword list catches the obvious cases and the model handles the remainder.
• Deciding the dunning level and sending the letter: rules, because a deadline and an amount hang on it. No model belongs in that step.
• Bringing four supplier quotes onto the same quantities: AI, because units, discount tiers and surcharges are written as prose. See comparing supplier quotes with AI in Excel.

Rules cost 20 euros a month, AI costs 22.81

The price gap is small. Everything below assumes 1,000 cases a month, both paths in the same automation platform, list prices as of 28 September 2026.

• Rules path: n8n Cloud Starter at 20.00 euros a month with 2,500 executions, which is 2.0 cents per case.
• AI path: the same 20.00 euros plus tokens. At 1,000 fresh input tokens, 2,000 cache reads and 400 output tokens per case, Claude Haiku 4.5 comes to 3.20 dollars a month, or 2.81 euros at the ECB reference rate of 1.1403. Total 22.81 euros, 2.3 cents per case.
• Microsoft path: Power Automate Premium lists at 15.00 dollars per user per month, so 13.15 euros; unattended flows as Power Automate Process cost 150.00 dollars per bot, or 131.54 euros.
• By hand: 1,000 cases at two minutes each is 33.3 hours. At the average gross hourly wage of 28.19 euros that is 939.67 euros a month, 94 cents per case.

Rework decides this, not the subscription

A 14 percent premium for the model settles nothing. The real gap sits in the error column. Assume 1 percent wrong routings on the rules path and 5 percent on the AI path, three minutes to correct each one. Both rates are figures set for this calculation, not measurements.

On those numbers the rules path produces 10 corrections, half an hour, 14.10 euros a month. The AI path produces 50 corrections, 2.5 hours, 70.48 euros. Totals of 34.09 against 93.28 euros, a factor of 2.7 behind a 14 percent price gap.

The kind of failure matters more than the count. A rule that does not fire leaves an empty cell. A model that knows nothing does not go quiet, it writes.

Both paths stay far below doing it by hand: 905.58 euros saved per month on rules, 846.39 euros on AI. The per-case arithmetic is in working out AI cost per case.

Two costs only surface in year two

Your model will be switched off. Anthropic promises at least 60 days of notice before a model retires and retired three during 2026: Claude Sonnet 3.7 on 19 February, Claude Sonnet 4 and Opus 4 on 15 June, Opus 4.1 on 5 August.

Parameters disappear too. Setting temperature, top_p or top_k returns a 400 error on Claude Opus 4.7 and later models. A rule-based flow has no such problem: it runs until somebody edits the rule.

The second cost is the data route. On the AI path, free text from customer mail leaves the building. To strip names and numbers first, there is a free tool for redacting prompts, and the limits of that masking are in removing customer data before a prompt.

Classification also drives your duties. The European Commission published guidelines on the definition of an AI system on 6 February 2025, and what follows for your staff is in proving AI literacy under Article 4.

Three recommendations, two of them against AI

Here is how we would decide in your position.

• Yes to rules, no model: anything with a deadline, an amount or a booking. Dunning runs, appointment bookings, routing by sender. The flow stays auditable and a breakdown shows up.
• Yes to AI, with sign-off: free text that has to end up in a form. Invoice line items, quote comparisons, service requests. The condition is a person who approves the draft, or you pay for the rework twice.
• No to both below 21 cases a month: a 20.00 euro subscription only earns itself back at 21 cases, assuming two minutes of manual work per case. Below that, a list kept by hand is cheaper.
• The mixed path pays off from roughly 200 similar cases a month, when the hours spent on the keyword list come back.

Common questions

Four questions that come up every time.

• Do we need an automation platform of our own? Not necessarily. n8n ships a self-hosted community edition at no licence cost, so you pay for a server and its upkeep instead of a 20.00 euro subscription.
• What about the 63 percent who fear unpredictable costs? Tokens are not the risk. In the same Bitkom survey, 8 percent named token usage as a major cost block, against 51 percent for infrastructure and 50 percent for data preparation.
• Can we combine rules and AI? Yes, and that is the right answer most often. A worked example with sorting rules and a model check is in pre-sorting a shared mailbox with AI.
• What if the process is full of text but rarely runs? Then it stays manual. What an agent costs over twelve months is in what an AI agent costs.

Next step: keep a tally for one week

Take a sheet of paper and count for a week. One line per case, three columns: did the information arrive structured, could the decision be written as an if-then line, did an amount or a deadline hang on it.

On Friday, count the marks. Above 80 percent yes in all three columns means rules, and you have saved yourself a model. Below 50 percent means AI with sign-off. In between sits the mixed path, costed at the same 28.19 euros an hour.

If you would rather not sort this alone, we walk through it in a free first call and build the workflow as an internal AI service.