On most AI projects the technology is not the problem. The problem is that nobody wrote down how long the process takes today, before anything changed. Without that there is nothing to compare against later, and the question of whether it was worth it becomes a matter of opinion.

This article gives you a formula, six rows to track alongside it, and one worked example. The figures are typical industry ranges, not the results of any particular company. Your own number comes out of the template, not out of us.

Why you never find out whether it paid off

Three mistakes show up most often, and all three happen before any measuring starts.

• There is no baseline. The time before rollout gets estimated after the fact. Estimates typically differ from measured values by 20–40 per cent, usually upwards, because everyone remembers their worst day.
• Saved time is treated as money automatically. Forty saved hours are money only if something specific happens to them: less overtime, a hire that does not need making, or more work delivered with the same headcount.
• Running costs get left out. API usage, licences, oversight and error fixing are a monthly line item. Together with the one-off build cost, these two numbers are what decide the payback.

Sort those three out and payback becomes simple arithmetic.

The formula

Two steps. First work out how much money comes in each month, then divide the one-off cost by it.

Monthly realised saving = (minutes saved per month / 60) × fully loaded hourly cost × realisation rate

Payback in months = one-off build cost / (monthly realised saving – monthly running cost)

Two items need explaining.

The fully loaded hourly cost is not gross pay divided by 160 hours. It includes employer contributions, holiday and non-productive time. For an office or administrative role in Slovakia and Hungary this is typically 14–22 euros per hour. Your own figure takes ten minutes to pull from your accounts.

The realisation rate tells you how much of the saved time actually turns into money. If the freed capacity goes nowhere, the rate is zero. In practice 50–80 per cent is the typical band, and you sit at the top end when the process is a bottleneck, so freed time immediately means more work delivered. If you cannot say where the freed time will go, use 50 per cent. Underestimating is the safer error.

Measurement template: six rows per process

One table, one column per process. That is enough, and nobody will keep more than that up to date.

1. The process and the unit measured. A quote request, an incoming invoice, a customer question. One unit, one item.
2. Volume per month. From the system or the mailbox, not from memory.
3. Manual time per item before rollout. Measured across three working days, with a stopwatch. This is the most important row, and it is the one everyone skips.
4. Time per item after rollout. Checking and correction included, not just the machine run.
5. Error rate and rework, before and after. If the error rate goes up, correction takes back part of the time you saved.
6. Costs. The one-off build, and the monthly running line: licences, API, oversight time.

There is no way round the third row. If you cannot spare three days, measure for one and note that it is a one-day sample. Even that is more comparable than a number pulled from memory.

A worked example

Processing inbound quote requests at a 40-person trading company. The figures are example values from a typical range, not any customer’s data.

Item	Value
Quote requests per working day	18
Working days per month	21
Manual time per item, before	11 minutes
Time per item after, including checking	4 minutes
Fully loaded hourly cost	18 euros
Realisation rate	65 per cent
One-off build	2,800 euros
Monthly running cost	126 euros

The calculation step by step:

• Time saved per item: 11 – 4 = 7 minutes.
• Time saved per month: 7 × 18 × 21 = 2,646 minutes, so 44.1 hours.
• Gross value of that time: 44.1 × 18 = 794 euros per month.
• Realised saving at 65 per cent: 516 euros per month.
• Net monthly benefit: 516 – 126 = 390 euros.
• Payback: 2,800 / 390 = 7.2 months.

This calculation is useful even when the answer is no. Run the same process at four requests a day and the net monthly benefit falls below 126 euros, so the project never pays back at all. Better to know that before than after. You do not need a round of quotes to get the order of magnitude: prices are published in the price list, and you can size your own case up front with the calculator.

What to leave out of the maths, and what not to automate

Keep hard-to-measure benefits out of the calculation. A better customer experience, less stress and faster response times are real advantages, but if the payback rests on them, then numbers did not decide the project. Track them on a separate list, as notes.

In three places the question is not payback at all, but whether it is allowed.

• Final decisions without a human. Hiring, dismissal, screening out a candidate, credit assessment. Some of this falls into the high-risk category under the EU AI Act and requires human oversight. Here AI produces a suggestion, a person makes the decision, and that has to be documented.
• Sending to customers without review. In the first month, have a person read every outgoing text. After that, sample 10–20 per cent on an ongoing basis.
• A rare or undocumented process. Below eight runs a month the build cost rarely comes back. And if nobody can describe the process, write it down first and automate second. Automated chaos is just faster chaos.

We wrote separately about which processes to leave alone, and you will find it on the blog.

The first thirty days

Four weeks, four steps, and a number at the end.

1. Week one. Measure. Fill in the six rows for two or three processes. No purchases, no tool demos.
2. Week two. Calculate. Run the formula on each one and rank them by net monthly benefit. The number sets the order, not enthusiasm.
3. Week three. Take the lowest-risk part of the top-ranked process and start there, with human review.
4. Week four. Measure again with the same method. If there is no improvement, the project stops rather than expands.

Training carries the same weight as the tool. If the colleagues who run the process do not understand what the system sees and what it does not, after a month or two they go back to the old way and the measured saving evaporates. That is not a technical question. It is a question of whether measurement and training sit inside the same project.

Fill in the six rows, run the formula, and six months from now you will have data about your AI project rather than an opinion.

In manufacturing and logistics, the conversation about AI almost always starts at the machines: vision systems on the line, predictive maintenance, route optimisation. These are real, but in a company of 10–250 people they are usually not where the fastest measurable gain sits. The measurable gain is in the paperwork around the line.

There is a simple reason for that. A project tied to the production line needs sensors, data collection, integration and months of work, and it carries the risk of interfering with output. Paperwork, by contrast, is already digital, it runs on written rules, and when it breaks, production does not stop. That is why it is the sensible place to start.

Why manufacturing and logistics differ from an office

Three things are different. First, the documents do not come from one place. In a mid-sized plant, the incoming material arrives with a delivery note, a material certificate, an order confirmation and an invoice, from four suppliers in four different formats, some of it scanned from paper. Second, shift work means information is handed over two or three times a day, and it loses something at every handover. Third, most of the know-how is written down but hard to find. Machine manuals, SOPs, old fault reports, all in PDF, somewhere on a network drive.

All three are text processing problems, not engineering problems. That is good news, because text processing is the most mature and the cheapest part of AI today. It needs no new machine and no shutdown.

Four processes where the impact is genuinely measurable

The figures below are typical industry experience ranges, not guaranteed results. Your own numbers can easily differ, which is why measurement always comes before implementation.

Process	Typical time today	After implementation	Time to live
Processing incoming supplier documents	3–5 minutes / document	1–2 minutes / document	3–5 weeks
Shipping paperwork (waybill, packing list)	6–10 minutes / shipment	2–3 minutes / shipment	4–6 weeks
Shift handover and fault log summary	15–25 minutes / shift	5 minutes / shift	2–3 weeks
Internal document search (manuals, SOPs, drawings)	20–40 minutes / person / day	5–10 minutes / person / day	4–8 weeks

Incoming supplier documents

Reading the delivery note and order confirmation, matching them against the purchase order, and listing the discrepancies. This is the classic case where AI does not decide, it only prepares: it fills the fields and flags where quantity or price differs. Approval stays with a person. The gain shows up because purchasing stops searching and starts checking.

Shipping paperwork

Producing waybills, packing lists and accompanying documents from order data, and matching returned documents back to the shipment. Most of the time here does not go on filling forms, it goes on collecting data from three systems. Once that comes together in one place, the filling itself is quick.

Shift handover and fault log

The fastest payback and the least popular project, because it does not sound exciting. End-of-shift notes, downtime records and fault reports become a one-page summary for the next shift supervisor. After two weeks it usually turns out the recurring faults were visible all along, nobody had just read them together.

Internal document search

A search tool that understands your own manuals, SOPs and earlier fault reports, and answers with a citation. This is not a chatbot, it is a search tool: every answer states which page of which document it came from. Without the citation it is unusable, because it cannot be verified.

A worked example: a 60-person plant

Take a manufacturing company with 60 staff, three shifts and 22 working days a month. They start two processes, not four.

• Incoming supplier documents. 900 documents a month, 4 minutes each on average, which is 60 hours. At a 50 per cent saving, 30 hours a month come back.
• Shift handover. 3 handovers a day over 22 days is 66 handovers. The time drops from 20 minutes to 5, so 15 minutes each, 16.5 hours a month in total.

That is roughly 46 hours a month. At an internal hourly cost of 18 €, it comes to about 830 € a month. The one-off implementation cost at this size is typically 4,500–7,000 €, with running costs of 60–150 € a month. Net monthly saving is therefore around 700 €, and payback lands at 7 to 10 months.

This calculation is a frame, not a promise. What matters in it are the two numbers you have to measure yourselves: how many units, and how much time per unit. Everything else follows from those. If you want a quick order-of-magnitude figure, the calculator runs the same logic on your own data.

What not to automate in manufacturing

There are four areas where the cost-benefit ratio is poor, or the legal risk is out of proportion.

1. The final quality decision. AI can flag that a batch is probably defective. Whether it ships to the customer is signed off by a person. If the decision runs without a signature, accountability disappears with it.
2. Safety functions and machine control. Emergency stop, guard logic, interlocks. These are standardised, deterministic systems. A language model does not belong in that chain.
3. Assessment of people. Performance reviews, shift allocation on disciplinary grounds, candidate screening. The EU AI Act treats these uses as high risk, and the cost of compliance is considerably higher than the benefit.
4. One-off, non-repeating work. If a process runs twice a month, maintaining the automation costs more than the work itself. The rule is simple: without at least weekly repetition and at least 20 units, leave it in human hands.

On top of that comes one general limit: wherever the output goes straight to a customer or an authority, there must always be an approval point. Not because AI errs often, but because even a rare error is expensive and hard to explain after the fact.

How to start: the first four weeks

In the first two weeks you build nothing, you measure. Three numbers need writing down: how many units go through the process, how much time goes on each one, and who does it today. Those three numbers set the order, not enthusiasm.

In the third week one process starts, the lowest-risk one. At the end of the fourth week you compare the measured time with the figure from before. If there is no improvement, the project stops rather than expands. This is the step most companies skip, and it is why they cannot say six months later whether it was worth it.

Training matters as much as the tool. If the plant manager and the shift supervisors do not understand what the system sees and what it does not, they will be back in Excel within a month. Article 4 of the EU AI Act also requires adequate AI literacy from staff, and we have written about that separately on the blog.

Our prices are published, so you do not need a round of quote requests to get an order of magnitude: they are in the price list. At the end of an audit you get a priority order, including which process is best left alone this year.

Accounting firms have an unusually good starting position with AI. Most of the work repeats, the input is a document, and the rules are written down. Payback shows up sooner here than in most industries.

This article walks through four processes in the order worth implementing them. For each one: the typical time saving, the rough cost of implementation, and the point where a person has to stay in the loop.

Why this particular order

The order does not follow the size of the saving. The first process comes first because it is cheap, live within two to three weeks, and it shows the team that this works. The biggest item comes second: it needs the routine from the first project, and it needs clean data.

One shared figure runs through the whole article: the internal hourly cost. Salary, contributions and overheads together typically sit between 15 and 22 euros for an accounting role in the region. We use 18 euros throughout. You can work out your own in a minute, and the calculator gives you the order of magnitude too.

One caveat: the figures below are typical industry ranges, not promises. Your own measurements apply to your firm, and two weeks of measuring is enough.

The four processes, in order

1. Chasing missing documents

Every month starts the same way: who has not sent what. Today that means a manual list, then emails, then phone calls, and someone keeping the gaps in their head.

What a machine can do: compare received documents against expected ones, send reminders client by client, watch deadlines, and deliver a summary of everything outstanding each morning.

Typical effect: 60–80% of the time spent chasing disappears. In a firm of 5–8 people that is 10–15 hours a month, or 180 to 270 euros.

Implementation: 2–3 weeks, and it is the cheapest of the four.

What to watch: the tone of a reminder is a client relationship, not a technical question. Write a template, and keep an exception list for clients a person should still contact personally.

2. Data extraction from incoming invoices, with a posting suggestion

This is the big item. The supplier, document number, date of supply, net amount, VAT rate and line items can all be read from the image of an invoice. The system then proposes a posting based on previous bookkeeping. The accountant no longer types, they check and approve.

Let us put numbers on it. A firm processes 1,200 incoming invoices a month, and entering one by hand takes 5 minutes. That is 100 hours a month. Extraction plus the suggestion typically cuts 50–60% of that, so settle on 55 hours. At an internal cost of 18 euros an hour, that is 990 euros a month. If implementation costs 4,000 euros, payback lands at around four months.

What to watch: recognition accuracy in the first month is typically 85 to 95%, and improves as patterns are learned. The key is not accuracy itself but making sure uncertain items go into a separate queue rather than disappearing into the approved pile. Ask for a confidence indicator on every line.

Implementation: 6–10 weeks, because tidying up formats and the supplier master data also takes time.

3. Internal document and knowledge search

The third process is the dullest of the four, which is exactly why it is good. The system searches the firm’s own material: internal policy, earlier written opinions, client folders, extracts from legislation, questions already answered by email. A colleague asks in one sentence and the system answers with the source named.

Typical effect: an accountant spends 20–40 minutes a day searching and asking colleagues. Half of that realistically disappears: 4–7 hours per person per month. At six people that is 25–40 hours a month, worth 450 to 720 euros.

Implementation: 3–5 weeks if the documents are in one place. If they are not, that is the first half of the project.

What to watch: a search tool only knows what has been written down. If the knowledge sits in people’s heads, write it down first: useful work in its own right. Permissions must be inherited: if someone cannot see a client folder, search must not show it either.

4. Suggested replies to client emails

A large share of incoming questions repeat: filing deadlines, what to send, where a matter stands, what data the authority wants. The system classifies the message, drafts a reply from your own material, and the accountant reviews, corrects and sends it.

Typical effect: response times get shorter, and writing time on routine messages drops by 30–50%. Reviewing one draft takes ten to forty seconds, so most of the saving survives.

Implementation: 4–6 weeks.

What to watch: this is the riskiest of the four, because the output goes to a client. There is no automatic sending. The machine suggests, a person sends, and that part is not negotiable.

What it adds up to

Taken together, the four typically cut into a monthly workload of roughly 90–130 hours in a firm of 5–8 people, and total implementation lands between 7,000 and 12,000 euros. Overall payback is usually 5–9 months if you roll them out one after another.

Process	Typical saving	Implementation	Risk
Chasing documents	10–15 hours a month	2–3 weeks	low
Invoice extraction	50–60 hours a month	6–10 weeks	medium
Internal search	25–40 hours a month	3–5 weeks	low
Suggested replies	8–14 hours a month	4–6 weeks	higher

These are ranges, and the top end requires clean data. If supplier master data is messy or folders are scattered, the top end will not materialise.

In an accounting firm the question is not whether a machine can do the books. The question is what no longer has to be retyped by hand.

What not to automate, and where a person stays in the loop

Accounting is a regulated profession and mistakes carry a price. Four points where the machine stops and a person decides.

• Filing with the authorities. Final approval of a return, report or submission stays with a person. The machine prepares, a person presses the button.
• Posting unusual or disputed items. A suggestion on routine items is fine. An unusual case, a borderline case or a classification question is professional judgement, not pattern recognition.
• Tax advice and written opinions. When a client asks what they should do, there is liability behind the answer. The machine can draft it, a person signs it.
• Payroll exceptions. An ordinary month can be automated. Sick pay, severance, deductions and legal disputes cannot.

Two rules that work every time. Keep a log of what the system suggested and who approved it, because you will need it after the fact. And set a value threshold above which a second pair of eyes is mandatory, typically 500 to 2,000 euros depending on firm size.

Data protection and the AI Act, briefly

An accounting firm handles other people’s data, which means stricter rules than in an average company. Sort out three things before the pilot.

1. Where the processing happens. Processing in the EU, an account that does not train on your data, a written data processing agreement. Client data does not belong in a free consumer account.
2. Who sees what. The AI tool’s permissions must not be broader than a colleague’s. Inherited access per client folder is the right baseline.
3. AI literacy. Under Article 4 of the EU AI Act, staff must have an adequate level of AI literacy. A documented internal training session twice a year satisfies it; a separate article on the blog has the details.

How to start

Do not launch all four at once. Measure for two weeks: how many invoices arrive, how long one entry takes, how much time goes on chasing documents. Then start with document chasing: it is live within three weeks, and after that the team stops arguing in principle and starts looking at numbers.

If you want the order of magnitude in advance, our prices are published, so a decision does not need a round of quote requests. At the end of the audit you get a sequence, including which process to leave alone this year.

Most automation projects do not fail on the wrong tool. They fail on the wrong process. The technology works, the invoice gets paid, and six months later nobody opens it.

This article is not about what to automate. It is about what not to automate. We go through five typical dead ends, each with a test that filters them out in a few minutes.

When automation is not worth it

The payback of an automation rests on three numbers: how long the process takes per run, how often it runs, and what it costs to build. We publish the third one, you have to measure the first two. A stopwatch and two weeks are enough.

Take a good candidate. The process takes 25 minutes per run and runs twelve times a week. That is roughly 21 hours a month. If you use an internal hourly cost of 18 euros (wage, contributions and overhead together), the process costs 378 euros a month. Automation typically removes 60–70% of the time spent, so let us stay with a real saving of around 250 euros a month. A 1200 euro implementation then pays for itself in about five months.

Now take a bad candidate. The process takes 40 minutes per run, but runs only twice a month. That is 80 minutes a month, or 1.3 hours, which at the same rate is 24 euros. If the implementation costs 2500 euros, payback runs past eight years. The process is genuinely annoying in the meantime, but the invoice is real.

These figures are typical industry ranges, not promises: for your processes, your own measurements are what count. You can estimate the order of magnitude in a few minutes with the calculator.

Annoying and expensive are not the same thing. It is the expensive one that is worth automating.

The five dead ends

1. The rare process

A task that runs once or twice a month almost never pays off. The implementation costs the same as it would for a daily process, but the saving is a fraction of it. A rarely used automation also breaks quietly: by the time you need it again, a form or an access right has changed.

The test: if the process runs fewer than eight times a month, put it at the end of the list and come back to it once the frequent ones are done.

2. The process nobody has written down

If three colleagues do the same thing in three different ways, you do not have a process, you have a habit. Automation will not tidy up that chaos, it will set it in concrete, in the version used by whichever colleague happened to sit in the kick-off meeting.

The test: can you describe the process on one page, step by step, well enough for a new employee to follow it? If not, write that first. It is an afternoon of work, and it often cuts the time on its own.

3. The final decision

A machine may produce a recommendation, a person makes the decision. This is not philosophy, it is risk management. Hiring, dismissal, a pay rise, a credit assessment, individual pricing, rejecting a complaint, terminating a contract: for these, the cost of a wrong decision is orders of magnitude higher than the ten minutes saved.

The EU AI Act also names several of these cases explicitly and treats systems supporting employment decisions as high risk. If the topic is new to you, there is a separate article on the practical steps under Article 4 on the blog.

The test: if the output concerns somebody’s money, job or rights, it can be a recommendation, but not an automatic action.

4. The data you do not have

An automation is worth exactly as much as the usable input it receives. If the information sits in a colleague’s head, deep in a mailbox or on paper, what is missing is not automation but data. The first half of the project then quietly turns into data cleanup, and the budget ends up twice what was planned.

The test: can you show where all the data the process needs is kept, and is it in machine-readable form? If the answer takes more than fifteen seconds, start with the data.

5. The process that should be scrapped

The fastest and cheapest automation is deletion. Many companies run a weekly report nobody reads, an approval round that made sense ten years ago, and copying between two systems, one of which is no longer in use.

The test: stop the process for two weeks. If nobody says anything, it does not need automating, it needs scrapping.

Where a human has to stay in the loop

Beyond those five there are a few points where the question is not whether it pays off, but where the machine stops. For these, the output always stays a recommendation and a person releases it:

• anything that goes out to a customer in the company’s name: a quote, a complaint response, a notice, a payment reminder
• movement of money: initiating a payment, changing supplier details, approving a discount
• HR decisions: screening, promotion, performance review
• legal and contractual text, as well as any filing with an authority
• anything irreversible: deletion, publication, sending a bulk email

In practice this takes less time than you would expect. The machine prepares the recommendation, a person checks it and releases it. Reviewing a good draft text takes between ten and forty seconds, so most of the saving survives.

Two rules that work in every such case. Keep a log of what the system recommended and who approved it. And set a value threshold above which a second approval is always required, typically between 500 and 2000 euros depending on company size.

Four questions before you start

Before you ask for a quote on any process, run through these four. A no on any of them is not necessarily disqualifying, but it does mean that is the thing to fix first.

1. Does it run at least eight times a month?
2. Is it written down on one page, and does everyone do it the same way?
3. Is all the data it needs available in machine-readable form?
4. Can the output stay a recommendation, or must it be executed automatically?

Criterion	Good candidate	Bad candidate
Frequency	several times a week	once or twice a month
Rules	written down, stable steps	every case is different
Data	in a system, searchable	in someone’s head, email, paper
Risk	errors are cheap to fix	legal or financial consequence
Typical payback	3–8 months	several years or never

What to do instead

If your first idea did not survive the four questions above, that is good news, not bad. Three areas remain where the effect is quickly measurable in most companies of 10–250 people:

• Internal document search. Boring, but the fastest paying project there is, because every department looks something up every day.
• Draft text for repeated correspondence. Quotes, replies to enquiries, support templates. The machine drafts, a person sends.
• Data entry and document processing. Incoming invoices, delivery notes, forms. Stable rules, high volume, checkable output.

The same order works for all three: measure first, write it down second, and only then pick a tool. That is why every project with us starts with an audit, and our published prices let you work out the cost in advance. The other articles are on the blog.

In most companies the question is already settled: colleagues use AI chat, they just do not always say so. A ban changes little, it only makes the usage invisible. What works is this: you give them an account you can attach a rule to, and you know where the data goes.

What happens to the data you paste in

Three things are worth separating, because the question „is this safe?“ has a different answer for each.

• Training: whether the provider uses your text to train the model further. On free and personal paid accounts this is often on by default and you have to switch it off yourself. On business plans (Team, Business, Enterprise, API) your data is not used for training by default.
• Retention: how long the conversation is stored. Under a business contract this is typically 30 days for abuse monitoring, and larger contracts can be set to zero retention.
• Place of processing: which regional servers the request runs on. EU data residency is available with most large providers today, but it is not always the default and it does not cover every feature.

From a GDPR point of view the essence fits in one sentence. If personal data goes into the chat, the provider is a processor and you need a written data processing agreement with them. A free account does not come with one. This is not a technical question but a legal one, and that is exactly why good intentions will not get you around it.

A five-minute check: open the settings and look at whether model training is switched off, where the organisation’s data region sits, and whether a signed processing agreement exists for the account. If you cannot answer even one of those three, you are still on the lowest level, regardless of how much you pay each month.

The three levels and what they cost

In practice there are three setups. For most companies of 10–250 people the middle one is enough, and the third is needed less often than you would think. The prices below are typical market ranges, not quotes.

Level	Typical cost	What it covers
Personal account (free or individual subscription)	0–25 € / month / person	Public information, drafting, brainstorming. No customer data.
Business subscription with a processing agreement and an EU region	25–30 € / month / person	Internal documents, customer correspondence under rules, daily work.
Model running in your own environment (private API endpoint or on-prem)	3 000–8 000 € one-off, then 300–900 € / month	Regulated data, where a contract on its own is not enough.

The third level is more expensive, slower and needs more maintenance. Choose it only if you hold data that may not leave the company even under a contract. If you do not, the second level is the better decision, because it can start tomorrow.

What not to paste, and where a human stays in the loop

Keep the rule short, or nobody will read it. One page is enough, as long as it contains this list.

The other half of the policy is not about what you must not paste, but about where the machine must not have the last word. A human stays in the loop in four places, without exception.

• Decisions about people: hiring, CV screening, promotion, dismissal, disciplinary cases. The EU AI Act treats these as high-risk uses, and the liability sits with the company even when the machine wrote the recommendation.
• Financial or legal positions that go to a customer. AI writes a good draft. A person signs it.
• Anything leaving the company: quotes, pricing, complaint handling. A quick read-through is enough, but it has to happen.
• Any figure the chat returns as fact. Check numbers against your own systems, not against the answer.

These four points do not slow the work down. Reading a quote takes two minutes. Withdrawing a bad quote takes two days.

One thing is worth saying out loud: liability does not transfer to the provider. If a quote written with AI carries the wrong price, it is your quote. If a candidate screened with AI challenges the decision, it is your decision. So it is not enough to write the control points down. Build them into the process, for example as a line in the quote template that somebody has to tick.

A worked calculation

Take a 20-person company where 12 people get a business account. Based on typical experience ranges, a daily user saves 15–30 minutes a day on correspondence, summaries and drafting. Let us use the middle of the range, 20 minutes.

12 people × 20 minutes × 20 working days = 80 hours per month

80 hours × 18 € full hourly employer cost = 1 440 € per month

Licences: 12 × 28 € = 336 €. Net per month: roughly 1 100 €.

The one-off items: writing the policy and running a three-hour practical training session typically come to around 600–1 200 €. Against the net figure above, that pays back within a month.

The calculation only holds if the tool is actually used. If four of the 12 accounts are active, the net drops to a third while the licence fee stays. An account on its own is therefore not the solution: training is not an extra line item, it is the condition for payback. Use the full employer cost for the hourly figure rather than net pay, and measure the minutes saved instead of estimating them. A week of measurement is more accurate than an hour of debate.

Two weeks is enough to roll it out

This does not need a project. Two weeks, half an hour a day.

1. Days 1-3: inventory. Ask who uses what today, and for what. Not as a punishment: this is how you find the three tasks that genuinely need it.
2. Days 4-5: decide the level. Using the table above. In most companies that is the second row.
3. Days 6-8: set it up. Assign accounts, switch on the EU region, switch off training, sign the processing agreement, close down work use of private accounts.
4. Days 9-10: the one-page rule. The do-not-paste list above, the four human control points, and one name people can ask.
5. Days 11-14: training. Two to three hours on your own tasks. Not prompt theory, but the three use cases that came out of the inventory.

Article 4 of the EU AI Act has required since February 2025 that staff working with AI have an adequate level of AI literacy. Training is therefore not optional but an obligation, and it happens to be the exact step that delivers the payback as well.

Review it once a quarter: who uses it, for what, and whether the list needs changing. The whole thing takes ten minutes.

---

If you want to see which level is justified in your company, our published prices let you work out in advance what the policy and the training will cost. The other articles are on the blog.

The figure on the quote is the price of the project, not the price of the rollout. The gap between the two is typically 30–50%, and it is the part most companies only notice halfway through. Not because anyone hides it, but because most of the missing items do not arise at the supplier. They arise at your end.

The four items that make up the price of a pilot

The cost of automating one process fits into four boxes. Two are one-off, two are monthly. A miscalculation almost always comes from looking at the first box only, because that is the one printed on the quote.

There is a fifth item that never appears on a quote and is nonetheless real: your own internal time. On a pilot this is typically 10–20 hours. Collecting sample documents, talking through the exceptions, testing, giving feedback. It is not an optional part. If you leave it out, the workflow will not model your process but a generic one, and that difference can only be fixed later, at a higher price.

A model calculation: a 25-person company, quote preparation

The example below is a model calculation, not a client result. The inputs come from the middle of typical industry ranges, and every one of them can be checked against your own data.

The process: an incoming enquiry turned into a quote and sent out. Three people work on it, 90 enquiries a month, an average of 50 minutes each. After automation, the AI drafts the quote from previous quotes and the current price list, and a person checks it, corrects it and sends it: 20 minutes each.

One-off costs

Item	Amount
AI opportunity audit (credited against a pilot order)	€790 → €0
Building the pilot automation	€2,400
Half-day training	€990
Your own internal time (14 hours × €16)	€225
Total	€3,615

Monthly costs

Item	Amount
Licences (4 seats)	€90
Operations	€190
Total	€280 per month

The savings side: 90 enquiries × 30 minutes saved = 45 hours a month. The full employer cost of an office role in Slovakia and Hungary typically sits between €13–20 per hour; let us use €16. 45 × 16 = €720 a month.

That makes the monthly net 720 minus 280, or €440. Payback: 3,615 / 440 = 8.2 months. The first twelve months come to €8,640 saved against €6,975 in total cost, so €1,665 in the black. In year two, with no one-off items, it is €5,280.

Eight months. Not three weeks. Anyone promising three weeks either has not done the arithmetic, or has not done all of it.

When the calculation says no

The same process, at the same price, but with 30 enquiries a month instead of 90: 30 × 30 minutes = 15 hours, so €240 saved. The monthly cost is still €280. The monthly net is minus €40. This project never pays back, however long it runs, and no amount of good implementation fixes it, because the problem is the volume.

That gives you a simple threshold. A pilot is worth doing if the process frees up at least 20 hours a month. Below that, the licence and operations fees eat the entire benefit. At that point there are two workable routes: start with a higher-volume process, or push the monthly cost down with fewer seats, a simpler workflow, and no operations package for the first few months.

For a quick check you need three numbers: how many times the process runs each month, how long one run takes now, and how much of that would remain. The calculator takes you through it in a minute.

What not to automate

Not every expensive process makes a good pilot. Four cases where a person stays in the loop, even where the technology could do otherwise.

• Pricing and discount decisions. A machine can write the draft. A person signs off the final number, because a mistake here is money directly, and it does not surface on its own.
• Outgoing customer text with no review. A suggested reply, yes. Automatic sending, no. The review can shrink from 20 minutes to 20 seconds, but it cannot drop out of the process.
• Recruitment screening. Machine ranking of CVs is a high-risk use under the EU AI Act, with obligations of its own. A first pilot is not the place to take that on.
• Exception handling. If a third of the runs are one-off cases, automate the regular two thirds and leave the rest to people. Maintaining a workflow that tries to cover every case costs more than it returns.

A good first pilot is high-volume, repetitive, has a checkable output, and its errors are cheap. If any one of those is missing, look for another process. The technology is not what decides this. Those four conditions are.

What the calculation usually leaves out

1. Cleaning up the data. If half of your older quotes are scanned PDFs, the first step is not AI but tidying up. Typically 5–15 hours of internal work, once.
2. Running in parallel. For the first two to three weeks the old and the new method live side by side, until trust builds. So the saving does not start on day one but from roughly the second month.
3. Article 4 of the AI Act. Anyone using AI on behalf of the company needs adequate literacy, and has done since February 2025. Training is therefore not an extra line item but an obligation, and it is worth making it serve both purposes at once.
4. The cost of leaving. Ask up front: who owns the workflow, where does it run, and what happens to it if you change supplier. On your own infrastructure that cost is zero. On a closed platform it is not.

Work it out yourself in twenty minutes

Two formulas are enough.

Monthly net = (hours saved per month × hourly rate) − monthly licences − monthly operations

Payback in months = total one-off cost / monthly net

For the hourly rate, do not use gross salary but the full employer cost, divided by 160. And do not let the hours saved be an estimate: measure for a week how often the process runs and how long it takes. A week of measurement is more accurate than an hour of arguing about what you think it is.

If the payback lands inside 12 months, it is worth starting. Between 12 and 24 months, narrow the scope and recalculate. Above 24 months, do not start at all; find another process instead. Framed this way the decision takes twenty minutes, not three meetings.

---

The numbers differ from process to process, but the structure is always the same. To work through it on your own processes, see the published prices or the other articles on the blog.

When you ask a team who uses AI at work, fewer hands go up than the number of people who actually use it. In a company of 20–30 people, 8–15 different AI tools are typically in use, mostly on free accounts, and management does not know about most of them. That on its own is not the problem. The problem is that there is not a single page anywhere saying what may be put into them and what may not.

Why a verbal agreement is not enough

A verbal rule fails on three counts. First, nobody can point to it. If someone pastes customer data into a free chatbot, there is nothing to measure that decision against afterwards. Second, it does not scale. A new joiner does not know the established practice, so they fall back on their own judgement. Third, Article 4 of the EU AI Act has, since February 2025, required AI literacy from anyone who uses AI on the company’s behalf, and in an inspection a short policy plus an attendance sheet is the simplest evidence you can produce.

A policy is not a legal shield, though. It is a working tool. A good policy is one a colleague can skim in three minutes and then know what they are allowed to do. A policy nobody reads protects you from nothing.

The seven chapters an SME needs

You do not need a forty-page document. Three to four pages with seven chapters covers the large majority of companies with 10–250 employees.

1. Scope and purpose. Who it applies to (employees, subcontractors, interns) and to what: all generative AI tools, not only the ones named.
2. Approved tools. A short list, and next to each the kind of data it may be used with. Have a one-sentence process for introducing a new tool.
3. Data classification. Three levels are enough: public, internal, confidential. Plus which level may go into which tool.
4. Human oversight. What may not leave the company without a person’s approval, even after an AI has drafted it.
5. Labelling and transparency. When you must tell a customer that AI was involved. The transparency rules in Article 50 of the AI Act have applied since 2 August 2026.
6. Roles. Who approves a new tool, who takes questions, who reports an incident.
7. Training and review. A short mandatory briefing on joining, and a one-hour review every six months.

Three sample paragraphs you can reuse

These are general wordings; adapt them to your own company. They do not replace a legal review, but starting from zero they will put you weeks ahead.

Data classification

Confidential data (customer names, contract prices, personal data, source code, business plans) may only be entered into a tool approved by the company and running on a business subscription. Confidential data must not be copied into a free or personal account. If you cannot decide which category something falls into, ask before you paste it.

Approved tools

The list of approved tools is held in the internal document store. Anyone may propose a new AI tool by email, stating what task, what data and what account they would use it with. A decision follows within five working days. Before approval, the tool must not be used with company data.

Human oversight

Text, calculations or code produced with AI may leave the company only once a named colleague has read and approved it. The approver is responsible for the content. A tool’s suggestion is never a justification, so „the AI wrote it“ is not an acceptable explanation for a faulty output.

What to leave out, and where a human is required

The most common mistake is a blanket ban. It does not work, it only makes the usage invisible: people carry on using AI, just from their phones, and they stop asking questions. The second mistake is a list of named tools. Six months later it is out of date and the policy loses its credibility. Think in categories, not in brand names.

The third mistake is a promise you cannot keep. If the policy prescribes monthly logging but nobody logs anything, that is worse than not writing it at all.

There are, however, a few areas where a human really does belong in the decision chain. Name them explicitly:

• Hiring and pay decisions. CV screening sits in the high-risk category under Annex III of the AI Act. An AI suggestion may at most rank candidates; a rejection is issued by a person.
• Price, quote, contract terms. AI can draft it, a person approves the number.
• Rejections, dismissals, closing a complaint. If a decision has to be justified to a human being, it has to be made by one.
• Finalising an accounting entry. A suggested posting is a good accelerator, but the close is a human responsibility.
• Medical, legal or safety advice to a customer. Here the cost of a wrong output is disproportionately high.

How much work it is, and when it pays back

Written as internal work, the policy typically takes 6–8 hours. In our experience the split looks like this. Treat it as a typical industry range, not a promise:

Step	Typical effort
Tool inventory and short interviews	2–3 hours
Writing the policy	3–4 hours
Management review and sign-off	1 hour
Introductory training (2×90 minutes)	3 hours per person

The payback usually does not come from reduced risk. It comes from the ad hoc questions. In a 30-person company the question „can I put this into ChatGPT?“ typically comes up 4–6 times a week, and clearing one up costs a manager 15–20 minutes. That is 60–120 minutes a week, or 4–8 hours a month. At an internal cost of 25 €/hour that is 100–200 € a month. The one-off 6–8 hours therefore tends to come back within one to two months, and onboarding new colleagues gets faster after that.

If you want to see what the number looks like for you, the calculator lets you run it on your own hours.

The first 30 days

1. Week 1: inventory. Ask team by team who uses which AI tool, and on what account. Do not punish, just collect. The list is always longer than you expect.
2. Week 2: draft. Write the three to four pages against the seven chapters. One chapter, one page at most.
3. Week 3: review. A management round plus one read from a data protection angle. If personal data is involved, get a legal eye on it here.
4. Week 4: publish and train. Two short sessions with practical examples from your own processes, and an attendance sheet. That sheet doubles as evidence for Article 4 of the AI Act.

After that, one hour every six months: update the tool list and fold in the questions that have come up. Do not plan for more, because it will not happen.

---

The policy is not the goal. It is the frame inside which automation can safely start. If you want to know which process is worth tackling first, see the other articles on the blog or the published prices.

In the summer of 2026 a lot of managing directors breathed out: the Digital Omnibus deferred the rules for high-risk AI systems. Article 4 of the AI Act did not move, though. It has applied since 2 February 2025, and it applies to you too if your team uses ChatGPT, Copilot or any other AI tool at work. Here is what that means in practice for a 20-person company.

What Article 4 says, and what it does not

Article 4 of the AI Act (Regulation 2024/1689) is about AI literacy. It covers two roles: whoever develops an AI system and puts it on the market (the provider), and whoever uses it in their own activity (the deployer). The vast majority of companies with 10 to 250 employees fall into the second category.

The obligation itself is simple: your staff, and anyone operating an AI tool on the company’s behalf, should understand at an appropriate level what the tool does, where it gets things wrong, and who carries the responsibility. The regulation does not prescribe a course type, a number of hours, an exam or a certificate. There is no official „AI Act certificate“, and anyone selling you one is selling marketing.

What does have to be taken into account: the prior knowledge of the people involved, the type of system in use, and who is affected by how that system operates. A graphic designer and an HR officer do not need the same training. Article 4 also reaches beyond your own payroll: it covers anyone operating the tool on the company’s behalf, including subcontractors, freelance marketers and agencies.

What changed in the summer of 2026

The Digital Omnibus was signed on 8 July 2026 and entered into force on 27 July. What it brought:

• High-risk systems (Annex III): the deadline moved from 2 August 2026 to 2 December 2027.
• AI embedded in products (Annex I): 2 August 2028.
• Transparency rules (Article 50): unchanged at 2 August 2026, with the watermarking duty applying from 2 December 2026.
• Article 4: the wording was softened. Previously an appropriate level had to be „ensured“, now its development has to be „supported“. The duty became effort-based rather than result-based. It did not disappear, and its date did not move.

In practice this means you do not have to prove that everyone understands AI. You have to be able to show that you did something about it, regularly and traceably. In a 20-person company that is not months of work, it is a few working days, provided you do it in the right order.

Eight steps that are enough for a 20-person company

1. Tool inventory. One table: who uses which AI tool, with what account, with what data. This is the step most often skipped, and the source of most surprises.
2. Risk classification. Check whether any of them is a prohibited practice under the AI Act (Article 5), and whether any of them decides about people in HR, lending or access management.
3. An approved toolset. Decide what may be used: EU data processing, company accounts, data not released for training. Private-account use is the single biggest risk in a small company.
4. A two-page AI policy. What may be entered, what may not, what needs approval, and who to turn to when something goes wrong.
5. Basic training for everyone who touches it. Half a day is enough: what the tool is good for, where it hallucinates, how to check the output, and what must never be pasted into it.
6. Targeted training for the riskier roles. HR, finance, customer service, legal. They need another hour or two on their own cases.
7. Documentation. Date, participants, training material, the name of the person responsible. That is the evidence, not the good feeling.
8. A six-monthly review. Put it in the calendar. A tool list goes stale within half a year.

How much time and how much money

Take a 20-person company where 12 people touch an AI tool. The figures below are typical ranges based on industry experience, not guaranteed values:

Step	Effort	Note
Tool inventory and roles	2–3 hours	one person, one morning
AI policy (2–3 pages)	3–4 hours	faster from a template
Basic training for 12 people	3 hours × 12 people = 36 person-hours	one group, half a day
Documentation	1–2 hours	this is the evidence
Six-monthly review	2 hours per round	booked in the calendar

That is 44 to 47 person-hours in the first round. At an internal cost of 20 to 25 euros an hour, roughly 900 to 1200 euros, plus an external trainer’s fee if you do not run it in house. Our published price list shows what it costs with us.

The other side: if ten of those twelve people save 15 to 30 minutes a day on routine text work (correspondence, summaries, minutes), that is 550 to 1100 hours a year. You can run the same figures with your own headcount in the payback calculator.

Article 4 is not satisfied by everyone becoming clever. It is satisfied by the company doing something about it, regularly and traceably.

Who supervises in Slovakia and Hungary

Since 2 August 2026 the market surveillance authorities of the member states have held their formal powers.

• Hungary: the AI Market Surveillance Authority (MI Piacfelügyeleti Hatóság) acts, as an organisational unit of the Ministry of Science and Technology, on the basis of Act LXXV of 2025 and Government Decree 344/2025.
• Slovakia: MIRRI SR is the designated general market surveillance authority and the single point of contact. The national implementing act was still in the legislative process at the time of writing.

One detail worth knowing: the AI Act’s catalogue of fines (Article 99) does not list Article 4 separately, so there is no EU fine directly attached to it. The detailed penalty rules are set by the member states themselves, though, and a failure on Article 4 surfaces in every case where something went wrong: an incident, a customer complaint, an employment dispute, a public tender or a large-company supplier audit.

What training does not solve

Literacy does not replace control points. There are a few things a small company must not hand to a machine without human approval, even with a well-trained team:

• CV screening and pay decisions. This is the high-risk category under Annex III. The deferral buys time until 2 December 2027, not an exemption.
• Price, deadline, contractual term. AI may draft the proposal, a person sends it.
• Rejection, dismissal, closing a complaint. If a decision has to be justified to a person, a person has to make it.
• Personal or customer data in a private account. That is a GDPR question first, an AI Act question second.
• Publishing AI content unlabelled. The Article 50 transparency rules have applied since 2 August 2026, watermarking from December.

---

If you got stuck on the first of the eight steps, that is entirely normal: in most companies it is the tool inventory that produces the surprise. More practical articles on the blog.

In most companies AI gets stuck at the stage where „someone uses ChatGPT“. That is not a bad start, but it is not a system. The five processes below are the ones a company of 10–50 people can get hold of fastest, with measurable results, and nobody on the team will hold them against you.

1. Handling incoming requests for quotes

The typical case: an enquiry arrives by email, someone reads it, copies it into the CRM, looks up a similar earlier quote and writes the new one. That is 20–40 minutes per customer, and most of it is mechanical copying.

• What we automate: pulling the data out of the email, the CRM entry and a draft quote based on your earlier quotes.
• What we do not: the pricing and the sending. A person approves those, in every case.
• Typical saving: 60–70% of the admin time, 15–25 minutes per quote.

2. Preparing incoming invoices for bookkeeping

From PDF to data, from data to a bookkeeping entry. AI has been doing this reliably for years, only in most companies nobody wires it in. The key is a review step where the bookkeeper corrects things with two clicks, not a decision to „just trust it“.

Good automation does not decide for you. It prepares the decision and leaves the responsibility with you.

3. Repeat customer questions

If 40% of the incoming questions are the same eight questions, you do not need to write the answers again every time. The right order is this:

1. Collect the 30 most frequent questions from the past six months.
2. Write consistent, approved answers for them. That is an afternoon’s work.
3. Build the suggested replies on those, not on free text generation.

4. Finding documents and earlier material

If the team spends 15 minutes a day looking for the latest version of a contract or last year’s price list, that is roughly 60 hours a year per person. A company search across your own documents is the „boring“ project that pays for itself fastest.

5. Putting reports together

A monthly report that someone copies together by hand from three systems. This is the most boring task in the company, which is why it makes the best first pilot: nobody will defend it, and the saving shows up in the calendar straight away.

---

If two of these sound familiar, that is already enough for a pilot. With the return on investment calculator you can work out what it would bring in your company, and in the use case catalogue you can see where your industry stands.