Real clients, real dates, and where possible, the client's own words.

Case Studies.

Seven pieces of real work. Two clients put their name to it; for four we keep the name confidential, but not the work. Next to them, our own company, where the same thing runs every night. Below them, nine scenarios: how the same would be built at your company.


01  /  Phone + the whole operation

Kakucsi Pékség Kft.

It started with an AI that takes phone orders. Today the till, the stock, supplier ordering and the daily close run on their own as well.

Client
Sándor Kakucsi, owner
Industry
Bakery and pizzeria, Gyál
First call
7 April 2026
Live
Every day since April 2026

The situation

A bakery and pizzeria in Gyál, open since 1997. Orders come in by phone, and the phone rings exactly when the queue at the counter is longest. Whoever can't pick up at that moment doesn't lose a call. They lose an order, and the customer who then dials the next number.

Sándor's first two questions

Can we keep our existing phone number? And won't the AI place phantom orders? The answer to both is built into the system, not promised. Customers call the same number they always have. An order is only recorded once every detail is there: if the name or phone number is missing, Rozi asks again, and the customer gets an SMS confirmation.

Step 1  ·  Rozi, the phone colleague  ·  April 2026

Three days after the first call we signed, and we built Rozi. Rozi doesn't take messages, she takes orders. She sees the menu and the prices live, carries the conversation through, and the finished order lands among the webshop's orders. She sends the invoice link by SMS and takes complaints too. She knows the opening hours by Sándor's rules: no pizzeria on Mondays but the bakery is open, and after 7pm she only takes pre-orders.

Step 2  ·  the whole operation  ·  August - September 2026

In the summer Sándor came back, and he didn't ask for an add-on. He asked us to move the shop's operation onto AI. Today, in the main shop, an order comes in by phone, online or at the counter, the till issues the receipt at the right VAT rate, the system deducts it from stock, and supplier orders follow the delivery days and cut-offs. If a goods receipt stays open too long, it flags it. In the evening the daily close is produced on its own and emailed to finance. Staff performance is tracked with a points system, and Rozi's support moved into the same system.

Step 3  ·  the website  ·  September 2026

In September he came back a third time, this time for a new website for the bakery. We built that too. Along the way he recommended us to another shop, which has since signed with us as well.

In numbers

22 callsanswered by Rozi on the busiest day in September
78 / 0operations and errors on one September day in the live till system
90%less paper, in Sándor's own words
3 projectsone after another, plus a referral to another shop

“Within a week it was handling phone orders, invoicing, stock management, supplier orders and reports automatically. Our paper use dropped by 90%, which not only made us more efficient but is a huge step forward in cost and sustainability too. The best feeling is that the AI works instead of us.”

Kakucsi Sándor Owner, Kakucsi Pékség Kft.

Sándor didn't say he was satisfied. He paid three times, and he recommended us.


02  /  Email customer service

Profitszakértő Kft.

Their old AI assistant stopped overnight. It was running again the same day, on a stronger engine.

Client
Ádám Szendrei, managing director · Éva Szabó, customer service
Industry
Financial education, Szeged
Rescue
26 August 2026, same day
Live
Since 8 September 2026

The situation

A company in Szeged offering financial education and a membership club. Questions, club matters and technical issues all land in one customer-service inbox, and Éva's team wants to answer every email quickly. An AI assistant helped with that, one we hadn't built. On 26 August 2026 the technology it ran on was shut down for good.

The same day  ·  26 August 2026

Éva wrote to our partner the day before. We spoke the next afternoon at two, and that same day, before the shutdown, we saved the full setup of both their assistants, every rule and every template. The same day we rebuilt it on the new technology, and it ran. Since everything had to be rebuilt anyway, we also built a second version alongside it, on a stronger AI engine.

The test week

The two versions ran side by side. In the first test round it put all six emails in the right category, and on payment notifications the new engine was clearly better. Along the way we found that the old system had been answering from an October 2025 template, and we fixed that too. We built 36 screenshots into the step-by-step answers, so customers don't just read the steps, they see where to click. At Éva's request the system checks the inbox every five minutes instead of once a day.

What broke, and what we fixed that same evening

On 5 September, a Saturday training day, the heavy traffic used up part of the system's allowance. That is something we should have seen coming. That same evening we rebuilt it so that every incoming email triggers it immediately, and the 15 emails that had piled up had their replies ready in six minutes. Éva's answer: “Get some rest instead :) Thanks for the quick response on a Saturday ❤️”

How it works today

Every new email goes into one of six categories: technical question, product enquiry, club cancellation, feedback, automatic notification or irrelevant. Where a reply is needed, the finished reply is waiting for the team in the email thread, with screenshots and signature, ready to go in one click. Every processed email gets a label, any error triggers an alert straight away, and the old version stays on standby, one click to switch back.

In numbers

Same daysaved and running again, on the day of the shutdown
5 minhow often it checks the inbox
6 / 6correctly categorised in the first test round
6 minto reply-ready for 15 piled-up emails

“Our customer-service AI assistant stopped from one day to the next, because the technology it ran on was discontinued. That same day they saved the full setup, took over the system, got it working again and moved it to a stronger AI engine. Since then the replies are much more accurate, friendlier and more human.”

Szendrei Ádám Managing director, Profitszakértő Kft.

“Since customer service was automated, the replies are much more accurate and need minimal human intervention.”

Szabó Éva Customer service, Profitszakértő Kft.

03 - 06  /  Confidential clients

The same kind of work, without the name.

Our contracts commit us to using a client's name only with written consent. These companies haven't given it, so we don't show the name, but we do show the work: what the problem was, what we built, and where it stands today.

03 Mobility provider · Budapest · Email customer service

[Client name confidential] Zrt.

More than 4,000 customer emails a month, in Hungarian and English. Thirteen days after signing, the system was working.

Signed
17 August 2026
Working
30 August 2026
Status
Test operation

The situation

According to the company's own figures, more than 4,000 emails a month arrive in the customer-service inbox, in Hungarian and English. The team answers the same questions over and over, while urgent cases must not get lost in the queue. They had one condition: customer data may not leave the company's own systems.

What we built

The colleague runs in the company's own cloud account, so the data stays where it always was. It categorises every email, detects its language and attaches a confidence score. It writes the reply from the company's 87-question knowledge base. Where it is certain, it may reply; where it is unsure, it hands the case to a person with a finished draft; anything urgent it forwards immediately. There is a kill switch, a daily report, duplicate protection and a full log. For acceptance we put together 50 test cases it has to pass.

Where it stands today

In test operation. The system works in draft mode on a test inbox, and the company's team is trying it with real emails. At the end of September they raised three points, one of them a Hungarian phrasing that didn't sound natural enough. We fixed all three in 70 minutes.

04 Accounting firm · Slovakia · Invoice processing

[Client name confidential] s.r.o.

An average of 6,000 invoices a month, entered by hand until now. The goal: the accounting software's own import file, with no typing.

Signed
July 2026
Volume
~6,000 invoices a month
Status
Being rolled out

The situation

A Slovak accounting firm keeping the books for several client companies. In their own words, they enter an average of 6,000 invoices a month by hand. The accounting software runs on their own server, with no outside access, and it has to stay that way.

What we built

The colleague takes in invoices arriving as PDFs and scans, reads the data, and checks the total, the VAT, and whether the invoice has already come in once. At the end it produces the accounting software's own import file, plus a review spreadsheet and a list of exceptions. A connector was installed on the firm's machine: it only reads the live data, imports into a test copy, and compares, item by item, what the machine recorded with how the firm booked it themselves.

Where it stands today

Being rolled out. In the test round it ran on 221 real invoices, and all 218 totals that could be compared matched. At the end of September the accountant found four discrepancies, one of them in the pre-coding of postings. After the fix, the posting codes matched the accountant's own decision in 87 cases out of 90, against none before. Before go-live, the accounting software's own response decides, not our test.

05 Marketing agency · Hungary · Website chatbot

[Client name confidential] agency

A chatbot that doesn't speak in generalities, but says what the agency itself would say.

Started
June 2026
Handed over
June 2026
Status
Handed over, closed

The situation

A web and marketing agency. The answers about their services already existed: in an introduction video, audio recordings and an FAQ document. A website visitor, though, asks questions, they don't watch videos.

What we built

We trained the website chatbot on those materials: the content of the video and audio recordings, and the FAQ. Alongside it we built a workflow that lets the agency add new material to the chatbot's knowledge without a developer.

Where it stands today

Handed over. Eight days after the start the updated version was ready, the final invoice was paid, and the project was closed.

06 Interior design studio · Abroad · Content system

[Client name confidential] Studio

The question was what to post, and what works in their market. The answer became their own content system, in English and Ukrainian.

Completed
September 2026
Status
Being handed over

The situation

An interior design studio working in two countries. Their work is strong, but social media got neither time nor method: it wasn't clear what to post, or what works in their market.

What we built

We built their own web-based content system, running on the studio's own AI subscription, so it adds no monthly cost. The first research round went through 18 competitors and picked out 24 standout posts, which became 14 ideas, 8 ready-to-shoot scripts and a four-week content calendar. Their own account data also showed that their 5.1% reach rate is several times that of the competitors examined (0.14-1.44%). 111 automated tests run on the system.

Where it stands today

Completed; the system runs at its own address, and the handover is in progress.


07  /  Our own company

Automating

Six AI colleagues, six jobs. What we build for clients, we run on ourselves, every day.

Team
2 people, 6 AI colleagues
Rhythm
Every night, hourly on workdays
Live
Since August 2026

Why our own company is here

Because here we see everything. Whatever we build for a client, we try on ourselves first, and our own operation runs on it too. Six AI colleagues work at Automating, and they aren't tools, they are jobs.

One night, while we sleep

  1. Leó, systems administrator. Goes first: checks every system and the website, and in every run fixes or improves something, including the others' work.
  2. Konrád, finance. Keeps the finances in order and tracks incoming money, deadlines and invoicing.
  3. Félix, marketing. Measures Instagram and TikTok, and writes a ready-to-shoot script for the next video.
  4. Viktor, new business lines. Measures how the new business lines are doing, and proposes what to double down on and what to close.
  5. Alfréd, operations. Goes through the whole company, updates the client records, prepares the emails and the calls, and by the time we start the day he has written, in three to five sentences, what we need to do.
  6. Rudolf, sales. Runs hourly on workdays: classifies every incoming email and handles the daily outreach.

The shared memory

Behind them is a shared memory we call the AI Brain: every client, every call, every decision in one place. All six work from it and write back into it. When we make a mistake, the mistake goes in as a rule, and it doesn't happen a second time.

Where we don't compromise

No outgoing email leaves without a machine check: right thread, right recipient, right sender, and no promise without a source. On its own, the system only sends clients a short acknowledgement. Anything about appointments, money or commitments waits for a human decision, with the reply already written. We build the same logic in for our clients: decisions stay with people.

In numbers

6AI colleagues, in six separate jobs
Every nightthe full round runs, with no one starting it
0outgoing emails without a machine check

Scenarios

How we would build it at your company.

Nine jobs an AI colleague can take over, and how we would build each one at your company. These are plans, not client case studies: the worked examples start from a stated assumption, so you can redo them with your own numbers.

For importers, wholesalers, accounting firms

Invoice and document colleague

Watches the inbox and the scan folder. Reads the invoice, delivery note or product data sheet, whatever form it arrives in. Checks the total, the VAT, and whether it has already come in once. At the end it produces the accounting or ERP system's own import file, ready to load in one click. Anything doubtful it doesn't guess: it puts it up for human approval, next to the original document.

We have worked on systems like this for an import distributor, turning product data sheets into data packages for their ERP, and for an accounting firm, turning invoices into accounting-software imports.

Worked example

If 1,500 invoices arrive a month and entering one takes 3 minutes on average, that's 75 working hours a month, just for typing. Close to half a full-time job.

For manufacturers, contractors, B2B webshops

Quote colleague

An enquiry arrives by email or form. The colleague reads what is being asked for and how much, builds the line items from your own price list and product data, and produces a finished quote in your branding, which the salesperson reviews and sends. If a detail is missing, it doesn't make it up: it writes the question to the client too.

The quote goes out in minutes, while the prospect is still at their desk, not days later when they have already asked someone else.

Worked example

If 20 quotes go out a week and putting one together takes 45 minutes, that's close to 65 hours a month of your best specialist's time.

For clinics, salons, service centres

Phone appointment booker

The same technology that has been taking orders at Kakucsi since April books appointments here. It answers the phone at night and at weekends too, finds out what is needed, finds a free slot in the calendar, books it and confirms by SMS. Anything sensitive or unusual it passes on with a summary, so reception picks up already in the picture.

The existing phone number stays; customers call the same number as before.

Worked example

If 5 out of 30 calls a day go unanswered, and every second one of those is a lost booking, then with an average treatment of 15,000 Ft and 22 opening days a month that is more than 800,000 Ft of revenue a month going elsewhere.

For staffing agencies, HR teams

Applicant screening colleague

Reads every incoming CV, whatever form it arrives in. Compares it with the open role's requirements, ranks with a short rationale, asks back for missing details, and offers suitable candidates an interview slot from the recruiter's calendar. The decision stays with the recruiter, who only reviews the ranking and the reasoning.

Worked example

If 300 applications arrive a month and screening one takes 6 minutes, that's 30 hours a month before a single interview happens.

For law firms, real-estate agencies, insurance brokers

Contract preparation colleague

Collects the details from emails, forms and attached documents, and fills in your own contract templates. Flags anything out of the ordinary: a missing annex, conflicting dates, an unusual clause. The finished draft waits on the case handler's desk with the risky points highlighted, and the case handler approves it.

Worked example

If 15 contracts are prepared a week and preparing one takes 40 minutes, that's close to 43 hours of specialist time a month.

For logistics companies, wholesalers, manufacturers

Order entry and coordination colleague

Orders arrive by email, PDF, Excel, everyone sends them differently. The colleague reads them and enters them into your system, sends the confirmation, and raises a flag if an item is out of stock or a delivery is slipping. The dispatcher deals with the exceptions, not the typing.

Worked example

If 40 orders arrive a day and entering one takes 4 minutes, with 22 working days a month that's close to 59 hours a month.

For estate agents, educators, service companies

Lead response colleague

Writes a personalised reply to every enquiry from the website, Facebook or email within minutes, at night and at weekends too. A few questions tell it how serious the interest is, and serious prospects get a slot in the salesperson's calendar straight away. Those who aren't ready yet are noted and contacted again at the agreed time.

Worked example

If, out of 60 enquiries a month, faster replies turn just three more into clients, at an average deal value of 200,000 Ft that's 600,000 Ft of extra revenue a month.

For managing directors, in any industry

Morning management briefing

Every morning it gathers the numbers from the till, the webshop, the accounts and the inbox, and writes in five sentences what happened yesterday, what is slipping and what needs attention today. Anything unusual goes to the top. The same runs for us every morning: Alfréd writes it.

Worked example

If putting together the weekly reports and spreadsheets takes 4 hours a week today, that's close to 17 hours a month of the manager's time.

For webshops, retailers

Stock and purchasing colleague

Watches sales and stock, and drafts the next order in line with supplier cut-offs, before a popular product runs out. Flags a goods receipt left open, or a product that hasn't moved for weeks. The same logic runs at Kakucsi, in supplier ordering.

Worked example

If your best-selling product runs out twice a month for three days, and it brings in 40,000 Ft a day, that's 240,000 Ft of missed sales a month.

Yours could be next.

A 30-minute call. We look at which job an AI colleague takes over first at your company, and tell you what it takes.