Case studies

AI case studies: 28 real projects, from problem to solution

I’m Michal Juhas, an AI implementation consultant based in Bratislava, Slovakia. This page documents 28 AI projects I’ve delivered for 16 companies and my own business: AI agents, custom CRMs, web apps, Excel macros and AI-built websites for accounting, controlling, recruiting, sales, marketing and management reporting.

Each case study gives the problem the team had, what I built, the form it took, and two scores from 1 to 10: how complex it was to build and how much value it brings. Every client is named except one, which is anonymized.

documented AI projects
28
companies
16
business areas
7
Key takeaways

What these projects show

  • Complexity doesn’t predict value.

    The most valuable project on this list, a VAT check across 15+ restaurants, is an Excel macro scored 1/10 for complexity and 10/10 for value. A restaurant’s inquiry app, by contrast, scored 6/10 for complexity and 3/10 for value.

  • 7 projects delivered a lot for little.

    That many projects scored 3/10 or less for complexity and at least 6/10 for value: an Excel macro, scripts and an agent in Codex, a custom GPT, a one-day hackathon, team training and a sales research agent.

  • An agent isn’t always the answer.

    Where the rules are fixed, as in processing applicants or school meal payments, deterministic software does the job. I use AI where something has to read text, judge or write.

  • Agents draft, people decide.

    The partner and donor agents find and qualify companies and people, then draft the email. A person still decides what gets sent.

  • 10 of 28 projects scored 8/10 or higher for value.

    And many clients have several projects with me: seven for a restaurant group and five for a trading firm.

All projects at a glance

Complexity vs. value

Each project, scored 1–10

Each dot is a project, plotted by how hard it was to build and how much it’s worth to the people using it. Tap or hover a dot to see which one.

Value →
Quick winsBig bets

Complexity →

5 projects

AI for accounting, controlling and finance

Finance teams lose days every month to retyping, downloading and cross-checking numbers between systems. These projects take that routine off accountants and controllers at a group of 15+ restaurants, from VAT checks and cash reconciliation to school meal invoicing.

Food & beverage

VAT check across 15+ restaurants

City Gastro

Complexity1/10
Value10/10
Problem
For every one of 15+ restaurants, the accountant downloaded the eKasa data and repeated the same admin steps by hand.
Solution
An Excel macro he runs on his own computer that does all of it in one go.
Built as
Excel macro
Tools and data
ExceleKasa

Food & beverage

Weekly cash reconciliation

City Gastro

Complexity6/10
Value7/10
Problem
Every week, the office worked out by hand how much cash each restaurant should be holding.
Solution
Automated the weekly calculation of how much cash each location should hold.

Food & beverage

Controlling for 15 locations

City Gastro

Complexity5/10
Value6/10
Problem
Every month, controlling spent several days pulling documents from different systems and checking that the numbers matched.
Solution
A custom web app where several people log in, upload the documents and run the checks.
Built as
Multi-user web app

Food & beverage

Management reports for 15 locations

City Gastro

Complexity2/10
Value5/10
Problem
One employee spent whole days retyping data from Pohoda into management spreadsheets.
Solution
A ChatGPT agent she runs on her own computer that builds the reports for her.
Built as
ChatGPT agent run on the user’s computer
Tools and data
ChatGPTPohodaExcel

Food & beverage

School meal payments from EduPage

City Gastro

Complexity4/10
Value5/10
Problem
Processing payments and meals served for three schools and 1,000+ diners, then invoicing the schools, took accounting several days a month.
Solution
A web app that takes the EduPage exports and produces the report for accounting and invoicing.
Built as
Cloud web app with deterministic processing
Tools and data
EduPage

3 projects

AI for operations and project management

Bids, menus and project obligations follow the same steps every time, which makes them good candidates for an agent. These tools prepare the repetitive parts, so people only review and decide, and they keep multi-company projects on track.

EU-funded project

Project management agent

EU-funded project consortium

Complexity6/10
Value8/10
Problem
After winning an EU-funded tender with several partner companies, management had to keep everyone within the tender rules, and nobody owned it.
Solution
An agent that knows every project contract, inside a web app where each person sees their own milestones and tasks and the admin sees everyone’s.
Built as
Multi-user web app with a project management agent

Food & beverage

Tender bid pricing agent

City Gastro

Complexity3/10
Value6/10
Problem
A solid bid for a public tender took two to three days, so the company could only enter a few.
Solution
An agent with a trained knowledge base that he runs in Codex on his own computer to draft and price bids.
Built as
Codex agent with a trained knowledge base
Tools and data
Codex

2 projects

AI reporting for management

Leaders need current numbers without asking someone to build a spreadsheet. These projects connect to the source systems and deliver the view a manager or CEO actually wants, live or on a schedule.

High-frequency trading

Live recruiting reports

Wincent

Complexity4/10
Value4/10
Problem
Leadership couldn’t see what the recruiting team was doing. The ATS reports fell short, and copying data into Excel by hand didn’t stick.
Solution
A custom site connected to the ATS that pulls live data and shows it exactly the way the manager wanted.
Built as
Custom web app connected to the ATS
Tools and data
ATS

Physiotherapy clinic

Weekly CEO report

SportWell

Complexity4/10
Value4/10
Problem
Data sat in several systems, and the CEO couldn’t easily pull it out or make sense of it.
Solution
A Claude Managed Agent that knows the business, runs every Tuesday morning, analyzes data from the source systems and emails the CEO a report.
Built as
Claude Managed Agent on a weekly schedule
Tools and data
Claude Managed Agents

3 projects

AI in recruiting and sourcing

Recruiting teams handle more candidates than they can read, and the best ones are often hidden in public lists. These projects process applicants at volume and find candidates in public competition results.

High-frequency trading

Olympiad participant sourcing

Wincent

Complexity5/10
Value9/10
Problem
Olympiads like IOI, IMO, IPT, IYPT and EGOI publish long results lists. Going through them by hand would take recruiters days.
Solution
AI-built software that scrapes competition sites, extracts prospective candidates and enriches them with contact details from ContactOut and CoreSignal.
Built as
Scraper with contact enrichment
Tools and data
ContactOutCoreSignal

High-frequency trading

Candidate processor

Wincent

Complexity4/10
Value8/10
Problem
A recruiter got over a thousand applicants a month for a single role and couldn’t keep up.
Solution
Deterministic software in Google Cloud that processes applicants, and keeps working while she’s on vacation.
Built as
Deterministic software on Google Cloud
Tools and data
Google Cloud

Recruitment agency

Autonomous recruiting agents

Synculario

Complexity3/10
Value5/10
Problem
Sourcing candidates took up most of the recruiters’ day.
Solution
Mapped what could be automated with the team, then deployed agents that help them source candidates faster.
Built as
Autonomous sourcing agents

6 projects

AI for sales, leads and partnerships

Finding, qualifying and answering prospects is repetitive work that decides revenue. These agents and custom CRMs route inbound leads, research partners and sponsors, and draft outreach for a person to approve.

E-commerce

YouTube partner sourcing

Hungry Minds

Complexity6/10
Value8/10
Problem
Several people were searching for YouTube channels to partner with, and it took far too long.
Solution
A web agent with an admin interface that searches on a schedule, qualifies channels against set criteria and keeps going until it hits the target number of partners.
Built as
Scheduled web agent with an admin interface
Tools and data
YouTube

Conference organizer

Sales research agent

Fleming

Complexity3/10
Value6/10
Problem
Finding the right companies and people for each conference took sales a lot of manual research.
Solution
An agent that researches the market, finds likely attendees and sends them personalized outreach. The research is strong, though the outreach hasn’t turned into replies yet.
Built as
Research and outreach agent

Non-profit

Fundraising prospecting agent

Gotická cesta

Complexity5/10
Value6/10
Problem
The chair of the non-profit had no efficient way to find partners who might donate.
Solution
A cloud agent that searches Finstate for companies that fit, qualifies them by location and finds the right people inside them to contact.
Built as
Cloud agent
Tools and data
Finstate

Real estate agency

Lead management CRM

BOND Reality

Complexity4/10
Value5/10
Problem
Dozens of leads a day came in from property portals, and two dispatchers handed them out to agents by hand.
Solution
A custom CRM for the whole team where AI reads each inquiry email and routes it to the right agent.
Built as
Custom CRM web app
Tools and data
nehnuteľnosti.sktopreality.sk

Restaurant

Inquiry inbox and booking overview

Hradná

Complexity6/10
Value3/10
Problem
The manager of one of Bratislava’s largest restaurants was buried in emails from prospective guests, mostly asking the same things, and business slipped away.
Solution
A cloud app that gives her an overview of every lead and answers the simple emails on its own.
Built as
Cloud app with automatic replies

5 projects

AI for marketing and websites

AI now builds production websites, writes in a company’s voice and runs ad campaigns through the API. These projects replaced an expensive agency redesign and manual campaign setup.

High-frequency trading

Website rebrand built with AI

Wincent

Complexity7/10
Value10/10
Problem
After a rebrand, the company didn’t want to pay a design studio €10–20k for a new website.
Solution
AI generated twenty website concepts, management picked from a shortlist, and the site moved off WordPress to AI-written source code.
Built as
AI-generated website, migrated off WordPress
Tools and data
WordPress

High-frequency trading

Copywriting GPT

Wincent

Complexity3/10
Value8/10
Problem
Marketing struggled to write strong copy for social media and the website, and managers had no time to write it themselves.
Solution
A custom GPT trained on company material, including podcast interviews, that writes in the company’s voice.
Built as
Custom GPT
Tools and data
ChatGPT

Online training

Automated Meta campaigns and reporting

AI with Michal

Complexity6/10
Value8/10
Problem
Setting up the twentieth campaign and the twentieth landing page by hand got so tedious that nobody wanted to do it.
Solution
A cloud agent that works with Meta through the API and emails the manager a report.
Built as
Cloud agent with Meta API access
Tools and data
Meta API

Non-profit

Sales website the CEO edits herself

TrustWomen

Complexity5/10
Value7/10
Problem
The old CMS was complicated, inflexible and dated, and every change needed a developer who was hard to reach and expensive.
Solution
An AI-built website the CEO updates herself through automatic pipelines, with Stripe payments built in.
Built as
AI-built website with automatic deployment
Tools and data
Stripe

E-commerce

AI segmentation quiz

Axolt

Complexity3/10
Value5/10
Problem
Every visitor saw the same offer, whatever problem they came with.
Solution
A quiz that uses AI to segment visitors by their problem, so each one gets a tailored offer on the first visit.
Built as
AI-scored quiz

4 projects

AI in software development and team training

Sometimes the most valuable project is teaching a team to work with AI themselves. These engagements covered agentic engineering for developers, AI-assisted Odoo development, and hands-on training for recruiters and leaders.

Defense tech

Agentic engineering hackathon

SEC Technologies

Complexity2/10
Value8/10
Problem
Management wanted the development team to ship faster.
Solution
A full-day hackathon where the team learned to build software with an agentic engineering methodology.
Built as
Full-day hackathon

IT consulting

AI-assisted Odoo development

26HOUSE

Complexity8/10
Value8/10
Problem
Consultants spent much of their time hand-building and modifying Odoo modules.
Solution
Supported the internal team as they built a secured environment where AI creates and modifies modules, so consultants deliver substantially more in the same time.
Built as
Secured AI development environment
Tools and data
Odoo

Recruitment agency

AI sourcing training

PeopleLink

Complexity3/10
Value6/10
Problem
A 20+ person recruiting team wanted to handle more clients without growing headcount.
Solution
Trained the team to source and process candidates with AI, and to systematize and automate their work.
Built as
Team training

Non-profit

One-on-one AI workshops with the director

Dobrý Anjel

Complexity1/10
Value3/10
Problem
The director of a well-known Slovak non-profit wanted leaner administration and reporting.
Solution
A series of one-on-one workshops with the director on where AI helps in back office and marketing.
Built as
One-on-one workshops
Method

How the scores work

Complexity (1–10)
How hard the solution was to build and how much it takes to keep running: the number of systems it connects to, how many people use it, whether it runs on one computer or in the cloud, and how much maintenance it needs.
Value (1–10)
How much the solution is worth to the people who use it every day: the hours it saves, the work it makes possible, and how much the business depends on it.
Who scores it
I score every project myself once it’s delivered, on the same scale, so projects can be compared with each other. The scores are my assessment, not a client rating.
Clients

Companies I’ve delivered for

Defense tech

SEC Technologies

Ran a full-day agentic engineering hackathon for the R&D and development team to speed up how they ship software, then worked with management on where AI helps internal processes across the company.

sec-technologies.com

IT consulting

26HOUSE

Supported their internal team as they built a solution that uses AI inside a secured environment to create and modify Odoo modules, so consultants deliver substantially more work in the same time.

26house.com

Recruitment agency

Synculario

Mapped what could be automated with the recruiting team, then deployed autonomous recruiting agents that help the team source candidates faster.

syncular.io

Physiotherapy clinic

SportWell

Cleaned up data processing and built an agent that sends the director detailed reports on a schedule. Less retyping between systems and, for the first time, visibility into marketing campaigns.

sportwell.sk

Real estate agency

BOND Reality

Built a custom CRM in which AI scores each inbound inquiry and assigns it to the right agent. More qualified leads without adding manual morning work for anyone.

bondreality.sk

E-commerce

Axolt

Built a quiz that uses AI to segment visitors by the problem they are trying to solve, so every visitor gets a tailored offer on their first visit.

axoltbrain.com

High-frequency trading

Wincent

Built new AI-assisted reporting for the recruiting function, GPTs for the marketing team to produce content, and sourcing automations plus a system for processing hundreds of candidates quickly.

wincent.com

Non-profit

TrustWomen

The director was drowning in tasks. Built three agents for her: one watching administration and calendar, one preparing marketing campaigns, and one finding partners.

trustwomen.eu

Recruitment agency

PeopleLink

Trained a recruiting team of more than twenty people on using AI to source and process candidates. They learned to systematize and automate their work, so they handle more clients at once.

peoplelink.lt

Conference organizer

Fleming

Built a sales AI agent that researches the market on its own, finds target companies and the people likely to care about a given conference, and sends them personalized outreach. The research is strong, though the outreach hasn’t turned into replies yet.

fleming.events

Non-profit

Dobrý Anjel

The director of one of the best-known Slovak non-profits wanted to make administration and reporting more efficient. In a series of one-on-one workshops with the director, we went through where AI helps in back office and marketing.

dobryanjel.sk

Non-profit

Gotická cesta

Built an AI agent that runs on its own in the cloud, finds companies and the people inside them who might support the non-profit financially, and drafts the outreach emails to them.

gotickacesta.sk

E-commerce

Hungry Minds

Built a sourcing agent for the partnerships team that finds YouTube channels worth partnering with, qualifies them against several criteria, and surfaces the most promising ones.

hungryminds.com

Food & beverage

City Gastro

Built eight agents across accounting, controlling, HR and sales that take over manual routine work, so each team gets more done.

citygastro.sk

Restaurant

Hradná

Built a cloud app that gives the manager an overview of every inquiry from prospective guests and answers the simple emails on its own, so she spends far less time on repetitive replies.

hradna.com
In depth

What these systems look like

Real estate

A custom lead-processing system for a real-estate agency

A mini-CRM for an agency with 15 agents that processes inbound inquiries automatically and assigns them to agents in real time.

Recruiting & HR

Internal AI sourcing tool searching 50,000+ candidates

A system that lets managers search an internal candidate database in natural language instead of complex keyword strings.

E-commerce

AI lead scoring through a personalized marketing quiz

A dynamic quiz with AI evaluation that turned anonymous email addresses into richly profiled prospects.

Trading & finance

Sourcing olympiad winners for a trading firm

An AI automation that collected and enriched data on 18,000+ participants of mathematics and physics olympiads.

Recruiting & HR

Management reporting for a recruiting team, straight from the ATS

An AI automation that pulls data from Ashby ATS and prepares accurate reporting for management.

Training & education

An AI orchestrator running operations across 10+ systems

Deployed inside a training company, connecting 10+ systems and running operations semi-autonomously with a human in the loop.

FAQ

Common questions

What kinds of companies does Michal Juhas work with?

Companies from 10 to 500 people across defense tech, high-frequency trading, recruiting, real estate, e-commerce, healthcare and the non-profit sector. Most are based in Slovakia, Czechia and the wider EU. I deliver on-site from Bratislava, Slovakia, or remotely.

What does a typical AI project look like?

Most start with a workshop that maps where AI helps in a specific team, followed by building two or three concrete automations on the company’s own processes rather than demo examples. A follow-up four weeks later makes sure the work is finished and adopted, not abandoned.

How long does an AI implementation take?

A workshop is one day. First deployments go live in the first or second week afterward, with tuning in week four. Larger builds, such as a custom CRM, an internal sourcing tool or an orchestrator across many systems, run over several months as a fractional engagement.

Which AI projects deliver the most value for the least effort?

In my portfolio, the best ratio came from small, focused tools: an Excel macro that runs the VAT check for 15+ restaurants (complexity 1/10, value 10/10), a ChatGPT agent that builds management reports from Pohoda, and scripts in Codex that turn menu drafts into web-ready menus. Bigger builds such as custom CRMs and multi-user web apps pay off when a whole team uses them every day.

Should an AI agent run on my computer or in the cloud?

It depends on who uses it and when. A tool that one person starts when they need it can run on their own computer through ChatGPT, Codex or Excel. Anything a whole team logs into, or that has to run on its own schedule, belongs in the cloud, for example on Google Cloud or as a Claude Managed Agent.

Is an AI agent always the right solution?

No. Where the rules are fixed, deterministic software is cheaper and more reliable: the candidate processor and the EduPage meal payment tool contain no language model at all. I use AI where the work means reading free text, qualifying leads or candidates, or writing drafts.

Which tools and platforms do these projects use?

ChatGPT and custom GPTs, Codex, Claude Managed Agents, Google Cloud, Excel macros, and APIs and data sources such as the Meta API, ContactOut, CoreSignal, Finstate, EduPage, Pohoda, eKasa and Stripe.

How are complexity and value scored?

Both are on a 1–10 scale. Complexity reflects how hard the solution is to build and run: the systems it connects, the number of users, where it runs and how much maintenance it needs. Value reflects what it’s worth to the people who use it. I score every project myself after delivery, so the scores are comparable across projects.

Are these results verifiable?

Every company named on this page is published as a reference with its own website linked. I also hold 4.7/5 across 198 public Trustpilot reviews.

What languages are projects delivered in?

Slovak and English. On-site across Slovakia, Czechia, Poland, Austria, and remote worldwide.

Results like these, in your company

Book a short call. We will look at which of your processes are worth automating first, and whether I am the right person to help.

Book a call