AI agents and AI integration
We don’t introduce AI. We solve a problem – sometimes with AI.
A well-built AI agent is not a chat window but a colleague who works from your own data: it extracts, sorts, summarises, drafts replies – and tells you where it got the answer. We start with the process that takes the most time, and only build when it measurably pays off.
Internal assistant EXAMPLE
What discount does Example Ltd. get?
Under the framework agreement 12%, and 15% for orders above HUF 500,000.
Partner records · Example Ltd. Framework agreement 2026-014
And who pays for shipping?
I couldn't find a clear rule for this in the documents. I suggest asking the sales manager.
No source – I will not guess
Six situations where an AI agent gets to work right away
They all have the same thing in common: repetitive text-based work that eats up your colleagues’ time without needing their expertise.
-
„We type in every invoice and delivery note by hand.”
The agent reads the document, fills in the fields and only flags what it is unsure about.
→ Your colleague checks and approves instead of typing.
-
„We get the same twenty questions over and over.”
An internal or customer assistant built on your own documents, with the source next to every answer.
→ Your experts get time for the real questions.
-
„We can’t find anything in eight years of documents.”
Search by meaning: you don’t need the exact word, just describe what you are looking for.
→ Seconds instead of minutes.
-
„We sort incoming e-mails by hand.”
Automatic categorisation, urgency and a suggested reply – the decision stays with your colleague.
→ Your colleague decides instead of reading.
-
„The monthly report takes three days.”
The agent collects the data from your systems and writes a summary that highlights the outliers.
→ Management gets the essentials, not a spreadsheet.
-
„We have the system, but half the team doesn’t use it.”
A built-in assistant that helps inside the interface: where to find things and how to do them.
→ Faster onboarding, fewer mistakes.
What can a well-built agent do?
An agent is not one big “smart” system but a set of well-defined tasks that we connect to your existing systems.
-
Document extraction
Invoices, contracts, delivery notes, forms – text becomes structured, verifiable data.
-
Search by meaning
Across your own knowledge: policies, contracts, e-mails, product descriptions.
-
Draft replies and summaries
For customer e-mails, support tickets and internal questions – approved by a colleague.
-
Triage and routing
Automatic categorisation of e-mails, requests and tickets, routed to the right person.
-
Connected to your systems
The agent reads from and writes back to your existing systems – via APIs, with permissions.
-
Multilingual by design
Translation and multilingual replies in your company’s own terminology.
What it is good for – and what it must not be used for
The deciding question: is it a problem if it is wrong in two cases out of a hundred? If not, AI is a huge gain. If it is, AI is not the answer there.
What it is good for
- Extracting data from documents
- Summarising and condensing
- Classifying and prioritising
- Searching your own knowledge
- Draft replies approved by a person
- Guidance inside your system
What we don’t use it for
- Calculations and bookkeeping – code does that
- Final financial decisions
- Legal opinions
- Unsupervised messages to customers
- Decisions about employees
- Anything that has to be 100% right
With us, AI prepares, suggests and finds – but it does not decide and it does not calculate.
What if it makes a mistake?
We build the agent so that a mistake cannot stay hidden and cannot leave the company unchecked.
-
It works only from your data
It does not answer from the internet but from your documents and systems.
-
A source with every answer
You can see which document or record the answer comes from – one click to verify.
-
It says when it doesn’t know
If there is no source, it does not guess – it tells you who to ask.
-
A person approves what goes out
Anything addressed to a customer or partner goes out only after approval.
-
EU data processing
Under a data processing agreement; the model does not train on your data. Pseudonymisation or a self-hosted model for sensitive data.
-
Measured accuracy before go-live
A pilot on real cases – at the end you get a percentage, not a promise.
Four steps – and you can stop after any of them
-
01
Choosing the process
1–2 hours, free of charge
We look at which process takes the most time, where AI is worth using – and where it isn’t.
-
02
Pilot on real data
2–3 weeks, fixed price
We measure accuracy on real cases. If it isn’t good enough, we stop here.
-
03
Integration and go-live
3–8 weeks
Connected to your existing systems, with permissions and a gradual rollout.
-
04
Monitoring and improvement
ongoing
Cost and quality monitoring, and switching models when a better or cheaper one appears.
Frequently asked questions
Isn’t it enough if the team uses ChatGPT?
A general chat answers from its own knowledge, not from your data, and is not connected to your systems. An agent works from your documents, cites sources, respects permissions, and the work happens inside your existing process, not in a separate window.
How long until it is usable?
For a well-defined process the pilot takes 2–3 weeks and a full rollout typically 1–3 months.
What happens when a better model comes out?
We build the system so the model can be swapped: moving to a newer or cheaper model takes days, not a rebuild.
Where will our data be?
With an EU-based provider under a data processing agreement, or with a self-hosted model for sensitive data. The model does not train on your data.
What does this mean under the EU AI Act?
Most internal agents (extraction, search, summarisation) are low-risk uses. We review the classification and transparency duties for each project.
What if it doesn’t pay off?
Then we stop after the pilot. That is why we always start by measuring: the decision rests on numbers, not enthusiasm.
Let’s start with the process that takes the most time.
In a short conversation we will look at where AI is worth using – and where it isn’t.