AI agents · Documents · Knowledge base · ERP/CRM/TMS
We design and ship AI agents, AI processing of invoices, orders and CMR documents, and RAG assistants over your internal documents – all connected through proper APIs to the systems you already run. Models stay in the EU or on your own servers, pilots run on your real data, and results are measured in hours saved and error rates, not gut feel.
Overview
AI process automation means putting large language models (LLMs) and AI agents to work on the routine admin inside your organisation: reading emails and documents, re-keying data between systems, routing requests and answering repeat questions. Rule-based workflow automation and RPA follow fixed scripts – the bot clicks the same button every time and breaks the moment a supplier changes its invoice layout. AI copes with unstructured input such as PDFs, scans, free-text emails and contracts, and handles judgement steps like deciding where a request belongs or whether a document matches its order.
AI does not replace your systems or your people's accountability. We build it as a layer between your inbox, your documents and your ERP, CRM or TMS: the model proposes an output, code validates it against master data, and a person decides whenever confidence is low or the step is sensitive (human-in-the-loop). Every step lands in an audit log, so you can trace what the system did and why. Where a plain rule does the job, we use the rule – it is cheaper and more predictable.
A ChatGPT or Microsoft Copilot licence is often enough for drafting, summarising and searching everyday documents – and if that is your situation, we will tell you. Custom integration pays off when AI has to work inside your systems: posting documents to the ERP, reading orders from the TMS, respecting access rights or processing a high volume of documents every day. We integrate through proper APIs rather than fragile no-code connectors, you own the code, and the model can be swapped without rewriting the application.
We measure value from the pilot onwards: hours saved, error rate before and after, processing time per document or request, and throughput – how many cases the team handles without hiring. That is why we start with a single process where the benefit can be proven quickly and within a limited scope.
What we deliver
Our AI workflow automation services ship the way all our software does – integrated with your systems, tested, monitored and owned by you. Start with one building block and add the others once the value shows up in the numbers.
Agents that complete multi-step tasks across systems: triaging incoming email, drafting replies, creating CRM records, moving order data into the ERP or kicking off approval workflows. Sensitive steps wait for human sign-off, and every action is written to an audit log.
Invoices, purchase orders, contracts and CMR/transport documents: AI reads PDFs, scans and photos, extracts header and line-item data (supplier, VAT ID, amounts, loading address), classifies the document and validates it against ERP master data. Confidence thresholds route uncertain documents to a person; the rest flow straight through.
An assistant that answers from your policies, manuals, contracts and knowledge base using retrieval-augmented generation (RAG): it finds the relevant passages first and only then writes an answer, citing the source. It respects access rights, so sales never sees payroll policies. Suitable for customer support and the internal helpdesk.
We wire the language model directly into the systems you already use – through their APIs, database or an integration layer we add. The AI works with live data on customers, orders and stock and writes results back, instead of people pasting text into a chat window. Built on experience from 100+ production projects.
We choose the model by data sensitivity: cloud APIs in EU regions, or an on-premise LLM deployment with open-weight models such as Llama or Mistral. We pick services and settings under which your data is not used for training, and treat GDPR compliance for every AI agent as an architecture question: data minimisation, retention, access control, audit logs, human oversight and documentation.
An AI system is a new attack surface. We defend against prompt injection (malicious instructions hidden in an email or document), give agents least-privilege permissions and can put the finished solution through a penetration test. After launch we track output quality, model costs and error rates, and tune prompts and models based on real usage.
Use cases
Typical scenarios in which AI takes over the routine and people handle the exceptions. The actual benefit is always proven in a pilot on your own data, not estimated from a slide.
Finance and accounts payable
Invoices arrive by email as PDFs, scans and in every layout imaginable. AI extracts header and line items, checks the supplier and VAT ID, matches the invoice to the purchase order or goods receipt in the ERP and prepares it for posting. The AP team only deals with mismatches and documents where the model is unsure.
Sales order entry
Customers send orders as attachments, spreadsheets or free text in the body of an email. An AI agent identifies the customer, maps line items to your product codes, checks prices and availability and creates a draft order in the ERP or CRM. Sales staff review and confirm it, so the customer gets a confirmation sooner.
Transport and logistics
Transport orders from customers and partners come in many formats, and CMRs often arrive as scans with stamps and handwriting. AI extracts loading and unloading points, dates, goods and reference numbers, creates the shipment in the TMS and, after delivery, attaches the signed CMR for invoicing. We know this domain from building our own TMS platform.
Customer service
An AI chatbot on your website or in email support answers from current terms, product guides and the order status it pulls from your system. It clearly identifies itself as AI, hands complaints or frustrated customers over to a person with the full context, and drafts replies for support agents.
Internal helpdesk and HR
New colleagues no longer have to ask five people how travel expenses are approved or where the current price list lives. The assistant answers from policies, wikis, SharePoint or shared drives, links to the exact document and shows only what the user is authorised to see.
Procurement and legal
AI pulls key terms from contracts – parties, term, notice periods, penalties and payment terms – into a contract register with reminders before expiry. For new contracts it flags deviations from your standard terms. This is not legal review: the final call stays with your lawyer.
How we work
We don't start with a big programme but with one process and a measurable result. During delivery you get weekly demos, daily stand-ups, an open backlog and full repository access.
Tell us which processes slow you down. We will say straight away where AI makes sense, where a ChatGPT or Copilot licence is enough and where conventional automation will do. We reply within 24 hours and sign an NDA on request.
We walk through the process with the people who do the work and review sample documents and the APIs of your systems. Together we pick the first use case with high volume and a clear metric, measure the baseline and propose the architecture, including model hosting.
We build a working pilot on your real documents and emails, anonymised where needed. Output quality is measured against a test set, confidence thresholds are tuned and we agree when a person takes over. At the end you know how much work the solution actually takes on.
We take the pilot to production: ERP, CRM or TMS integration, permission management, audit logs, monitoring, backups and fallbacks for when a model or API is unavailable. Before go-live the solution goes through a security review, including prompt-injection testing.
We track accuracy, model costs, processing time and the share of cases routed to people. We refine prompts, add document types and further processes, and stay with you for the long run – bug fixes, model updates and new requirements.
Why Megu
Every engineer has 10+ years of production experience – no juniors, no anonymous subcontractors. With 100+ production projects behind us, we build AI as software that still has to work three years from now: tests, logs and monitoring included.
We won't recommend AI where a rule, a script or a licence you already have does the job. We choose proven tools, measure value in hours, error rates and processing time, and tell you openly when a project isn't worth it.
You get full repository access, and code ownership stays with you. The architecture lets you swap the model – from a cloud API to an EU-hosted or on-premise model – without rewriting the whole application.
We build our own transport management platform, so CMRs, transport orders and dispatch are familiar ground. Based in Košice, we work with companies in Slovakia, Czechia, Austria, Germany and across the EU, under EU jurisdiction, in English, German or Slovak.
Technology
We pick the technology per project based on data sensitivity, processing volume and the systems you already run. We are not tied to any single model provider.
FAQ
AI process automation is the use of large language models and AI agents for work that used to need a person: reading documents and emails, moving data between systems, routing requests and answering questions from internal documents. Unlike a chat window in the browser, the AI is connected to your ERP, CRM or TMS and writes results back into them. Sensitive steps are approved by a person and every action is recorded in an audit log.
RPA follows fixed rules, while AI can handle unstructured input and simple judgement calls. A rule such as ‘when file X arrives, move it to folder Y’ is reliable, but it fails on an invoice in a new layout, a scanned CMR or an email written in someone's own words. AI can read and interpret that content, yet it can also be wrong, so we validate its output against system data and route uncertain cases to a person. In practice the two approaches are often combined.
Typical AI workflow automation examples are processing supplier invoices, turning emailed orders into ERP sales orders, extracting data from CMRs and transport orders, and answering customer or employee questions from internal documents. Others include triaging a shared inbox, drafting replies for support agents and pulling key dates and terms from contracts. What they share is high volume, repetitive steps and messy input that fixed rules struggle with. We pick the first one together with you during process mapping.
Cost and timeline depend mainly on the number of connected systems, the number of document types, how the model is hosted and processing volume, so we give a concrete estimate only after mapping the process. A pilot on real data typically takes a few weeks; the production rollout depends on the integrations involved. On top of development, plan for running costs – usage fees for a model API or hardware for an on-premise deployment. The first call, where we clarify the scope, is free.
ROI is measured by comparing a baseline taken before the pilot with the same metrics afterwards: hours of manual work, error rate, processing time per case and throughput. We record the baseline during process mapping, so the comparison rests on your own numbers rather than industry averages. On the cost side we count development, model usage or hosting, and the time people still spend reviewing exceptions. That gives you a solid basis for deciding whether to extend the solution to further processes.
ChatGPT is neither compliant nor non-compliant in itself – it depends on which offering you use and how you use it. Business and API offerings from major providers typically come with a data processing agreement and terms under which your data is not used for training, and some offer EU data residency; consumer accounts are a different matter. Your organisation remains the controller, so you still need a lawful basis, retention rules, access control and, for higher-risk processing, a DPIA. Always check the provider's current terms; the legal assessment stays with your data protection officer or lawyer.
No – running a model on your own servers removes the risk of transferring data to a third party, but it does not make the processing compliant by itself. You still need a lawful basis, data minimisation, defined retention, access control, a way to honour data-subject rights and, where the risk is higher, a data protection impact assessment (DPIA). We design the technical side to support these requirements; the legal assessment stays with your data protection officer or lawyer.
The EU AI Act can create obligations both for organisations that develop AI systems and for those that use them, and what applies depends on what a system is used for and how. The regulation sets different rules for different categories of use, and its obligations are phased in. Whether and how it applies to your specific system, including any deadlines, is a question for your legal counsel; we do not provide legal advice. What we do is design solutions with the AI Act in mind: human oversight, transparency towards users, logging, access control and technical documentation that makes the legal assessment easier.
Contact
Describe it in a few sentences, ideally with a sample document stripped of sensitive data. We reply within 24 hours, the first call is free and no-strings, and we sign an NDA on request.
