When AI genuinely helps and when rules are enough
AI earns its place where data is unstructured: free text, scans, photos, conversation transcripts, customer behaviour. If a process can be described by a handful of conditions such as “if the amount exceeds the limit, send it for approval”, plain workflow automation will be cheaper and more predictable. Models always carry some chance of error, so use them where that cost is acceptable and controllable.
A good AI candidate has three traits:
- a repetitive task performed hundreds or thousands of times a month – answering the same questions, keying in invoice fields, sorting tickets;
- representative data – sample documents, images or conversations on which quality can be measured before launch;
- a process owner who defines a correct result and what should happen when the system is unsure.
If one is missing, we start with a short analysis rather than code – sometimes concluding that the data needs tidying before AI makes sense.
Six types of AI solutions we build
Chatbots for customers and internal teams
An assistant on your website, in WhatsApp or in Teams that answers from an approved knowledge base, cites sources and escalates out-of-scope questions to a person. Details: AI chatbot development.
Document assistant
Search and cited answers across large sets of contracts, procedures or technical documentation, respecting each user's permissions. See the AI document assistant.
OCR and document recognition
Turning scanned invoices, delivery notes, CMR consignment notes and forms into validated data in your ERP or accounting system, with a manual review queue. More: OCR and document recognition.
Computer vision
Image and video analysis for one narrow operational task: inspection on a production line, counting objects, sorting site or damage photos. Service page: computer vision software.
NLP and translation
Classifying and routing messages, extracting data from text, summarising, and translating with glossaries and translation memory. See NLP and translation software.
Recommendation engines
Product, content or offer suggestions based on user behaviour and catalogue data, with business rules and A/B testing. More: recommendation engines.
Matching the solution to the data you already have
Start from the input and the expected output:
- customer or employee questions in conversation form → chatbot;
- hundreds or thousands of files people search through → document assistant;
- scans and photos of documents whose fields must be keyed in → OCR;
- photos or video streams and a “pass / check” decision → computer vision;
- a stream of emails, tickets, reviews or content to translate → NLP;
- a product catalogue plus click or purchase history → recommendations.
Many projects combine two types. A document assistant often needs OCR for scans, and a support chatbot benefits from ticket classification. One specification then covers both modules, and the shared integration layer is built once, usually as an API available to your other systems.
Models, architecture and where data is processed
We are not tied to one model vendor. The choice follows tests on your data, weighing unit cost, latency and data residency requirements. Typical options:
- models via API (e.g. OpenAI, Anthropic, Google, Mistral), including EU cloud regions – the fastest start and strong quality on language tasks;
- open-weight models on your own infrastructure – when documents must not leave the company, or volume is high enough that API fees stop making sense;
- classic machine learning and specialised engines (classifiers, OCR, object detection) – cheaper to run for narrow, high-volume tasks.
The AI layer usually runs as a separate Python service connected to your existing system through a REST API or job queue, whether that system is built on Node.js, .NET or PHP. Search data lives in a vector database or PostgreSQL with pgvector, and logs let you trace where any answer came from.
Quality measured before launch, not in a demo
An impressive demo on five examples says little about behaviour on a thousand real cases. Every project therefore has an evaluation set prepared with a subject-matter expert on your side, plus agreed acceptance thresholds. The measure depends on the solution: correctness and faithfulness to sources for a chatbot, field accuracy for OCR, precision and recall for image detection, NDCG and conversion for recommendations.
The same set later serves as a regression test before any change of model, prompts or content. Low-confidence cases go to a queue where a person makes the call.
GDPR, the EU AI Act and AI security
AI projects touch personal data more often than expected: in email bodies, on scans, in chat logs. The specification states what data reaches the model, where it is processed, how long logs are kept and which data processing agreement is needed with the vendor. Where possible, data is masked before reaching an external API.
The EU AI Act introduces, among other things, a duty to tell users they are interacting with an AI system and specific requirements for high-risk uses such as recruitment or creditworthiness assessment. We flag requirements that may apply but do not give legal advice. On the technical side we test for risks typical of LLM applications listed in the OWASP Top 10 for LLM Applications, including prompt injection and data leakage in responses.
What an AI project costs and what is excluded
An AI module, chatbot, OCR, NLP component or proof of concept is indicatively from $2,900. When requirements are unclear, we start with discovery and a technical specification for $490–$1,500. If the AI module is part of a new application, the reference point is a web application MVP from $4,900. Proposals come in three options – Essential, Recommended and Scale – in USD, EUR or GBP.
Excluded unless the proposal says otherwise: model and API usage (billed by the vendor per token, page or request), cloud GPU capacity, commercial engine licences and third-party data labelling. We estimate monthly running costs before work starts. Maintenance and monitoring start from $249 per month. Full terms are on the software pricing page.
How we start working together
MaViAl is a software company based in Wrocław, working with clients in Poland, the EU, the UK, the USA and Canada. We communicate in English, Polish, Russian and Ukrainian. Every project follows the same stages: discovery, specification, design, iterative build, testing and acceptance, launch and maintenance. The commercial model – fixed scope after discovery, time and materials, or dedicated capacity – depends on how much uncertainty the project carries.
To begin, describe the task AI should take over, the rough volume and the systems it needs to connect to. Send us your brief and we will reply with clarifying questions and a proposed next step. Other services are listed in the software development section.
AI Software for Business: From Idea to a Working Process
AI Chatbots That Answer From Your Company's Knowledge
Details →AI Document Assistant for Company Documents
Details →Computer Vision: Automated Image and Video Analysis in Your Process
Details →NLP and Translation Software for Multilingual Text
Details →OCR and Document Recognition: From Scan to Data in Your System
Details →Recommendation Engines for Products and Content
Details →Frequently asked questions
Where do we start if we don't know which process suits AI?
With a short discovery phase. We review two or three candidate processes, their volume, available data and cost of error, then recommend one for a pilot. It costs $490–$1,500 and ends with a written specification.
Isn't ChatGPT enough instead of a custom solution?
For individual work it often is. A custom solution makes sense when AI must work on data from your systems, with access control, logging, CRM or ERP integration and quality measured against a test set.
Will our data be used to train the vendor's models?
That depends on the vendor and plan; many business API offerings do not train on customer data by default, but we verify it for the specific service and region. If that is not enough, the model can run on your own infrastructure.
How long does an AI implementation take?
A proof of concept with limited scope usually fits into a few weeks; a full production rollout ranges from weeks to a few months. Most of the time typically goes into preparing data and the test set rather than writing code.
Who is responsible when the AI gets something wrong?
The rules are set in the specification: confidence thresholds, situations where a person decides, and operations the system may not perform on its own. In our projects AI supports staff rather than replacing decisions with legal, financial or employment consequences.
Can you take over an AI project started by another vendor?
Yes, after reviewing the code, prompts, data, vendor configuration and deployment. Often the first step is building the missing test set.