Custom AI software development for companies

What custom AI software do we build?

Software built around one process of yours, where an off-the-shelf tool does not fit: your documents, your systems, your rules. These are the builds we are asked for most:

If the work is a fixed chain of steps between systems, it is AI process automation. If the software must decide which tool to call and then act in your systems, it is a custom AI agent. Many projects combine the three.

AI assistants and chatbots on your company's data

An AI assistant on company data, often called a RAG chatbot, answers your staff's or your customers' questions from your own documents, procedures and knowledge base, and shows the passage it used. It is the custom AI build we are asked for most, and the one where a generic chatbot fails most visibly: it answers fluently, without a source, and sometimes wrongly.

An assistant is ready for production when it has:

• Answers only from your sources. Every answer cites the document and passage; when nothing relevant is found, it says so instead of improvising.
• Permissions. Each user sees only the documents they are allowed to see.
• Fresh content. Documents are re-indexed when they change, so the assistant does not answer from last year's procedure.
• An evaluation set. Real questions with the correct answers marked by your people, re-run on every change of model, prompt or documents.
• The channel your people use. Web chat, Microsoft Teams, Slack, WhatsApp or inside your application, with hand-off to a person.
• Data where you decide. Public cloud, your own cloud region or on-premise.

We build assistants like this for the public too: ai-aflat.ro answers questions about Romanian law from 220,000+ legislative acts, updated daily, and links the official source with every answer. Earlier, we built one of the first ChatGPT-based customer-support systems in production, for a startup serving retail brands.

The step-by-step version is in how to build a chatbot that knows your company data. If the assistant must also check an order status or open a ticket, it becomes an AI agent.

How does a custom AI software project start?

1. The contact form, a short call, a custom offer. Tell us about the process in the contact form; after a short call we send you an offer that fits. Every offer is sized to the project's complexity.
2. A Tech Audit when the scope is unclear. If you are not sure what to build first, we start with the Tech Audit: four hours on site with the people who do the work, then a report within three working days with 3–5 opportunities ranked and what not to build.
3. Scoping on your data. One use case, a written specification, the first evaluation cases and a test on a sample of your documents or systems. You leave with a fixed price for the build.
4. Fixed-price build, usually 6–12 weeks. Integrations, evaluation set, interface and monitoring. The software runs alongside your team before it takes over the work.
5. Production and handover. Code, documentation, runbook and evaluation set in your repository. For the next use cases, an AI engineer can work inside your team.

What comes with every build?

• An evaluation set from your real cases, re-run on every change, so you know whether a new version is better or only different.
• Monitoring in production, with an alert when quality drops.
• A person approving wherever a mistake costs money or reaches a customer.
• An EU AI Act classification of the use case and GDPR notes on the data flow. General information, not legal advice.
• No lock-in. The code, documentation and runbook are yours; your team, we or anyone else can maintain them.

Which custom AI software have we built?

• ai-aflat.ro, built and run by Sapio: a free AI assistant over Romanian legislation, used by more than 15,000 people. How we built it.
• A Romanian speech-to-text model for a voice-services startup, trained from about 5,000 hours of recordings. It had the best word error rate on the market at delivery, and we handed over the full training pipeline. The case study.
• A music genre classifier for a record label, trained on about 2 million tracks across 150 genres and sub-genres, with 95% accuracy on the test set, delivered with a containerised pipeline for retraining. The case study.
• Sapio was founded in 2021 by Vlad Tudor (Forbes Romania 30 Under 30, 2025), who leads every engagement personally. More on the projects page.

What drives the cost of custom AI software?

We do not publish prices: an assistant over a hundred procedures and a speech model trained from scratch are different projects. The cost depends on:

• The scope of the use case, and how many user groups and edge cases it has to cover.
• The state of your data. Clean and structured, or scanned, versioned and contradictory.
• The systems to integrate, and whether they have APIs.
• The accuracy you need, and how bad a wrong answer is.
• Whether an API model is enough or a model has to be trained on your data.
• Where it runs and how much it is used. Cloud, your region or on-premise; usage sets the monthly running cost.