shardana.ai handles AI development for companies, startups and SMEs: it designs and builds custom AI software that enters real processes instead of staying a demo. The studio is based in Cabras (Oristano), in Sardinia, and works with an international team, on site and remotely.
Many companies have already tried artificial intelligence: a subscription to a generic assistant, a few prompts, a demo that convinced everyone in a meeting. Then the project stalls. The assistant does not know the internal procedures, cannot read the management system, answers confidently even when it is wrong, and nobody can say what it did yesterday. Our work starts from here.
What does it mean to develop artificial intelligence for a company?
Developing artificial intelligence for a company means building a system that uses language or predictive models inside a specific process, with that company's data and rules. It is not about switching on a generic assistant and hoping it understands the context. The work includes choosing the sources the AI can read, the actions it can take and the points where it must stop and ask a person. It also covers everything around the model: APIs, databases, authentication, interfaces, testing, monitoring and maintenance. The result can be a chatbot on your website, an internal assistant that answers questions about procedures, an agent that routes incoming requests or an AI component added to an existing application. In every case it is custom software, and it is judged on the process it improves, not on the quality of a single answer in a demo.
When is an artificial intelligence project worth it?
An AI project is worth it when there is recurring work that depends on written information, on text to interpret or on handovers between different tools, and when the cost of that work is clear to the people who do it. This happens when the same questions reach support every day, when company knowledge is scattered across PDFs, emails and shared folders, or when requests and documents must be read and sorted by hand before anyone can work on them. It is also worth it when you already have a prototype that works in a demo but cannot handle real users, permissions and volumes. If instead the problem can be solved with explicit, predictable rules, we will tell you: a deterministic automation often costs less, is easier to maintain and does not need a language model. Using AI is a project decision, not a starting point, and we assess it together with you.
Which business processes can artificial intelligence improve?
Artificial intelligence is most useful where people spend a lot of time searching for, reading, summarizing or sorting information. Every department has typical cases: sales needs quick answers on products and terms, customer care receives repetitive questions, and administration works on documents with fields to extract and check. Management and human resources also have tasks where AI can lighten the daily workload. Not every case deserves the same investment. To choose them we look at three elements: how often the activity repeats, how much a mistake costs and how reliable the starting data is. The table collects typical examples of application by department, useful as a basis for the first discussion. They are not promised results or client cases: every process must be analyzed before deciding whether AI is the right choice and what level of autonomy it can have.
| Department | Typical examples of AI use |
|---|---|
| Sales | Answers on products, price lists and terms; draft quotes; qualification of incoming contacts |
| Customer care | Chatbot for customer support on the website and in the customer area, request status, handover to a person |
| Administration | Data extraction from invoices, orders and documents; consistency checks before posting |
| Operations and back office | Routing of emails and tickets; updates of management systems and CRMs; report preparation |
| Human resources | Internal assistant on regulations and procedures; support for onboarding new hires |
| Management | Search and summaries across internal documents; forecasts and scoring with machine learning models |
Which artificial intelligence solutions do we develop?
We develop different AI solutions depending on the problem, and often several components work together in the same project. A business chatbot answers customers or colleagues starting from approved content. An AI agent goes one step further: it takes actions in your systems, such as opening a ticket or updating a record, within defined boundaries. An MCP server is the layer that connects agents and assistants to company data and tools in a controlled way. Search across company knowledge, based on RAG, finds answers in internal documents and can cite the source. AI can then enter the applications you already use, to summarize, classify or extract data, or sit inside automations that connect several tools. Below you will find the main solutions, each with a detail page when available, so you can start from the one closest to your problem.
Business chatbots
Chatbots for customer support, websites and internal use, which answer starting from approved content: catalogues, regulations, procedures, FAQs. The chatbot follows your escalation rules, states when it does not have enough information and hands the conversation to a person. → Business chatbots for customer support
MCP servers for AI agents
The Model Context Protocol connects AI agents and assistants to company tools, APIs and data through a dedicated layer. We design MCP servers with small, verifiable tools, defined permissions and a log of every call. → MCP server development
AI agents
Agents that read a request, gather the information they need and prepare or carry out an action in your systems. Every action has defined permissions, is logged and, when it touches sensitive data or decisions, waits for approval.
Search across company documents (RAG)
Retrieval-augmented generation systems that index internal documents and databases, retrieve the relevant passages and build answers that can cite the source. They are the basis of many internal, onboarding and technical support assistants.
AI inside the applications you already use
We integrate language models into web platforms, portals and custom software: summaries, data extraction from documents, classification, draft generation.
AI process automations
When the goal is to automate a flow across several tools, the work shifts to automations: CRMs, management systems, email and documents connected in processes with clear rules and approval steps. → AI automations for businesses
Machine learning and deep learning: when are they really needed?
Machine learning and deep learning are needed when the problem is not answering a question, but reading data to predict, classify or recognize something. A machine learning model can estimate the probability that a customer leaves a service, assign a score to sales leads or forecast demand for a product from historical data. Deep learning comes into play with more complex data, such as images, audio or unstructured text, where neural networks and larger amounts of data are required. In both cases quality depends on the available data more than on the algorithm: before proposing a model we check how much data there is, how clean it is and how it will be updated over time. A simple, well-monitored model is often more useful than a complex architecture, and it is also easier to explain to the people who will use its results every day.
How do we make an AI system controllable and safe?
An AI system in a company must be useful, but also explainable to the people accountable for it. That is why we start from permissions: the assistant reads only authorized sources and takes only the planned actions, and if a user cannot see a document they must not be able to get it by asking the AI. Sensitive decisions stay with people, with thresholds, approvals and escalation rules wherever automation touches personal data, payments, contracts or customer relationships. Every system is observable: we record questions, answers, tools called and unresolved cases, so you can understand what happened and improve over time. We also assess with you which data can be sent to an external model and which must be filtered or excluded. Finally, we do not promise total autonomy or guaranteed savings: together we define what to measure and we verify it after release.
How is an AI project developed, step by step?
An AI project moves in short, verifiable steps, so that hypotheses are tested before the system grows. It starts with a discussion of the process and the tools in use, with no need for a technical solution to be defined already. An analysis follows that maps activities, exceptions, data and responsibilities, and sets a first scope with shared success criteria. Then we build a prototype on a real, limited case, to be tried with real users. If it works, we develop it and integrate it with company sources and tools through incremental releases, code reviews and CI/CD pipelines. Finally the system goes into production and is monitored: unresolved cases and new questions show where to improve content, rules and integrations. The table summarizes the phases and indicative durations, which we confirm for each project after the analysis.
| Phase | What happens | Indicative duration |
|---|---|---|
| First discussion | Current process, goals, tools in use | One meeting |
| Analysis and scope | Map of activities, data, risks and success criteria | About one week |
| Prototype | First working version on a real case with a graphical interface. | 2-3 weeks |
| Development and integration | Connection to sources and tools, and optimization. | 2-4 weeks |
| Release and improvement | Go-live, monitoring, updates | Ongoing |
How much does it cost to develop an artificial intelligence solution?
The cost of an AI project depends on the agreed scope, which is why we do not publish a price list that fits everyone. Five factors weigh the most: the number and quality of the sources to connect, the integrations with existing software, the actions the AI can take, the security and privacy requirements and the usage volumes. On top of these come the recurring costs of models and infrastructure, which grow with use and should be estimated from the start to avoid surprises. A prototype on a limited case is often the fastest way to a realistic estimate of the full project, because it shows how much work data and integrations really require. After the first discussion you receive a written proposal with phases, deliverables and costs, and you can start with a limited first step before committing to the whole project.
| Factor | Simpler project | More complex project |
|---|---|---|
| Data sources | A few documents that are already in order | Many sources and formats, data to clean |
| Integrations | None or one documented API | Management systems, CRMs and ERPs without ready-made APIs |
| AI actions | Answers only | Actions in the systems with approvals |
| Security and privacy | Public content | Personal or confidential data, permissions by role |
| Volumes | Internal use by one team | External users and traffic peaks |
What do you receive at the end of the project?
At the end of each phase you receive something you can verify, not just a presentation. The analysis produces a document with the project scope, the risks identified and the agreed success criteria. The prototype is a working version on a real case, which your team can try and comment on. The final system is integrated with your sources and tools, with permissions defined by role and by action. You also receive logs and monitoring tools to see questions, answers, actions and errors, together with technical and operational documentation written both for the people who use the system and for those who will maintain it. After release, the real cases collected become an improvement plan with clear priorities. Code ownership, access and support arrangements are defined in the proposal, before we start.
Which technologies and what experience do we work with?
shardana.ai was founded in 2026, but its work rests on more than 25 years of experience of the founder, Maurizio Brioschi, in software engineering, backend development, systems architecture and leading technical teams. That experience predates the studio, and we keep it distinct from the age of the brand. In AI projects this foundation matters more than it seems, because the language model is only one part of the system: the rest is APIs, databases, authentication, work queues, testing, deployment and maintenance. The reference stack includes PHP, Laravel, MySQL, Node.js, TypeScript, React and Vue.js, with API-based architectures. On the AI side the studio works on language model integrations, RAG, agentic workflows and MCP servers. We do not publish client cases we cannot document: in the first discussion we go into the technical choices for your case.
“If you think good architecture is expensive, try bad architecture! That is what Brian Foote and Joseph Yoder used to say. It is all a matter of software design and architecture: a bug that blocks an application for two days over the weekend, amid a thousand curses, is not caused by a programmer who got a line of code wrong, but by a software engineer who did not properly join the use cases with the edge cases or underestimated a problem.” — Maurizio Brioschi, founder of shardana.ai
Where does shardana.ai operate?
shardana.ai works with companies, startups and SMEs across Italy, and much of the work is done remotely: analysis, development, releases and reviews take place online, with in-person meetings when needed. The office is in Cabras, in the province of Oristano, in the Sinis area, and Sardinia is the territory where the studio is rooted and where it follows local businesses, institutions and startups with particular attention. Being in Sardinia changes neither the method nor the tools: an AI project mainly requires access to data, people willing to try it and clear decisions, and all of this works well at a distance too. For projects in Sardinia, and in particular in the province of Oristano, in-person meetings can be arranged. Alongside the studio, shardana.ai develops DOMOS Network, a technology network project in Sardinia currently in the validation phase.
Frequently asked questions
- Where do we start if we have never used AI in the company?
- From a concrete, limited process: the most frequent support requests, search across internal documents, email routing. A small, measurable first case is more useful than a project that tries to change everything at once.
- Do we need to have our data and documents in order already?
- No. You can start from existing documents, FAQs, tickets and procedures. During the analysis we work out what needs tidying up and we improve it as real user questions emerge.
- Is our data used to train the models?
- It depends on the model provider and the type of contract. During the project we assess with you the available options, where data is processed and which information can be sent to the model. Sensitive data can be excluded or filtered before sending.
- What is the difference between a chatbot and an AI agent?
- A chatbot answers questions starting from defined content. An agent can also take actions: open a ticket, update a record, query an API. For this reason an agent requires stricter permissions, logs and approval points.
- What is an MCP server and when is it needed?
- It is a component that exposes company tools and data to an AI agent in a controlled way, through the Model Context Protocol. It is needed when the AI must work with real systems and you want a single layer to manage permissions, logs and maintenance.
- Can AI make mistakes?
- Yes. That is why we design the system to work on approved sources, to state when it does not have enough information and to hand uncertain or sensitive cases to a person. Logs help identify errors and correct content and rules.
- Can we integrate AI into the software we already use?
- Often yes, if the software exposes APIs, database access or usable exports. We check this in the analysis phase, before committing to timelines and costs.
Related services
- Business chatbots: customer support and internal assistants based on your content.
- MCP servers: the layer that connects AI agents to company data and tools.
- AI automations for businesses: processes across CRMs, management systems, email and documents.
- MVP development for startups and SMEs: from the first idea to a working product, with AI built in if needed.
- About us: the studio, the method and the founder.
Tell us about the process to improve
Describe how it works today, what you would like to achieve and which tools you use. We will reply with a first assessment and, if it makes sense, with a proposal for a first concrete case.