shardana.ai designs and develops AI automation for businesses and SMEs: workflows that connect the software you already use and hand to artificial intelligence the reading, classifying and preparing that a person does today. The studio is based in Cabras, in the province of Oristano, and works with companies across Italy, mostly remotely.
In almost every company there is someone who spends hours moving data from one place to another. They read an email, open the management software, copy a code, update a spreadsheet, notify a colleague. The work is not difficult, but it never stops, and the errors arrive precisely when the volume grows. Automation removes this friction and leaves the decisions to people.
What is AI automation?
An AI automation is a digital workflow that carries out a sequence of steps across several tools on its own and uses an artificial intelligence model wherever content has to be understood. Traditional automations follow fixed rules: when a form arrives, create a contact in the CRM. They work well as long as the data is tidy. AI comes in when the input is free text, a PDF, an email written differently every time: it reads, recognises what it is about, extracts the useful data and prepares the next step. An AI agent goes a step further and decides for itself which actions to take to reach a goal, within the permissions you give it. In practice the best automations combine all three: rules where they are enough, AI where interpretation is needed, people where decisions are made.
| Aspect | Rule-based automation | AI automation | AI agent |
|---|---|---|---|
| Typical input | Structured data, forms | Email, PDFs, free text | Goals and open requests |
| How it decides | Rules written in advance | The model interprets, the rules guide | The model chooses the actions, within defined permissions |
| Predictability | High | Medium, to be measured | Lower, to be monitored |
| When to use it | Steps that are always the same | Variable content | Tasks with several steps and tools |
Which processes should you automate first?
The first process to automate is not the most important one: it is the one where the balance between effort and risk is most favourable. We look for activities that repeat often, every day or every week, that start from digital inputs and that today require manual steps between two or more tools. It matters that the rule can be explained to a newly hired colleague: if nobody can say how the decision is made, the process has to be clarified first and automated afterwards. It also matters that an error is visible and can be corrected before it causes damage. That is why, at the start, we prefer automations that prepare the work, such as a draft, a classification or an extracted piece of data, over those that act on their own towards customers or suppliers. Once the workflow has proved that it works, the scope is widened.
Signs that a process is not ready yet: there are more exceptions than normal cases, the data is scattered on paper, or everyone does it their own way.
Examples of AI automation by department
The examples below are recurring patterns, not client cases. For each one we show where the AI works and where the decision stays with a person.
| Department | Example of automation | What the AI does | What a person decides |
|---|---|---|---|
| Administration | Supplier invoices received by email or certified email | Extracts supplier, amounts and due dates; matches the order | Approves the booking when the data does not match |
| Sales | Quote requests from the website and email | Classifies the request, fills in the CRM record, prepares a draft | Sends the offer and sets the price |
| Customer support | Sorting incoming requests | Recognises topic and urgency, proposes an answer with sources | Handles sensitive cases and complaints |
| Operations | Orders arriving in different formats | Reads the order and turns it into a line in the management software | Checks non-standard orders |
| Tourism and hospitality | Availability requests in several languages | Understands dates, guests and language; prepares the reply | Confirms the booking |
| Management | Weekly report from several sources | Gathers the numbers and writes a summary | Reads, comments, decides |
For customer support there is a dedicated page on business chatbots for customer support. For the method we use on a single workflow, from mapping to production, read the page on business process automation with AI.
n8n, Make, Zapier or custom development?
We do not start from the tool but from the process. Platforms such as Zapier and Make are cloud services, quick to set up and suited to simple workflows with moderate volumes; their cost grows with the number of operations. n8n can also be installed on the company's own servers, which helps when data must not leave the company perimeter or when volumes make a pay-per-use service expensive. Custom development makes sense when the logic is complex, precise security controls are needed, volumes are high or the workflow has to become part of an application. Often the solution is mixed: a platform to orchestrate the steps and custom components for the critical parts. In the first conversation we explain why we propose a given route and how much it will cost to maintain over time.
How do automations connect to CRMs, management software and email?
The preferred channel is the APIs of the software you already use: CRMs, management and ERP systems, e-commerce platforms, document management systems, email and certified email inboxes. Where the software supports them we use webhooks, which notify the automation when something happens without having to check continuously. When a system has no API, we evaluate alternatives that are more stable than automating the user interface: scheduled exports, read-only access to the database, file exchange through a shared folder. If the automation has to let an AI agent work across several tools, the connection often goes through an MCP server, which states explicitly what the agent can read and what it can do.
How do AI automations stay under control?
An automation nobody monitors is a risk pushed further into the future. That is why every workflow we build follows four principles.
- A person approves actions that cannot be undone. Payments, messages to customers, changes to master data: the automation prepares, a person confirms.
- The model says when it is not sure. Below an agreed confidence threshold, the case goes to a review queue instead of continuing.
- Every run leaves a trace. What arrived, what the model decided, who approved, what was written and where. If a piece of data is wrong, you can trace it back to the step that produced it.
- Least privilege. The automation only accesses the data and functions it needs, with dedicated, revocable credentials.
On top of this come the handling of personal data under the GDPR and, when a customer interacts with an automated system, the transparency required by the AI Act.
What are the risks and limits of AI automation?
AI automations work well, but they are not infallible, and it is better to know that beforehand. A language model can produce a plausible but wrong piece of data, for example an amount misread from a low-quality scan: that is why critical data is compared with other sources or goes through a review. Automations break when something changes upstream, such as a supplier's invoice format or an updated API, so they need monitoring. Incoming content, such as emails, may contain instructions written to trick a model: a well-designed automation treats that text as data to read, never as commands to execute. Then there are the usage costs of the models, which grow with volume, and the dependency on service providers. Finally, the most common limit: automating a confused process makes it confused faster.
How do you measure the return on an automation?
The return is measured by comparing the situation before and after, so the first thing to do is to measure the process as it is today. Three numbers are needed: how many cases per month, how many minutes each takes, how many end with an error or a correction. An example calculation: if booking an invoice takes 6 minutes on average and 400 arrive each month, the process takes up 40 hours a month. If the automation leaves a person only the check of doubtful cases, for example one in ten, the hours drop accordingly. From the time saved you subtract the recurring costs: platform, AI models, maintenance. The calculation should be redone after the first two months of operation with real data, not only with the initial estimates.
How we work, from the process to the automation in production
- Process analysis. We map steps, exceptions, data and responsibilities, and measure the starting situation.
- Prototype on a real case. We build the workflow on a subset of real cases and measure its accuracy.
- Integration with your systems. We connect CRM, management software, email and documents, with dedicated permissions.
- Shadow start. For a period the automation runs in parallel with the people, who compare and correct.
- Operation and maintenance. The workflow goes into production with monitoring, logging and periodic reviews.
How much does an AI automation cost?
The cost depends on the agreed scope, which is why we do not publish a one-size-fits-all price list. The main factors are the number of systems to connect and the quality of their APIs, the variety of documents and cases to handle, the share of the work that requires artificial intelligence, the security requirements and the monthly volume of cases. On top of development come the recurring costs: automation platform, AI models and maintenance. We estimate them before starting, together with the expected return, because an automation only makes sense if it costs less than the work it removes. After the first conversation you receive a written proposal with phases, deliverables and costs, and you can start with a single process.
What happens after go-live?
An automation is not delivered and forgotten. Suppliers change formats, software gets updated, AI models are replaced by new versions. Maintenance includes monitoring errors and cases that ended up in review, updating integrations when APIs change, checking accuracy when the model changes and a periodic review with you to decide what to improve or extend to other processes.
Where do we work?
We work with companies and SMEs across Italy, mostly remotely: analysis, development and reviews work well online. The office is in Cabras, in the province of Oristano, and for Sardinian companies in-person meetings can be arranged, especially useful during the process analysis phase.
Frequently asked questions
- Which process should we start from?
- From a repetitive process, with digital inputs and rules you could explain to a newly hired colleague. It is better to start with a workflow that prepares the work, such as extracting data or sorting requests, than with one that acts on its own towards customers or suppliers.
- Is artificial intelligence always needed?
- No. Where the data is structured and the rules are fixed, a traditional automation is cheaper and more predictable. AI is needed when free text, documents or emails have to be read, requests classified or replies prepared.
- Do automations replace people?
- They remove repetitive work and leave people the checks, the exceptions and the decisions. In our experience the bottleneck shifts: less time copying data, more time for the cases that need attention.
- What happens if the automation makes a mistake?
- Every run is logged, so the error can be found and corrected. Actions that cannot be undone always go through an approval, and uncertain cases end up in a review queue instead of continuing.
- Is our data safe?
- The automation only accesses the data it needs, with dedicated credentials. When needed, we use tools installed on the company's servers and model providers with contractual guarantees on data use. We discuss it in the first conversation, based on the kind of data you handle.
- Do you work with the management software we already use?
- In most cases yes, if the software offers APIs or allows exports. Before proposing a project we check how it can be connected, and if the only route is fragile we tell you.
Related services
- Business process automation with AI: the method for automating a single workflow.
- Business chatbots for customer support: answers based on your content.
- AI development for businesses: AI agents, RAG and custom AI software.
- MVP development for startups and SMEs: when the automation becomes a product.
- About us: the studio, the method and the founder.
Tell us about the process you want to lighten
Describe how it works today, how often it repeats and which tools it involves. We will reply with a first assessment and, if it makes sense, with a proposal to automate a first process.