Custom software
Can you start from a software project already under way instead of from scratch?
Yes. A software project already under way can be taken over from its goal, code, data and contracts, not from a blank page. The first step is reading the code and documentation with the people who use the software today; that tells you what to continue, what to connect and what to rebuild.
What should you prepare for the first meeting about a software or automation project?
You do not need a specification. It is enough to know which work you want to improve, who does it today and with which systems. Two or three real examples help a lot, even anonymised: a customer request, a quote, the spreadsheet that travels between departments.
How do you work with a custom software vendor, from the first steps to handover?
A custom project goes through four phases: analysis, first version, full development with testing, handover with documentation. Before building, the vendor puts in writing what the software will do, who will use it, what it costs and how long it takes.
How do you test custom software before handover?
Custom software is tested in three ways: automated tests written by the vendor, tests on the connections to other systems, and tests by the people who will use it every day. Tests matter on the difficult cases: missing data, duplicates, failed requests, users without permissions.
Can custom software connect to the CRM, ERP and archives you already use?
Yes. Custom software connects to the CRM, ERP, email and archives, both to read data and to write it. The simplest way is the API the system provides; without one, you use scheduled file exchanges or read the database directly, with more work and more checks.
Is it better to build custom software or adapt an off-the-shelf product?
If an off-the-shelf product covers the work well, buy it. Custom software pays off when you work in a way no product covers, when the subscription grows with users, or when code and data must stay with the company. Over three years it usually pays back above €15,000 a year in subscriptions; below €10,000, keep the subscription.
What budget do you need for a custom software or AI agent project?
With Yempik a first working version (MVP) starts at €8,000 and takes 2-4 weeks; a production system starts at €12,000 and takes 4-8 weeks. Projects that span several departments or critical work get a dedicated quote and are delivered in phases.
AI agents and voice agents
What can an AI agent do in a company?
An AI agent carries out a defined task using authorised data and systems: it reads a request, decides the next step and takes it, with human review at the agreed points. Typical examples: sorting incoming email and drafting replies, entering orders in the ERP, answering the phone, searching company documents and citing the source.
How do you decide which actions an AI agent may take and when it must stop for human review?
Start from a written list: every action the agent could take goes into one of three columns. It does it alone, it prepares it and waits for a yes, or it never does it. The test is the possible harm: an action that can be undone in one click can run alone; one that reaches a customer or moves money goes through a person.
How do you put an AI agent into production, not just a demo?
Today the model matters less than the work around it: talking with the domain expert and redesigning the workflow. You connect data to the right sources, make quality measurable with repeatable tests on answers and the expert’s sign-off, and start from a deliverable that takes weeks today.
How long does it take to have a first working AI agent in a company?
A first agent on a single task usually takes 2-4 weeks: the time for a first working version on the company’s data. An agent in production, connected to the systems and tested by its users, takes 4-8 weeks.
How does a voice agent for a call center work?
A voice agent answers or calls in natural language: it listens, understands the request, replies following the script and logs the outcome in the call center’s systems. It works well on first contact: calling back people who left an online request, collecting details, answering common questions.
Can a voice agent call back leads after an online request and hand the conversation to an operator?
Yes. The website form opens a case, the agent calls within minutes, collects the details and sets the next step. If the customer asks for something out of scope or wants a person, the agent transfers the call or books a callback, and the operator finds the summary in the CRM.
Company brain and company knowledge
What is a company brain?
A company brain is a company’s operating context: decisions, rules, project status and open questions, in one place that people and AI assistants read, with the source next to each item. Unlike a wiki it is updated as work happens, and new information becomes official only when a person approves it. Yempik builds it as a project for each client, who keeps ownership.
Do I need a company brain, or is a tidy folder enough?
Start with the tool you already have, putting decisions and rules in one tidy place. A custom company brain comes later, once you know which information people really look for, who updates it and whether AI assistants need to read it.
Is a company wiki enough to manage knowledge?
A wiki is enough when the context is stable, the team is small and no AI agent needs to read it. It becomes false comfort when nobody updates it: pages age and the real answers go back into people’s heads. Before switching tools, test with the one you have.
How do you keep what your best employee knows inside the company?
With two or three recorded one-hour interviews in which the employee walks through real cases rather than abstract procedures. Rules are written from the interviews, and the expert reviews and corrects them. That way what they know stays in the company when they change role or retire.
Who owns the company knowledge I put into AI?
It depends on the contract with the tool vendor. Two questions decide it: can you take everything back in a readable format, and may the vendor use your content to train models? If the knowledge lives only inside a tool’s memory, you are renting it.
AI training
How long is an AI training programme for executives?
Yempik’s complete format runs 52 hours across three modules: AI fundamentals, applications in business functions, preparing a project. You do not have to take it all: modules can be taken separately, and you can begin with the first, AI literacy.
Do you have to take the whole AI training programme, or can you start with one module?
You can start with one module. The 52 hours are Yempik’s complete format, not a requirement: beginners start with AI literacy, people who already use the tools daily can start with applications in their own function.
What does an executive bring back to the company after AI training?
Three things: which activities to tackle and in what order; which data is needed, who manages it and how to check results; a project with success criteria and a work plan for the following 6-12 months.
What does an executive need to know about artificial intelligence?
Where AI can help in their work, which information they can entrust to it and how to assess an investment proposal. They do not need the technical details: they need criteria for data, tools and checks, and a number to measure results by.
What happens in an AI course for people without a technical background?
People try the tools on cases close to their own work, learn where they go wrong, which data to share and when human review is needed, then each defines a first idea tied to their activity. Exercises use invented cases and sample data.
Can executive AI training be funded through Fondirigenti?
Yes, for Italian companies that belong to Fondirigenti, the training fund for executives, if the plan meets its rules. There are two routes: the Conto Formazione, where member companies submit plans during the year against their available balance, and calls with their own themes, amounts and deadlines.
What does the law require of companies on AI training for employees?
Article 4 of the EU AI Act, in force since 2 February 2025, asks companies that use AI to ensure a sufficient level of AI literacy among the people who use it. It prescribes no course, number of hours or certificate. Training already done counts if you record who, when, on what, with a signature.
AI adoption and results
Which department should you start with when introducing AI?
The department with repetitive, frequent and measurable work, and a manager willing to try: often finance, sales or customer service. Where the work gets stuck matters more than who complains loudest.
Which AI tool should you buy, and should everyone get it?
If you work with Microsoft or Google, start from the assistant included in your subscription, Copilot or Gemini. Buy the paid licence for people who work inside documents all day; the included chat is enough for the rest. Give everyone the company account on the same day, with a date when personal accounts are switched off.
How do you measure whether AI delivers results?
With one number measured before and after on a specific piece of work: hours, errors, response time, sales. ROI is calculated on total cost, including subscriptions, development and people’s time, and against a group that keeps working as before, to know how much is really due to AI.
Which business process should you automate first?
The one that repeats often, follows clear rules and holds up the rest of the work, which is rarely the most annoying one. The four-step method: map the processes, filter by signals (volumes, errors, waiting), find the bottleneck, score it.
Why hasn’t revenue grown after giving ChatGPT to everyone?
Because a tool given to everyone saves scattered minutes, which do not turn into revenue on their own. To see an effect, pick one activity, measure the starting time and leave one case in ten to today’s method for comparison. Then let the numbers decide.
What should you bring to a meeting to decide on AI?
A one-page sheet with seven fields, from the decision to take to the tool, with the five figures needed to present a credible number. Before the meeting, have a colleague ask you the sheet’s seven questions.
Rules, data and liability
Can I upload a company document to ChatGPT?
Before uploading, ask whether it contains customer or employee data, trade secrets or information received under confidentiality. A business plan where data is not used for training lowers the risk, but personal data stays subject to GDPR. When in doubt, remove names and figures first.
Who is liable if AI gets a quote or a customer reply wrong?
Towards the customer, the company is liable, not the tool vendor and not the AI. That is why every price, offer or customer reply goes through a person who checks it against the source before sending.
Which AI rules should you give employees?
A few written rules: which tools to use, which data never to upload, who checks before text reaches a customer, whom to ask when in doubt. Yempik proposes six to send to everyone in five minutes and ten for management to discuss and sign.
How do you manage data and access when using AI at work?
Start with an inventory of the AI tools in use, including the ones nobody authorised. Then decide where each type of data may go, use business plans with a data processing agreement, and give tools only the access they need. A governance plan can be set up in 30 days.
How can a defense SME use AI without data leaving the company?
Start from information classification: what may leave, what stays in the company, what never touches an external tool. For data that cannot leave, use models installed on the company’s own servers. The rules can be written in an afternoon and become software when knowledge and rules sit in one place the company owns.
Vendors and partners
Which questions should you ask an AI vendor before signing?
Twelve questions in four groups: security, how data goes in and out, what changes in the company, what it really costs. A disqualifying answer on security blocks the tool; elsewhere a weak answer becomes a written condition in the contract, with a name and a date.
How do I choose a reliable software house for an AI project?
Look at three things: AI projects in production that you can verify, a clear way of handling data and human review, and a contract that leaves you the code and the data. Ask for a case similar to yours and, if possible, to speak with the people who use it. A demo is not enough.
How does white-label software development work for agencies and software houses?
In white-label development an agency or software house sells the project under its own brand, and a technical partner builds it. The client relationship stays with the agency: the commercial proposal, project communication, decisions on how to present the team.
In white-label development, who manages the client relationship?
The agency. It presents the proposal, answers the client and decides how to communicate the project; the technical partner joins technical meetings only as agreed and does not contact the client on its own.
Is your question missing?
Write it in the contact form: we reply within 24 hours, and if many people ask it, it ends up here.
Ask your question