The short answer
To bring artificial intelligence (AI) into your company, you start from a process, not from the technology: you choose a repetitive, measurable activity, check the available data, build a fixed-scope pilot with a point of contact inside the company, measure the result and only then extend it. The first result takes a few weeks, provided you do not try to do everything at once.
- Start from a specific process, with a number to improve.
- Check the data first: AI reasons on what you give it.
- Run a small, fixed-scope pilot, with a point of contact in the company.
- Keep people in the loop: AI proposes, the person responsible decides.
- Prepare the people who will use it: the AI Act requires companies to take measures for the AI literacy of their staff.
Why so many companies do not get started
Almost every company has tried an artificial intelligence (AI) assistant. Few have brought it into a process. In between there are always the same three obstacles: nobody knows where to start, there are fears about the confidentiality of data, and the first “generic” trials produced nothing measurable.
The good news is that the method for starting well is simple, and requires neither large investments nor a dedicated department.
This guide describes the path, one step after another. If you have to decide how to organise adoption, that is, what falls to senior management and what to the departments, you will find the comparison in the guide AI adoption in companies: top-down or bottom-up.
Step one: choose the right process
The first project must be small, frequent and measurable. A good candidate has these features:
- it is repeated many times a week;
- today it involves searching for, copying, checking or re-entering information;
- it is done by skilled people who could be doing something more useful;
- a mistake is not irreparable, because there is a human check in any case;
- it can be measured: minutes per case, errors per month, days of waiting.
Typical examples: filling in customer and supplier records starting from a VAT number, reading orders and order confirmations in PDF, preparing the first draft of a quote, answering recurring questions about procedures and price lists, checking the consistency of contracts and documents, translating content, summarising a customer’s history before a phone call.
To be avoided as a first project: high-risk decisions, processes that nobody has ever described, vague objectives (“let’s use AI for sales”).
Step two: look at the data
A model reasons on what it is given. Before building anything, it is worth asking:
- Where is the information that process needs? In the ERP (the company’s management software), folders, email, Excel sheets, people’s heads?
- Is it accessible? Are there APIs, a readable database, or at least regular exports?
- Is it reliable? Who updates it, and when?
- Who can see it? The existing permissions must apply to AI as well.
This inventory is often the most useful result of the first weeks: it brings to light duplications and gaps that were costing time regardless of artificial intelligence. We cover this on the page data and integrations.
Step three: a fixed-scope pilot
The pilot is a reduced but real version: it works on real data, with real users, on a single process. The rules that make it work:
- A written scope. What it does and what it does not do. Whatever is left out goes on the list for later.
- A point of contact in the company, who knows the process and has a few hours a week.
- A starting number and a target number. Without a measurement there is no project, only a demo.
- People in the loop. AI proposes, prepares, flags. The person responsible confirms.
- A short duration. A few weeks: long enough to see real cases, not long enough to get lost.
At the end of the pilot there is only one question: has the number improved? If it has, you move into production. If it has not, you have spent little and learned where the real problem lies.
Step four: go into production and grow
Moving from the pilot to production means adding what can be left aside in a trial: role-based access, an action log, error handling, cost monitoring, instructions for users, maintenance.
From there you grow one process at a time, reusing what has already been built: the connections to the data, the permissions, the rules. The second project always costs less than the first.
Confidentiality: the three questions to ask any supplier
- Is my data used to train models? With the business APIs of the main providers the contractual answer is no: this must be checked and put in writing.
- Where does the data travel and where does it stay? Archives and documents can remain on the company’s servers; the model receives only what is needed to answer that question. When not even this is acceptable, the model can be run inside the company: the comparison is in the guide AI in the cloud or on-premise.
- Who sees what? AI must respect the same permissions as the users, and every action must leave a record.
People: training and the AI Act
An AI project fails more often because of people than because of technology. Two things help:
- Involve the people who do the work. Those who use the tool every day know where time is lost and recognise a wrong answer straight away.
- Prepare the people who will use it. As well as being common sense, it is required by law: Article 4 of the AI Act, which has applied since 2 February 2025 and was rewritten in July 2026, requires organisations that use artificial intelligence systems to take measures to support the AI literacy of the staff involved. It does not require them to certify each person’s level: it requires them to act, and to be able to show it. A long course is not needed: what is needed is that those who use the tools know what they can get wrong and which data not to enter, and that senior management decides this and records it. The details are in the guide AI Act: what a manufacturing company must do.
The most common mistakes
- Starting from the technology (“we need a chatbot”) instead of from the process.
- Trying to do everything at once.
- Not measuring: without a number before and after, nobody will be able to say whether it worked.
- Taking people out of the loop too soon.
- Neglecting the data, and then blaming the model.
- Buying a generic tool in the hope that it will adapt by itself to your way of working.
How we work
At FT Support we follow exactly this path, which we have described in our method: a first conversation with no obligation, a day with senior management for the process map and priorities, a fixed-scope pilot, the move into production. The first two steps are our AI consulting for companies. We apply it to ourselves first of all: the CRM with AI agents we use every day came about this way.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 4 and Article 50
- Regulation (EU) 2026/1744 of 8 July 2026 (the Digital Omnibus on AI), which rewrote Article 4
- European Commission, AI literacy: questions and answers (Article 4)
Checked on 28 October 2026. Rules and deadlines change: always check the official source before you decide.