The short answer
A CRM is the customer archive: contacts, quotes, contracts, deadlines. With artificial intelligence (AI) inside, it searches, summarises, fills in and checks: it prepares a draft quote, reads a PDF order, flags contracts at risk of cancellation. It does not decide discounts, blocks or renewals on its own: it proposes, and a person confirms. In Italy only 21.7% of companies with at least 10 employees use a CRM, against an EU average of 28.5%.
- AI in a CRM searches, summarises, fills in and checks. Decisions on discounts, blocks and renewals stay with people.
- The tasks that pay off soonest: draft quotes built from the catalogue, orders read from the PDF, contracts at risk flagged from the notes.
- In Italy only 21.7% of companies with at least 10 employees use a CRM (EU 28.5%); in metals, 14.5%. The ERP, the company's management software, is there instead: 49.5%.
- In the most widely used products AI is included only in some editions and agents are paid for by usage: before signing, read the licences, the credits and the terms on data.
- GDPR: the model provider has to be treated as a data processor; solely automated decisions about a person, as a rule, no.
- The Italian Workers' Statute and the AI Act: AI may read the notes to find the contracts at risk, not to assess the salespeople.
Note. The terms of the products mentioned are those published by the providers on 30 September 2026 and they change often; the guide gives no prices and no comparisons in figures. The rules are summarised for the people who run a company: for your own case you need a professional adviser. The national rules described are those of Italy; other EU member states have their own.
What a CRM is, and what changes with AI
A CRM is the program in which a company keeps the information about its customers: who they are, what they have asked for, what they have bought, which contracts they have and when these expire. ISTAT, Italy’s national statistics institute, defines it as follows: software for managing customer information that makes communication easier and helps to keep track of interests and buying habits.
With artificial intelligence (AI) inside, the archive stops being just an archive. A language model, the engine behind assistants such as ChatGPT or Claude, reads the CRM data and uses it to search, summarise, fill in and check. It answers a question written in plain language, prepares a summary page on a customer, writes a draft quote, checks that a contract is consistent, reads an order in PDF, flags the contracts at risk. It does the repetitive work that people do today, starting from the data the company already has.
What AI does not do, in a well-built CRM, is decide: which discount to grant, whether to block a customer, whether to renew. It proposes, explains why, and a person confirms. This is a technical choice, and it is also what the rules require, as we will see further on.
In Italy the CRM is missing, the ERP is not. According to ISTAT, in 2025 CRM software is used by 21.7% of companies with at least 10 employees: 21.1% of small and medium-sized ones, 56.5% of large ones, 19.0% in manufacturing. The European average is 28.5% and, on Eurostat data, Italy ranks seventeenth out of 27. With the ERP, the company’s management software, it is the other way round: 49.5% in Italy against 46.5% in the Union. In basic metals and fabricated metal products the CRM stands at 14.5%, against 22.1% in Europe. In many companies, then, sales work still sits somewhere between email, spreadsheets and the salespeople’s memory: it is one of the parts of the company least served by software, and the one where AI has the most room.
What AI does in sales, task by task
Providers describe AI in the CRM with different words (assistant, copilot, agent), but the tasks are few and concrete. This table says what the AI does and who confirms the result.
| Task | What the AI does | Who confirms |
|---|---|---|
| Customer summary | Reads the customer record, quotes, contracts and notes and prepares a page before a phone call or a renewal | Whoever reads it checks it against the data: it is a reading, not a decision |
| Search in plain language | Turns “customers in Lombardy with a contract expiring within two months” (written in Italian, in our case) into a filter on the archive | The salesperson checks the filter; the system runs it read-only |
| Draft email or quote | Prepares the text from the catalogue, the price lists and the customer’s history | The person who sends it, always |
| Contact priority (lead scoring) | Ranks the contacts to call back according to their history and the signals collected | The sales manager decides whom to call |
| Closing estimate | Gives every open deal a probability of closing | Senior management uses it as an indication, not as a certain figure |
| Check before saving | Flags inconsistencies in customer records, quotes and contracts that a fixed rule would miss | The operator corrects or confirms |
| Document reading | Extracts the number and date from an order in PDF, the company details from the VAT number | The operator checks before applying |
| Renewal signals | Reads expiry dates and notes and brings out announced cancellations, complaints, reminders | The salesperson decides what to propose to the customer |
The general rule: AI works well on structured, measurable tasks, where the result can be checked in a few seconds. This is also what the research says: the measured productivity gains are larger in structured work and smaller in tasks that require deep reasoning (Stanford University’s AI Index 2026).
Quotes and estimates with AI
It is one of the questions most often asked by people who sell: can you prepare quotes with artificial intelligence? Yes, on one condition: the AI must work from the company’s catalogue and price lists, not from its own memory. A model left to itself makes up plausible prices, and a quote with a made-up price does damage.
The right way takes three steps.
- Gathering. The assistant asks for or reads what is needed: the customer, the products or services, the quantities, the terms. If the customer already exists, it takes the data from the customer record and from previous quotes.
- The draft. It puts the quote together with the items from the catalogue, the prices from the current price list and the standard terms. If the catalogue has tiers or formulas (a licence by number of users, a price tied to an index), it applies the formula.
- The check. Before the quote is saved, a check flags what does not add up: a discount beyond the limit for the role of the person preparing it, an item without a price, a payment term different from the one in the current contract, a customer with a duplicate VAT number.
The quote starts as a draft. A person rereads it and sends it, and answers for it to the customer. The time saved lies in the gathering and in putting the quote together. For a manufacturing company with formula-based quotes (metal price plus processing, indexation, currencies) the part that pays off most is the check: a mistake in a quote is discovered later, when it costs money.
Customer contracts, renewals and cancellations
The work does not end with the signature: the contract has to be entered, the customer’s order recorded, the deadlines followed. Here AI does three useful things.
- It reads the documents. From the purchase order in PDF it extracts the number and date and proposes them in the right fields. From the VAT number it fills in the customer record by querying the public registers, and the operator checks.
- The contract is generated from the accepted quote, with the items, instalments and dates already consistent; the AI checks it before confirmation.
- It finds the contracts at risk. It reads the salespeople’s notes and the expiry dates and brings out the signals: a cancellation announced verbally, a complaint, a payment reminder. The result is a list of contracts to follow up, with the reason.
The last point deserves a clarification that returns in the section on the rules: the AI flags the contract, it does not judge the salesperson who wrote the note. These are two different uses of the same data, and the law treats them differently.
Administration (accounts): customer records, invoices, payments received
AI in accounting and finance is among the least widespread uses: according to Eurostat it is used by 1.7% of companies in Italy, against a European average of 4.6%; in administrative processes, by 4.2% against 6.2%. Yet this is where the tasks are most repetitive.
- Customer records that fill themselves in, with the tax checks done by the system and the final check by the operator.
- Checks before issuing: the invoice that does not match the contract, the instalment already collected, the customer with an administrative block. An AI check reads the case and flags it; the hard rules (VAT numbers, VAT groups, limits) stay written in the code, where they can be checked.
- Summaries for decision makers: a customer’s position, the contracts expiring in the next sixty days with their margin, the open reminders. Questions that today take three screens and a spreadsheet.
The principle is the same as in sales: the AI prepares and checks, the person confirms, the system records.
Chatbots and agents that talk to customers
The chatbot is the most talked-about form of AI, and the most misunderstood. For a company that sells to other companies, an assistant that answers customers on its own is rarely the first thing to do: the questions are few and specific, and the relationship matters. An internal assistant pays off more: it prepares the answer with the customer’s data and leaves it for a person to sign.
If, on the other hand, the chatbot answers on its own, for example on the website or about the status of an order, Article 50 of the AI Act applies: anyone interacting with an AI system must know it, unless it is obvious. Saying so in the first sentence is enough. And common sense applies: a chatbot that promises a delivery or a discount can commit the company as if an employee had said it.
A real example: FT Mercati’s CRM with AI agents
At FT Support we use, every day, a CRM built in-house for FT Mercati’s sales cycle: customers, quotes, contracts, invoicing, portal user accounts. It is the case study that many of the examples in this guide come from, and here it serves to show the limits, more than the features.
- A limited number of tools. The agent has nine tools: five read, three write, and one is deliberately blocked and sends the operator to the guided procedure (creating login credentials stays with people).
- Everything starts as a draft. A quote prepared by the AI is a draft; a contract is awaiting confirmation. Before writing, the assistant repeats the data it has gathered and waits for a yes.
- The hard rules are in the code. Duplicate VAT numbers, VAT groups, discount limits by role: decided by written rules.
- Filters are verified. The queries generated by the AI are checked before they run, and can only read.
- Roles apply to the AI as well. Anyone who cannot see a piece of data does not get it by asking the assistant either.
- The log. Conversations, tools used and checks remain on record, and it is possible to reconstruct who proposed and who decided.
The salespeople’s notes are used to bring out the contracts at risk, not to assess the people who write them. We use different models for different tasks: a fast one for conversation, checks and translations, a more capable one for analyses and complex filters.
Standard with AI inside, or custom?
All the most widely used CRMs now have AI included in some edition, and they all do more or less the same things: summaries of the customer and of the deal, draft emails, natural-language search, contact priority, closing estimates, agents that qualify contacts or answer customers. The table shows the terms published by the providers on 30 September 2026: where the AI is and how the agents are paid for. No prices: the price lists are public, change often and should be read on the day you decide.
| Product | Where the AI is | How the agents are paid for |
|---|---|---|
| Zoho CRM | AI agents already in the Standard edition, the cheapest paid edition; the Zia assistant in the higher editions | By usage |
| Microsoft Dynamics 365 Sales | Copilot and agents from the Enterprise and Premium editions | Copilot credits included in Premium, then by usage; the agents require an Azure subscription |
| HubSpot Sales Hub | Assistant and agents in the paid editions; mandatory onboarding for Professional and Enterprise | Credits included in every edition, then by usage |
| Salesforce Sales Cloud | AI in the higher editions; Agentforce for the agents | Credits by usage, plus a per-user licence for access to the agents |
Terms read on 30 September 2026 on the official pages of Microsoft, Salesforce (Agentforce), HubSpot and Zoho (the Microsoft and Salesforce pages are in Italian): they are not sources of the guide, they are the providers’ terms, and they change often.
Three things to read before signing.
- Agents are paid for by usage. The licences cover the assistant; the agents’ actions (qualifying a contact, answering a customer, updating a record) use up credits. Estimate them beforehand, with a one-month trial.
- The data sits with the provider. With a cloud product, customer data is processed in the data centres of the provider and of its model providers: read the contract and the privacy notice, as explained in the guide AI in the cloud or on-premise: how to choose.
- The process has to fit into the product. Formula-based quotes, tiered licences, renewals with rules of their own, data that must stay in the company: when the sales cycle does not fit into a standard product, forcing it costs more than going custom.
When custom pays off. When the ERP is already there and the CRM has to work on top of it (in Italy this is a frequent case: the ERP is in 49.5% of companies, the CRM in 21.7%); when the process is unusual; when the data must stay in the company. In many cases the best choice is not to change CRM: it is to add an AI agent to the CRM or ERP that already exists, if it exposes APIs (a way of connecting from outside) or an accessible database. That is how our CRM came about, and it is the foundation we start from when we build one for another company.
How you pay
Three items, in every case: the per-user licence, which grows with the number of people; the agents’ usage, which grows with the work they do and which all four providers measure in credits, partly included in the licences and then charged for; the connection to the existing systems, which depends on how open they are. For a custom agent the second item is the model’s usage, and the third is the largest part of the project.
A standard product takes little work if the company’s sales cycle resembles it, and a lot if it has to be bent to fit; a custom agent is quoted on the basis of the process. We do not publish prices or comparisons in figures: the assumptions change from company to company. We explain what the cost depends on in the guide What an AI project costs: what it depends on; before the price, it is worth looking at the process map.
Customer data: the rules apply with AI too
Even in a company that sells only to other companies, customer data is to a large extent personal data: the names of contact persons, email addresses, phone numbers, the notes on a phone call, self-employed professionals and sole traders. The GDPR applies in full, and AI adds one step: the data leaves the CRM to reach a model.
- Legal basis. For self-employed professionals and sole traders, quotes, contracts and renewals fall under the contract or pre-contractual measures (Article 6, point (b)); for the contact persons of a company, who are not party to the contract, and for analyses of customers, legitimate interest applies (point (f)), which has to be assessed and put in writing. Consent is needed for promotional communications: in Italy, under the marketing rules of the Garante (Italy’s data protection authority).
- Privacy notice. Articles 13 and 14 require you to say who the data goes to: if an AI provider processes it, that provider is a recipient, and if it processes it outside the European Union, this must be said too.
- The model provider is a data processor. Article 28 requires a contract with sufficient guarantees; the provider cannot pass the data to another provider without written authorisation. Read what the provider does with the data: whether it uses it to train its own models, how long it keeps it, and where.
- Solely automated decisions about a person: as a rule, no. Article 22 gives a natural person the right not to be subject to a decision based solely on automated processing which produces legal effects concerning them or similarly significantly affects them; the exceptions (a decision necessary for the contract, authorised by law or based on explicit consent) still require the possibility of human intervention and of contesting the decision. A credit block decided by AI on a sole trader is exactly this, and it is also a high-risk use under the AI Act. The practical rule holds for all customers: the AI proposes, a person decides, the system records who.
- If the CRM “learns” from the data. The European Data Protection Board clarified in Opinion 28/2024 that a model trained on personal data is not automatically anonymous: it has to be assessed case by case. If the provider trains its own models with the company’s data, the question has to be asked beforehand, in writing.
Three questions to ask any provider: does my customers’ data train your models? Where is it processed and how long does it stay? Who are your model providers, and under what contract?
Salespeople: AI reads the notes, it does not judge people
The CRM records the salespeople’s work: visits, phone calls, notes, times. AI that reads that data touches three sets of rules, and the difference lies in the use. The first two are Italian and are given here as the example of one country: other EU member states have their own rules on monitoring workers.
- The Italian Workers’ Statute, Article 4. Under this law, Italy’s basic statute on workers’ rights, the CRM is a tool used to perform the work: it does not need the trade union agreement required for monitoring tools, as long as it remains a work tool. The information it collects can be used, including for disciplinary purposes, only if the worker has been given adequate information about the tools and the monitoring, in compliance with the privacy rules. An AI that assesses salespeople from their notes turns it into a monitoring tool (paragraph 1): a trade union agreement or the authorisation of the Ispettorato del lavoro (Italy’s labour inspectorate) is needed, in addition to informing the workers.
- Law No. 132/2025 (Italy’s AI law), Article 11. The employer informs workers when it uses fully automated decision-making or monitoring systems that affect recruitment, the management of the employment relationship, the allocation of tasks, surveillance or evaluation. The guidelines of Italy’s Ministry of Labour (December 2025) recommend involving workers and their representatives before introducing AI.
- AI Act. A CRM is not in itself a high-risk AI system. It becomes one if it is used to monitor and evaluate the performance and behaviour of workers or to allocate tasks based on individual behaviour (Annex III, point 4), and also if it evaluates the creditworthiness of customers who are natural persons, sole traders and self-employed professionals (Annex III, point 5). In that case, from 2 December 2027, the deployer’s obligations apply, including informing workers and their representatives, and, for a CRM that a company has had developed and uses under its own name, the heavier obligations of the provider. By contrast, a system that performs a preparatory task or detects deviations without replacing human assessment (Article 6) is not high-risk, unless it profiles people: in that case it is always high-risk. And since February 2025 it has been prohibited to use AI to infer people’s emotions in the workplace, except for medical or safety reasons: Italy’s Garante has already sent, in May 2026, a warning to a start-up over a component that measured employees’ stress from company chats.
The practical line: AI may read the notes to find the contracts at risk, to summarise a customer, to suggest whom to call. It must not produce rankings of salespeople, “engagement” scores or analyses of tone. If the company wants to measure people, it does so with the usual numbers (orders, margin) and with the usual rules. And the people who use these tools must be put in a position to use them: Article 4 of the AI Act requires measures for the AI literacy of staff, as explained in the guide on the AI literacy obligation.
What to expect: the real numbers
On sales in the strict sense we have not found an independent study with a clean number, and the providers’ numbers are the providers’ own. The solid data comes from customer support.
- A 2023 study of 5,179 customer support workers measured, with the introduction of an AI assistant, 14% more requests resolved per hour, with a gain of 34% for the least experienced and a minimal effect for the most experienced (NBER, a US research institute). The AI Index 2026 takes it up: the measured gains are 14-15% in customer support, larger in structured work and smaller in tasks that require deep reasoning.
- In Italy AI is already used above all in sales: among the companies that adopt it, marketing and sales is the most frequent area, with 33.1% (ISTAT 2025). Banca d’Italia (Italy’s central bank), among companies with at least 20 employees, finds AI used first of all in commercial activities. But for 70% of the companies that use it AI has not yet affected productivity, and ISTAT notes that a third of the companies that say they use it do not indicate any purpose.
The reading is simple: the benefit is there, it comes on repetitive, measurable tasks, and it rewards above all those with less experience. It does not come by itself: it comes when the task has been chosen, the result is measured and the process around it changes.
Where to start in a manufacturing SME
You do not need to change CRM to begin. You need three repetitive tasks, a number for each and a person who confirms.
- The customer summary before a renewal or a visit. Measure: minutes of preparation before and after.
- The check on the quote before it is sent. Measure: errors found by the check in a month, and what they would have cost.
- The reading of orders in PDF and the filling in of the fields. Measure: orders recorded per hour.
For each one you decide what the AI can only read, what it can prepare as a draft and what stays with people; you write it into the usage rules; you inform the salespeople. After a month you look at the numbers and choose the next task. This is the path described in the guide AI adoption in companies: top-down or bottom-up, and it needs little: the archive that is already there, a connection and the rules.
The most common mistakes
- Buying the CRM for the AI. If the sales cycle does not fit into the product, AI will not save it.
- Letting the AI decide. Discounts, blocks and renewals decided by the model: a legal problem, and a problem of trust for the salespeople.
- Not reading the contract on data. Where the customers’ data goes, who keeps it, whether it trains models: you find out afterwards.
- Measuring people instead of contracts. Rankings and scores drawn from CRM data are monitoring, with the rules that apply to monitoring.
- Underestimating the credits. How much the agents use becomes clear only when they work for real: run a one-month trial before signing.
- Starting with the chatbot. For those who sell to companies, the internal assistant that prepares and checks pays off more.
How we work
At FT Support we start from what the company already has: the ERP, the CRM if there is one, the archives. In the first conversation we look at the sales cycle and say where AI can bring a measurable benefit and where it cannot; in the day spent on the process map, the second step of our AI consulting for companies, we choose the first tasks together with the people who sell and the people who handle administration; then, with a fixed-scope pilot, we build an agent that works inside the existing systems with the same rules as our CRM: a limited number of tools, drafts, confirmations and a log. The CRM with AI agents we use every day is the foundation, in technology and in experience, that we start from; the AI agents and Custom software services describe how.
Sources
- ISTAT (Italy's national statistics institute), Imprese e ICT, year 2025 (report of 15 December 2025): CRM and ERP software, areas of AI use (in Italian)
- Eurostat, integration of business processes: enterprises using CRM and ERP software (table isoc_eb_iip, updated on 27 February 2026)
- Eurostat, enterprises using CRM software by economic activity (table isoc_eb_iipn2)
- Eurostat, use of artificial intelligence by purpose (table isoc_eb_ai, updated on 15 June 2026)
- Eurostat, Statistics Explained: E-business integration
- Banca d'Italia (Italy's central bank), Annual Report for 2025 (29 May 2026), box on the adoption of artificial intelligence in firms (in Italian)
- NBER, Working Paper 31161, Generative AI at Work (2023): a study of 5,179 customer support workers
- Stanford HAI, AI Index Report 2026, Chapter 4 (Economy)
- Regulation (EU) 2016/679 (GDPR), Articles 4, 6, 13, 14, 22 and 28
- European Data Protection Board, Opinion 28/2024 on AI models and personal data
- Garante per la protezione dei dati personali (Italy's data protection authority), guidelines on promotional activities and combating spam (2013, in Italian)
- Garante per la protezione dei dati personali, warning to a start-up over a plug-in that detects workers' stress (28 May 2026, in Italian)
- Regulation (EU) 2024/1689 (AI Act), Articles 4, 5, 6, 26 and 50 and Annex III
- Regulation (EU) 2026/1744 of 8 July 2026 (the Digital Omnibus on AI): new Article 4 and new dates of application
- Italy's Law No. 300 of 20 May 1970 (the Workers' Statute), Article 4, on Normattiva (in Italian)
- Italy's Law No. 132 of 23 September 2025, Article 11, on Normattiva (in Italian)
- Italy's Ministry of Labour and Social Policies, Ministerial Decree No. 180 of 17 December 2025: guidelines for the implementation of AI in the world of work (in Italian)
Checked on 12 October 2026. Rules and deadlines change: always check the official source before you decide.