What can be automated, and on what evidence.

01No data: possible solution

Preparing quotes

Problem

For every quote someone gathers price lists, similar offers and the customer’s terms, then rewrites everything by hand. The reply arrives late and offers are not consistent with each other.

Data needed

Past quotes, price lists, customer records, standard commercial terms.

Tools

n8n or Make · a language model for the text · the ERP for customer records and documents.

What stays with people

Prices and terms come from the price lists, not from the model. A person reviews and signs before anything is sent.

What is known

There are no serious measurements on quotes. A related figure: in an experiment on professional writing tasks (not quotes), people using AI took on average 40% less time.

Noy and Zhang, Science 381, 2023 · experiment with 453 college-educated professionals, USA, writing tasks

02Adoption in companies

CRM and follow-up of leads and offers

Problem

Requests and offers sent go unanswered because nobody remembers to follow up. Customer information is scattered across email, spreadsheets and phone calls.

Data needed

Website forms, sales inbox, open offers with date and status, customer records.

Tools

n8n or Make · a CRM with APIs · a language model to classify and draft.

What stays with people

Follow-ups are limited in number and can be stopped. Contact with the customer stays with you.

What is known

21.1% of Italian SMEs use CRM software, against 56.5% of large companies. This is an adoption figure, not an effect.

Istat, Enterprises and ICT 2025 · SMEs with 10–249 employees and large companies with 250+ employees, Italy

03Value measured in studies

Incoming email and requests

Problem

Inboxes like info@ receive everything: customers, suppliers, certified email, advertising. Someone reads it all, sorts it and answers the recurring questions.

Data needed

Shared inbox, categories and recipients, standard replies already in use.

Tools

n8n or Make with Gmail or Microsoft 365 · a language model to classify and extract data.

What stays with people

Replies stay as drafts: at the start of the project nothing is sent on its own.

What is known

In a study of over 6,000 workers at large companies, people using the assistant spent half an hour less per week reading email and completed documents 12% faster. Vendor study: a commercial survey.

Dillon, Jaffe, Peng and Cambon, Microsoft Research MSR-TR-2025-18 · over 6,000 workers at 56 large companies · commercial survey (vendor study)

04Value measured in studies

Customer support on FAQs and documents

Problem

The same questions (opening hours, order status, documents to bring) keep staff busy, and the answer changes depending on who replies.

Data needed

FAQs, manuals, terms of service, history of emails or tickets.

Tools

n8n with document search · a language model · website, WhatsApp Business or email.

What stays with people

The most robust model is an assistant that suggests replies to the agent. Whoever writes in always knows they are talking to an AI system.

What is known

Across 5,172 customer support agents, an assistant suggesting replies increased issues resolved per hour by 14% on average and by 34% for the least experienced.

Brynjolfsson, Li and Raymond, “Generative AI at Work”, QJE 140(2), 2025 · 5,172 customer support agents, US software company

05No data: possible solution

Appointments and reminders

Problem

Booking, moving and reminding appointments by phone keeps the front desk busy. No-shows, or people arriving without documents, waste slots in the calendar.

Data needed

Calendars, appointment types and durations, contacts, documents to bring.

Tools

Calendar and email in n8n or Make · reminders by email, SMS or WhatsApp.

What stays with people

AI is usually not needed here: it is a classic automation. We say so upfront.

What is known

No measurement found in institutional or academic sources.

06No data: possible solution

New customers: documents and forms

Problem

To open a new customer account, someone collects documents, company records and forms, then chases what is missing and enters it by hand in the ERP.

Data needed

Checklists by customer type, forms, shared folders, documents in use.

Tools

n8n or Make · document reading with human review · the ERP (integration to be checked).

What stays with people

The automation collects and checks completeness. Assessments, including those required by law, stay with a person.

What is known

No measurement found in institutional or academic sources.

07Value measured in studies

Supplier invoices

Problem

Supplier invoices arrive as XML through Italy’s e-invoicing system, but they still have to be checked, matched to orders and delivery notes, coded and booked. Discrepancies become non-conformities.

Data needed

XML invoices from the e-invoicing system (SdI), orders, delivery notes, supplier records, coding rules.

Tools

n8n or Make to read the XML · reading of unstructured delivery notes · APIs of the ERP or the e-invoicing channel.

What stays with people

A person confirms the accounting entry.

What is known

In an automated order-to-payment scenario, savings are estimated at 58 minutes and €22 per cycle, and 86% less time spent handling invoices. This is a scenario estimate, and it grows with volume.

Digital B2b Observatory, Politecnico di Milano, 2024 · Italian companies · scenario estimate, not a measured result

08Value measured in studies

Bank reconciliation

Problem

Matching bank transactions to invoices is often done by hand or in a spreadsheet: it is slow, and with incomplete payment references it becomes a puzzle.

Data needed

Bank statements (CBI or CSV), payment schedule, issued and received invoices, IBANs.

Tools

n8n or Make with matching rules · a language model only for ambiguous payment references.

What stays with people

Matches are automatic only when they are certain. Doubtful cases go to a review queue.

What is known

Indirect evidence, on the manual method: in field inspections 94% of operational spreadsheets contained errors; in laboratory tests people get on average 3.9% of cells wrong.

Panko, EuSpRIG 2015 · 85 operational spreadsheets inspected; 14 laboratory studies

09Adoption in companies

Payment reminders

Problem

Overdue invoices get chased late or erratically, because it is manual work nobody likes.

Data needed

Reconciled payment schedule, contacts, graduated reminder templates.

Tools

n8n or Make for scheduled reminders by email or certified email (PEC) · approved texts.

What stays with people

We only start from a reconciled payment schedule: a reminder to someone who has already paid damages the relationship.

What is known

Among companies that use AI, only 10.2% use it in accounting, controlling or finance.

Istat, Enterprises and ICT 2025 · companies with 10+ employees, Italy

10Value measured in studies

Periodic reports and dashboards

Problem

The monthly report is put together by copying data from the ERP, the bank and various spreadsheets. It arrives late, and every copy can introduce an error.

Data needed

ERP exports or APIs, bank statements, written definitions of the KPIs.

Tools

n8n or Make to consolidate the data · a BI tool · a language model only for the commentary.

What stays with people

The numbers are calculated by verifiable formulas, not by the model.

What is known

Companies that analyse their own data went from 26.6% to 42.7% between 2023 and 2025; 16.0% use business intelligence software. On the manual method, the spreadsheet figure in use case 8 applies.

Istat, Enterprises and ICT 2025 · companies with 10+ employees, Italy

11No data: possible solution

Checking documents before delivery

Problem

Before a document or contract goes out, someone checks attachments, signatures, dates and cross-references. Omissions slip through easily.

Data needed

Checklists by document type, reference templates, documents in the folder.

Tools

n8n or Make · a long-context language model that flags anomalies, quoting the passage.

What stays with people

The check flags, it does not certify. Responsibility stays with whoever signs.

What is known

No data on document checking. A useful warning: in an experiment with 758 consultants, on tasks outside AI’s reach correct answers dropped by 19 points.

Dell’Acqua et al., Harvard Business School Working Paper 24-013, 2023 · experiment with 758 BCG consultants and GPT-4

12Value measured in studies

Recurring documents

Problem

Circulars, reports, letters, minutes: similar texts rewritten from scratch every time.

Data needed

Templates and texts already approved, case data, style guidelines.

Tools

A language model via API, connected to the structured data with n8n or Make.

What stays with people

Legal references, dates and amounts are checked by a person: the model can invent plausible references.

What is known

In an experiment with 453 professionals, on writing tasks AI cut average time by 40% and raised quality by 18%.

Noy and Zhang, Science 381, 2023 · experiment with 453 college-educated professionals, USA, writing tasks

13No data: possible solution

Finding and comparing suppliers

Problem

To compare offers, someone collects quotes in different formats and copies prices, lead times and terms into a table.

Data needed

Offers received, supplier records, purchase history, evaluation criteria.

Tools

n8n or Make to collect the offers · a language model that puts them in a table with a reference to the page.

What stays with people

Information found online is checked against official sources. The choice stays with you.

What is known

No measurement found in institutional or academic sources.

14Adoption in companies

Searching internal archives

Problem

Procedures, past cases, technical sheets and contracts sit in folders nobody can navigate any more. Work already done gets redone.

Data needed

Shared folders with defined permissions, metadata where available.

Tools

n8n with semantic search over documents · a language model that answers citing the source.

What stays with people

The system respects existing access permissions. An answer without a cited source does not count.

What is known

Extracting information from text documents is the most widespread AI technology: 70.8% of companies that adopt AI use it.

Istat, Enterprises and ICT 2025 · companies with 10+ employees, Italy

15Adoption in companies

Meeting transcription and summaries

Problem

Minutes of meetings, interviews and site visits are written from memory, or not at all. Decisions and tasks get lost.

Data needed

Recordings with the participants’ consent, list of participants, minutes template.

Tools

Transcription via API · a language model for summary, decisions and tasks · n8n or Make for filing.

What stays with people

Recording only with consent. The summary is reviewed by someone who was there.

What is known

41.3% of companies that use AI use speech recognition.

Istat, Enterprises and ICT 2025 · companies with 10+ employees, Italy