Service
AI agents in business processes
We design AI agents that work inside your ERP: they read data and documents, prepare drafts and fill in forms, always with confirmations, roles and a log.
In brief
An AI agent is software that uses a language model to understand a request, choose the right tool and carry out a task: finding a customer, reading a PDF order, preparing a draft quote. FT Support designs them to work inside the systems the company already uses, with permissions, explicit confirmations and a log of every action.
- They work in your systems: CRM, ERP (the company's management software), email, documents, databases.
- They use only the tools you decide on; sensitive actions ask for confirmation.
- Every answer and every action stays in a log that can be consulted.
- You start with a single process, with a measurable objective.
What an AI agent does in practice
There is only one difference from an assistant that just chats: the agent does things. It receives a request in plain language, decides which tools to use from those it has been given, uses them, checks the result and answers. Some examples taken from systems we have built and use every day:
- Sales. “Show me the customers with a contract expiring within 60 days and no open quote.” The agent turns the sentence into a filter on the CRM, runs it and returns the list. The filter can be saved and shared with colleagues.
- Administration (accounts). A purchase order arrives as a PDF: the agent extracts the number and date and proposes them on the contract. A new customer gives their VAT number: the system checks it on VIES, looks up the company details and pre-fills the customer record, which a person checks and confirms.
- Data quality control. Before a quote or a contract is saved, an AI check flags the inconsistencies that a fixed rule would not see: a duration that does not match the dates, an unusual discount, a note that contradicts a field.
- Renewals. The agent reads the notes left by the sales staff and highlights the contracts that need close attention: cancellation requests, reminders, complaints.
How we build them: AI proposes, the rules decide
An agent is useful only if you can trust it. The rules we apply in every project:
- Explicit tools. The agent does not have access “to the system”: it has a closed list of tools, each with a purpose. Most of them are read-only.
- Writing is the exception. The tools that create or change data are few, ask for explicit confirmation and produce draft records, which a person reviews.
- What is sensitive stays with people. Creating credentials and irreversible operations are not delegated: the agent takes the operator to the right screen and stops there.
- The same permissions as the users. The role of the person talking to the agent applies to the agent too: a salesperson cannot get through AI what they could not do by hand.
- Checks in traditional code. Duplicates, VAT numbers, formats, discount limits: everything that can be checked with a rule is checked with a rule, not with a model.
- A complete log. Conversations, tools used and outcomes stay on record: this is needed to correct, to improve and to answer anyone who asks “why did the AI do this?”.
Where they pay off most
AI agents pay off where today there are skilled people who spend hours searching for, copying, checking and re-entering information: sales and purchasing departments, administration, customer support, quality. They do not replace expertise: they clear away the repetitive work that keeps it tied up.
In companies that buy, process and sell raw materials there is an extra advantage: agents can also read FT Mercati prices, indexes and exchange rates, and use them to check a price list, update a quote or verify an indexation clause.
What you need to get started
- A specific process, not “AI in general”.
- A person who knows that process well and can give it a few hours a week.
- Access to the data, even read-only, even just through exports.
- A number to improve: hours saved, errors avoided, days gained.
The rest is our method: a first conversation, the process map, a fixed-scope pilot, then the move into production.
Frequently asked questions
AI agents: the most common questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions. An AI agent, as well as answering, uses tools: it queries a database, reads a document, fills in a form, creates a draft record. Precisely because it acts, it has to be designed with permissions, confirmations and clear limits.
Can an AI agent make mistakes?
Yes, like any colleague. That is why in our projects the actions that change data go through an explicit confirmation, records are created as drafts, formal checks are left to traditional code and every step is logged.
Do we need to change our ERP to use an AI agent?
No. The agent connects to the existing systems through APIs, databases or files. If the ERP exposes nothing, you start from the documents and exports it already produces.
How long does it take to have a first agent up and running?
A fixed-scope pilot project on a single process generally takes a few weeks. The timing depends above all on two things: the availability of the data and a point of contact in the company who knows the process well.
Do you have a process you would like to see working better?
Tell us about it in a first conversation, with no obligation: together we work out whether and where AI brings a measurable benefit, and which step to start with.