An AI agent for business procedures is software that uses artificial intelligence to carry out a task, make decisions within defined rules and interact with the company’s tools. It can check documents, compare information and prepare operations in the ERP. A person keeps control of the steps that require a decision.
In this article we describe the agent we built for the accounts payable of an engineering company in Emilia-Romagna, which receives about a hundred invoices a day. We cannot name the client.
What is the difference between an AI agent and automation?
Traditional automation runs a sequence of instructions set in advance. If an invoice arrives in the mailbox, the software can recognise the sender, save the attachment and start a procedure. Every step follows conditions that have already been programmed.
An AI agent can handle less rigid situations. It receives a goal, analyses the information available and decides which tools to use to reach it. It can, for example, read an invoice, look up the related order in the ERP, retrieve the delivery note and compare the data.
Anthropic also describes the distinction in its guide Building effective agents, published on 19 December 2024. The guide separates:
- Workflows: AI models and tools that follow paths defined in code.
- Agents: the model directs its own process and decides how to use the tools available.
The difference matters when designing a solution. A task with steps that are always the same could be handled by traditional automation. A task that requires interpreting different documents, searching several archives and handling exceptions could benefit from an agent.
Anthropic suggests starting from the simplest solution and adding complexity only when needed. The guide goes as far as recommending not to build an agent if a workflow is enough.
It is an important criterion for a business project too: artificial intelligence has to solve a concrete problem, not become one more step to maintain.
Which business procedures suit AI agents?
A procedure is a good candidate when it involves repetitive work, information spread across different documents or software, and checks that take up people’s time today.
Not every administrative or production task has these features. It is worth starting from those where a person has to read, interpret, compare or classify information before going ahead.
Finance and accounting
In accounts payable, for example, you need to check that a supplier invoice matches the order and the goods received. An agent could retrieve the documents, compare them and prepare a proposed posting.
Handling administrative communications can be another useful area. An agent could read a request received by email, recognise its content and prepare the information needed to deal with it.
Purchasing and procurement
An agent could review internal requests, look up information on orders and check that the necessary documents are there. It could also flag situations that need purchasing to step in.
Production and quality
The information needed to handle a non-conformity can sit in reports, job sheets and technical documentation. An agent could gather it, find the relevant references and prepare a sheet to be checked.
Document management and internal knowledge
Operating procedures, manuals, instructions and company documents hold information that often needs a manual search. An AI system can help find the relevant documents and answer questions based on the knowledge available in the company.
The selection criterion is simple: find a procedure where AI can do work that is recognisable and verifiable, leaving to people the decisions that carry responsibility.
How does an AI agent for accounts payable work?
An engineering company in Emilia-Romagna receives about a hundred supplier invoices a day, and each one has to be checked against the purchase order and the delivery note before it is posted.
In Italy electronic invoicing is mandatory for invoices issued from 1 January 2019, under article 1, paragraph 916, of law 205/2017. The e-invoice is issued in XML and sent through the Revenue Agency’s Exchange System (Sistema di Interscambio, SdI), which delivers it to the recipient’s electronic address.
The agent we built steps in after the invoice is received, using data already available in the company’s system. This is how it works.
Receiving the e-invoice
The agent receives the invoice through the flow the company already used for electronic documents.
Its first task is to read the invoice and pick out the information needed for the check: supplier, order references, description of goods or services, quantities and amounts.
Reading the XML gives a structured basis. The AI model interprets the information when the references are not immediately recognisable.
Finding the purchase order
Once the order reference is found, the agent looks up the document in the company’s ERP.
If the reference is incomplete or not directly available, it uses the other information in the invoice to find the matching order. The result is a proposed match to be checked.
Retrieving and reading the delivery note
The agent looks for the delivery note linked to the supply. The delivery note holds the information needed to check that what was invoiced matches what was actually delivered.
The agent reads the document and extracts the data useful for the comparison with the order and the invoice.
Comparing the documents
This is where the actual check begins. The agent compares:
- Invoice and order: checking that references, descriptions, quantities and amounts match.
- Invoice and delivery note: checking that the goods invoiced match those recorded on delivery.
- Order and delivery note: checking consistency between what was ordered and what was delivered.
The result is a check sheet with the matching items and any differences.
For example, if a quantity on the invoice does not match the one on the delivery note, the agent highlights the discrepancy and shows which documents the data comes from. If all the checks are consistent, it prepares the proposed posting.
Checking anomalies
Not all invoices look the same. There are missing references, unclear descriptions, unavailable documents or differences to look into.
The agent recognises these situations and stops. Instead of producing a result that only looks complete, it prepares a report with the information available, stating what needs a human check.
Even when the model is uncertain, the procedure defines a precise behaviour: ask for a check, without inventing missing data.
Preparing the posting
Once the checks are done, the agent prepares the data needed for the accounting entry.
The person in charge sees the invoice, the order, the delivery note, the outcome of the comparisons and any flags. They check the result and approve the proposal.
The posting is never carried out by the agent on its own: a person’s approval is always needed. Only after that approval does the data go into the ERP in the way the company has set up.
How do you design an AI agent to integrate into the ERP?
Design starts from the existing procedure, not from the choice of AI model. To build a solution like the one described, this is the path.
First, study the procedure
You need to understand how the accounts office works today: where invoices come from, where orders are kept, how delivery notes are retrieved and which checks are made before posting.
It is important to look at exceptions too. A procedure can look simple when all the documents are correct, and get more complicated when references are missing or differences appear.
This analysis separates the steps that can be automated from those that have to stay under a person’s responsibility.
Then, define the approval points
The agent has to know when it can go on and when it has to stop.
In accounts payable, the agent completes the reading and comparison of the documents, but posting always needs a person’s approval. Doubtful situations are also sent for a check, with an understandable explanation of the information that raised the doubt.
The data has to stay private
At dotenv, client data only passes through dotenv’s private AI services, never through external AI services.
It matters when working with invoices, orders, delivery notes and commercial information. The architecture has to take the confidentiality of the documents into account.
The agent has to fit into the software already in use
An AI agent should not force a company to duplicate its work on a separate platform.
At dotenv we work to integrate the AI model directly into the software. The model makes the first pass, while a person validates the doubtful cases. The results go into the ERP, the CRM or the platform already in use.
In accounts payable, the accounts office keeps working in its own ERP, and finds the proposed checks and postings in its usual procedure.
Finally, start with a single procedure
There is no need to involve every office and every process at once. A pilot on one procedure shows the quality of document reading, the accuracy of the matches and how anomalies are handled.
The economic value could be worked out with the company’s own data: invoices a month × minutes needed to check one × hourly cost.
At that point you could compare the time spent today with the time the agent-assisted procedure takes, counting the time needed for human checks too.
Who controls an AI agent’s decisions?
An agent’s autonomy needs clear limits. Software that steps into business procedures has to let people understand what it is doing and step in when needed. The principle matters most for high-risk AI systems.
Article 14 of the AI Act states that high-risk systems must be effectively overseen by people while in use. Whoever oversees them must be able, as appropriate and proportionate, to:
- decide not to use the system;
- disregard or correct its output;
- override the decision the system suggests;
- interrupt the system.
These provisions concern high-risk systems. For Annex III systems they apply from 2 December 2027.
The design principle is useful beyond these obligations too: an agent needs checkpoints, ways to step in and defined ways to stop.
In accounts payable, it means the person in charge must be able to consult the original documents, check the comparisons, change the proposal and decide whether to go ahead with the posting.
AI prepares the work. Responsibility for the decision stays inside the organisation.
How can you tell whether an AI agent suits your company?
A good starting point is to look at the procedures that take up time because they require retrieving information, reading documents and comparing data held in different systems.
There is already concrete experience showing how dotenv approaches this kind of project.
For a second-hand luxury goods marketplace we built a listing moderation system: 80% of listings are decided by the software, while 20% go to a person.
For a public administration in Emilia-Romagna we built a meeting minutes transcription system. The audio does not leave the client’s infrastructure and the speech recognition model runs in-house.
They are different projects, but they show two important things: AI can be built into an operational process, and the architecture can be designed around data confidentiality and the role of people.
Neurally, an AI company founded inside dotenv, also comes from this expertise. It governs company knowledge through AI document search and agents that carry out repetitive tasks.
And there is the accounts payable agent described in this article, which checks about a hundred invoices a day for an engineering company in Emilia-Romagna, with the final word always left to a person.
So the question to start from is a concrete one: which procedure in your company takes the most time today to read documents, search for information and check that the data matches?
If you have found a process like that, we can think it through together and work out how to build it into the software you already use. You can read more about dotenv’s work on artificial intelligence on our AI page.
Frequently asked questions about AI agents for business procedures
Can an AI agent work directly in the company ERP?
Yes. It can be integrated into the ERP or other software the company already uses. It analyses the information, prepares results and makes them available in the existing procedure, without necessarily needing a separate platform.
Can an AI agent post an invoice on its own?
In the case described, no. The agent prepares the proposed posting after comparing invoice, order and delivery note. A person always checks and approves before the data is posted.
What happens when the AI agent finds an error?
The procedure can have the agent stop, flag the difference and prepare the information needed for the check. The person in charge can look at the original documents and decide how to proceed.
Are company documents sent to external AI services?
At dotenv, client data only passes through dotenv’s private AI services, never through external AI services. The architecture is designed around the confidentiality of the information handled.
Which procedure is best to start with?
A repetitive task with identifiable steps and verifiable results. Accounts payable is a good place to start, because it requires reading documents and comparing information before a posting. It is best to start with a single process and assess the results.
How much does it cost to build a business AI agent?
It depends on the procedure, the documents to interpret, the software to integrate and the checks needed. To estimate the value, you can start from the time each task takes today and its hourly cost, and compare it with the AI-assisted process.
