The short answer
An AI agent can work in the ERP, the company's management software, on three conditions: the ERP can be queried via APIs, webhooks or stable screens; the data is well kept; a person confirms the proposals. It connects via API, webhook, MCP or RPA and works at three levels: it reads, proposes drafts, and writes only where the administrator allows it. Microsoft does so in Business Central: a quote before the order, a draft invoice before posting.
- Three levels, three risks: reading is freely allowed, proposing a draft is the norm, writing is an exception that the administrator grants object by object.
- Four ways to connect: the ERP's APIs, webhooks, MCP (the open standard created in 2024 and adopted by the major providers in 2026) and, as a last resort, RPA on the screens.
- This is how ERP providers do it: Business Central always creates a quote before the order, a new supplier is created blocked, SAP requires explicit approval for high-risk processes.
- Idempotency, minimum permissions, a log: if the agent retries, the order must not be created twice, and it must be clear who confirmed what.
- E-invoicing in Italy: the agent prepares; transmission to the SdI, Italy's e-invoice exchange system, goes through the company's credentials and delegated authorisations, with confirmation by a person.
- In Italy an ERP is already in place in one company in two and in three manufacturing companies in five; workflow automation with AI in fewer than three in a hundred.
Note. Informational content updated as of 13 October 2026; the providers' pages were reread on 17 November 2026. The features offered by ERP providers change every few weeks and are reported here only as declared on their official pages, without passing judgement; for e-invoicing, accounting entries and personal data, consult your own advisers.
Can I put AI to work in my ERP?
Yes, an artificial intelligence (AI) agent can work in the ERP, the company’s management software, on three conditions. It must be possible to query the ERP: through application programming interfaces (APIs), through automatic notifications (webhooks) or, in the worst case, through stable screens. The data must be well kept: clean customer and supplier records, consistent item codes, dated price lists. And a person must confirm what the agent proposes, at least until the numbers say that this can be relaxed.
The ERP is almost always there: in Italy 49.5% of companies with at least ten employees use ERP software, 59.9% in manufacturing and 59% in basic metals and metal products, in line with the rest of Europe (ISTAT, Italy’s national statistics institute, December 2025; Eurostat, 2025 data). Artificial intelligence inside that ERP is not: workflow automation with AI is found in 2.9% of companies in Italy, half the European average, and in metals in just over one in a hundred. Almost everyone who reads this guide is starting from scratch, and the guide starts from what can be done, not from how to optimise.
What it is not: a generic assistant “connected” to the ERP with no defined tools. An AI agent works with a closed list of tools, each with its own permissions, as explained in the guide AI agents: what they are and what they can do.
Reading, proposing, writing: three levels, three risks
- Reading. Summaries, answers to questions about the data, checks: “which of this month’s orders have a price different from the price list?” Low risk: nothing changes in the ERP. It is the level you start from, and it is already useful.
- Proposing. The agent prepares a draft (an order, a purchase invoice, a posting, a payment reminder) and a person reviews and confirms it: the draft action. It is the normal level of work.
- Writing. The agent saves or posts on its own, within tight limits and for selected objects. It is the exception, and the administrator has to grant it case by case.
This is not an extra precaution: it is the scale that ERP providers have written into their products. Microsoft documents that the Business Central MCP server starts out read-only, and that the administrator enables creation, modification, deletion and actions on selected objects, with the user’s identity and a trace of who did what. Its agent for sales orders always creates a quote first, a document that does not touch planning, availability or cash flows, and has every message to the customer approved by a person. Zucchetti, an Italian provider, describes its offering for manufacturing in four levels: natural language, assist, enrich, automate. Sistemi, another Italian provider, proposes in its product for accountancy practices the account, the subledger and the VAT code for each line of a purchase invoice, and these remain a proposal.
The four ways to connect
| Way | How it works | When it is the right one | Where it breaks |
|---|---|---|---|
| The ERP’s APIs | The external program requests and writes records through functions that the provider publishes; the agent does not touch the database | When the ERP offers them and documents them | If the provider changes the interfaces; if a field you need is not exposed |
| Webhooks | It is the ERP that gives notice when something changes: order confirmed, invoice posted, customer or supplier record changed | For agents that react to an event instead of querying all the time | Subscriptions that expire and have to be renewed; notifications lost if nobody checks |
| MCP | The open standard through which AI applications discover and use a system’s tools and data; in 2026 it is offered or accepted by Microsoft, SAP, Salesforce and UiPath | When the ERP has an MCP server or when you use a general-purpose assistant on company data | If whoever builds the application does not put in confirmations and permissions: the specification asks for the user’s consent before every tool, but cannot enforce it on its own |
| RPA on the screens | A software robot repeats a person’s actions on the screens | When the ERP has neither APIs nor webhooks and changing it is not an option | The screen changes, the robot breaks; it does not understand new cases |
There is a fifth bridge, for when the assistant is the one in the productivity suite and the ERP is a different product: connectors. Microsoft distinguishes two types: those that copy and index the data into the assistant’s system and the “federated” ones, which query it in real time without copying it, with the user’s identity and permissions; since October 2026 federated connectors, where the connector provides the tools, can also create, update or delete data “with the user’s approval”, and administrators can audit the activity. The difference between copying and not copying matters for confidentiality and for how up to date the data is, and it has to be explained to the people who decide.
The ERP providers’ agents, today
These are the features declared on the official pages as at the date of this guide; they change every few weeks and should be read again before you choose.
- Microsoft Dynamics 365 Business Central, Microsoft’s ERP for SMEs: Copilot and the agents are available only in the online version, not in the one installed on the company’s premises; Copilot inherits the user’s permissions and Microsoft declares that it does not use the data to train the models without explicit permission; for customers inside the EU Data Boundary the data does not leave that boundary. The Sales Order Agent reads customers’ requests, finds the customer from the email address, searches for the items first by exact code, then by attributes and description, and always creates a quote; the Payables Agent, validated and supported only in English as at the date of this guide, reads suppliers’ invoices from the mailbox using OCR, creates unposted purchase invoices for review and, if the supplier is new, creates it when the supervisor asks and creates it blocked: the page says that the agent does not offer supplier approval features. The MCP server and the REST APIs are documented; webhooks give notice when an entity changes. For large companies, Dynamics 365 Finance has agent management, an agent for account reconciliation and an MCP server, all declared to be in preview as at the date of this guide.
- SAP: Joule Agents carry out multi-step tasks on SAP and non-SAP systems using the MCP and A2A protocols; each agent has a verified identity and acts only within the user’s permissions, high-risk processes require explicit human approval, accounting controls block non-compliant postings and logs track every action. For SAP Business One, the product for small companies, the pages we read do not declare any agents: the route is integration.
- Oracle: NetSuite exposes REST services for reading, creating, modifying and deleting records and for queries; for its ERP for large companies Oracle announces, in an undated document, dozens of agents and speaks of operations “with minimal to no human intervention”. This is a different position from the one declared by Microsoft and SAP: who confirms, and when, is something to ask the provider product by product.
- Salesforce Agentforce, for the CRM: agents that work within guardrails and hand out-of-scope cases over to a person, connectors and verified MCP servers, a trust layer that masks personal data before sending it to the model and does not keep prompts and responses with the model provider. The CRM is covered in another guide: CRM with AI: what it does and what it does not.
- The Italian ERP products widely used by SMEs in Italy: TeamSystem presents its ERP as “enhanced by artificial intelligence”, with use cases in accounting, bank reconciliation and company knowledge; Sistemi declares, for its product for accountancy practices, the automatic proposal of account, subledger and VAT code on purchase invoices and an assistant on the documentation, while for the ESOLVER ERP the page mentions the support of Sistemi AI in administration and finance, without describing its features; Passepartout describes AI in its ERP for companies, from accounting automation to data analysis; Zucchetti, for manufacturing, describes document processing, predictive maintenance and quality, and planning. None of the pages we read states publicly how permissions, confirmations and logs are handled: it is the first question to ask.
RPA, automation and agents: what changes
RPA is automation carried out by software robots that repeat a person’s actions on the screens: suited to repetitive, rule-based tasks, even on old programs, and fragile when a screen changes. Agents read, understand and choose among tools. For ISTAT, automation without artificial intelligence components is not AI, and UiPath, one of the main RPA providers, now presents robots as the layer that executes what an agent has decided.
The rule is simple: rules where a rule is enough, an agent where understanding is needed. Anthropic, which builds models and agents, recommends looking for “the simplest solution possible”, even no agent at all, and increasing complexity only when needed: a workflow that a rule can handle on its own does not need a model.
The controls you do not skip
- A closed list of tools. The model decides when to call a tool and returns a structured call; it is the company’s program that executes it, not the model: that is where confirmations, permissions and the log sit.
- Minimum permissions and the user’s identity. The agent acts with the permissions of the person it is working for, never with an account that sees everything; least privilege is a security principle defined in the NIST glossary, and ERP providers apply it to the agent.
- Draft action. The quote before the order, the unposted invoice before posting, the new supplier blocked until someone verifies it.
- Human confirmation for sensitive actions. Tools declare whether they require approval; execution stops until a person decides, and if the arguments cannot be checked it fails safe (human in the loop). Excessive deference to what the system proposes, which NIST calls automation bias, is the risk run by anyone who confirms without reading: that is why confirming is a job, and you measure it.
- Hard rules in code, out of the model’s reach: an item code not found means the draft stays on hold, never “the closest match”; no automatic blocking of a customer; limits on amounts and quantities.
- Idempotency. If the agent retries after a network error, the order must not be created twice: a unique key for every request and a check in the ERP.
- Action log. Who proposed, who confirmed, what changed, with source and date; it is there to reconstruct an error and to measure, not to monitor people.
- Tests in an isolated environment and stopping conditions, including a maximum number of attempts, before touching real data.
- Untrusted instructions. An uploaded document or an external tool may contain instructions for the agent: the MCP specification says to treat tool descriptions as untrusted, unless they come from a trusted server, and OWASP devotes a list of the top ten risks to agents.
In the CRM with AI agents that we use every day at FT Mercati the assistant has nine tools, five of them read-only; actions that change data ask for confirmation and every step is in the log.
What stays with the company, and with people
- E-invoicing. In Italy, to take one country’s example, the invoice is a file in the format set by the revenue agency (Agenzia delle Entrate), with the technical specifications updated to version 1.9.1 since May 2026; issuing it means transmitting it to the Sistema di Interscambio (SdI, Italy’s e-invoice exchange system) and receiving the outcome. An agent can prepare the data and the draft; transmission goes through the ERP or the agency’s services, with the company’s credentials or delegated authorisations, after a person has confirmed. Legally compliant preservation of electronic documents is a service with a person responsible for it, as Italy’s Digital Administration Code (Codice dell’amministrazione digitale) provides: the agent does not replace it.
- Accounting entries and responsibilities. Responsibility for postings, tax returns and payments stays with the company and its directors. The agent proposes; in the products read for this guide posting goes through a person (Microsoft, SAP), and where a provider promises operations without human intervention it is the company that has to put the confirmation in place.
- Customer and supplier data. Customer and supplier records, contact persons and bank details go to the model only if the task needs them, with a provider engaged as data processor and a data processing agreement; masking personal data before sending it is a practice that one provider declares and that can be replicated in custom software.
- Training. People who work with the agent must be trained on its limits: Article 4 of the AI Act requires it, with no fine in Italy but with the obligation to take measures. And in Italy, if the agent changes the way an office works, since 23 October 2026 its use has been part of the workplace risk assessment and of safety training (Legislative Decree No. 179/2026, Article 42), as explained in the guide AI literacy: what the AI Act requires of companies.
- Transparency. If the agent writes to customers, people must be able to understand that they are interacting with an AI, unless it is obvious (Article 50, which asks this of whoever provides the system).
- High risk. An agent that handles orders, invoices and stock is not a high-risk AI system: those uses are not in Annex III. It becomes one if it evaluates workers on the basis of their operations or if it decides the creditworthiness of natural persons, and sole traders are natural persons; for these uses Article 6 excludes narrow procedural tasks and preparatory tasks, unless there is profiling. In Italy, since 23 October 2026, decisions on the employment relationship cannot be taken solely by an automated system and the market surveillance authority is the ACN, the National Cybersecurity Agency (Legislative Decree No. 179/2026): the details are in the guide AI Act: what a manufacturing company must do.
How to measure it
Before and after, on a single workflow: minutes per order entered, errors found late, days of delay in customer payments, hours spent by administration (accounts) on purchase invoices. And a criterion that Microsoft’s documentation for its own invoice agent suggests, when it says that the goal is a draft needing few or no corrections: how many proposals are accepted without corrections. The action log is the source of the measurement, the fixed-scope pilot the way to carry it out: a few weeks, one department, the drafts compared with what people would have done.
The most frequent examples in a manufacturing company: the order that arrives as a PDF and becomes a draft order with codes, quantities and dates; the order confirmation prepared from the “order confirmed” webhook with the dates from the production plan; the price list indexed to the metal, recalculated every month with the LME average and the exchange rate taken via API, the calculations done by code and the draft approved by sales; the payment reminders prepared from the schedule of payments due, excluding customers with open disputes; the purchase invoices matched to orders and delivery notes, with the account and VAT proposed and the discrepancies on indexed prices flagged, as in the guide on the agent for purchasing; the check that the material certificate matches the alloy and the batch on the delivery note, read-only and of the greatest value.
When it is not worth it (yet)
- When the ERP has neither APIs nor webhooks and the screens change with every update.
- When the customer and supplier records are dirty: for 45.2% of the companies in Italy that have considered AI without adopting it, the obstacle is data that is unavailable or of poor quality (ISTAT, December 2025).
- When nobody can confirm the drafts in good time: the drafts pile up and the workflow slows down instead of speeding up.
- When the process changes every month.
- When a rule handles the workflow on its own.
- When the workflow touches the data of natural persons and there is not yet a written rule on who can see and decide what.
How we work
We build AI agents that work inside the systems the company already uses, with the data and integrations service for the connections: APIs and webhooks where they exist, MCP where the ERP offers it, RPA as the last route. We do not sell ready-made connectors for one ERP or another: every connection is built for the case in hand, with the drafts, confirmations and log described here, and with FT Mercati’s metal prices when the workflow requires them. The CRM with AI agents that we use every day is our test bed, with its declared status. The path is the one in our method: one workflow, one pilot, one measurement.
Sources
- Microsoft Learn, Copilot FAQ, Business Central (permissions, training, EU Data Boundary)
- Microsoft Learn, Sales Order Agent overview (Business Central)
- Microsoft Learn, Payables Agent overview (Business Central)
- Microsoft Learn, Configure Copilot and agent capabilities (Business Central)
- Microsoft Learn, Business Central MCP Server overview and setup
- Microsoft Learn, API (v2.0) for Dynamics 365 Business Central
- Microsoft Learn, Working with webhooks in Dynamics 365 Business Central
- Microsoft Learn, Agents, Copilot, and AI capabilities in Dynamics 365 apps
- Microsoft Learn, Microsoft 365 Copilot connectors overview (synced and federated)
- Microsoft Learn, Federated connectors overview
- SAP, Joule Agents and Joule Assistants
- SAP, SAP Business One
- Oracle, Overview of SuiteTalk REST Web Services (NetSuite)
- Oracle, overview of the Oracle AI Agents for ERP (PDF)
- Salesforce, Agentforce: The AI Agent Platform
- Salesforce, Trusted AI: Key Principles (Trust Layer)
- TeamSystem, Artificial Intelligence (in Italian)
- Sistemi, Sistemi AI for PROFIS (in Italian)
- Passepartout, Artificial intelligence in Passepartout management software (in Italian)
- Zucchetti Digital Supply Chain, AI Factory (in Italian)
- Model Context Protocol, What is the Model Context Protocol (MCP)?
- Model Context Protocol, Specification 2026-07-28 (user consent, untrusted tools)
- Anthropic, Introducing the Model Context Protocol (25 November 2024)
- UiPath, What is Robotic Process Automation
- Microsoft Learn, Introduction to desktop flows (Power Automate)
- IETF, RFC 9110 HTTP Semantics, 9.2.2 Idempotent Methods
- IETF, The Idempotency-Key HTTP Header Field (Internet-Draft, revision 08 of 22 October 2026)
- NIST CSRC, glossary: least privilege
- Anthropic, Building effective agents
- Anthropic, Tool use with Claude
- OpenAI Agents SDK, Human-in-the-loop
- OpenAI Agents SDK, Guardrails
- OWASP GenAI Security Project, Agentic Security Initiative (Top 10 for Agentic Applications 2026)
- NIST AI 600-1, Generative Artificial Intelligence Profile (automation bias)
- AI Act, Article 6: classification rules for high-risk AI systems (AI Act Service Desk)
- AI Act, Annex III: high-risk AI systems (employment, creditworthiness)
- AI Act, Article 4: AI literacy
- AI Act, Article 50: transparency obligations
- Italy's Legislative Decree No. 179 of 7 October 2026, authorities and penalties, Gazzetta Ufficiale No. 234 of 8 October 2026 (in Italian)
- Italy's revenue agency (Agenzia delle Entrate), technical specifications for e-invoices, version 1.9 (in Italian)
- Italy's revenue agency (Agenzia delle Entrate), section on e-invoicing and electronic daily receipts (specifications 1.9.1, free services, delegated authorisations, SDICoop) (in Italian)
- FatturaPA, the e-invoice format
- Italy's Digital Administration Code (Codice dell'amministrazione digitale, Legislative Decree No. 82/2005), Article 44: preservation of electronic documents, on Normattiva (in Italian)
- The Garante (Italy's data protection authority), the text of Regulation (EU) 2016/679 (in Italian)
- ISTAT (Italy's national statistics institute), Imprese e ICT, year 2025, report of 15 December 2025 (PDF, in Italian)
- Eurostat, E-business integration (isoc_eb_iip): enterprises with ERP software, 2025
- Eurostat, Artificial intelligence by size class of enterprise (isoc_eb_ai): workflow automation with AI, 2025
Checked on 17 November 2026. Rules and deadlines change: always check the official source before you decide.