AI Integration with Business Systems: ERP, CRM and Legacy

Integrating AI with existing business systems is often the most complex (and most underestimated) part of an automation project. AI technology itself is relatively mature; the difficulty lies in making it work with a heterogeneous IT landscape: ERPs built in the 1990s, heavily customized CRMs, on-premise databases with proprietary data structures, and dozens of departmental applications.

The first step is to map the API catalog of the existing systems. Modern systems such as Salesforce, HubSpot, SAP S/4HANA and Microsoft Dynamics 365 offer complete, documented REST APIs. Legacy systems often have no APIs: in those cases teams turn to alternative techniques such as RPA to scrape the graphical interface, database connectors for direct data access, or integration middleware such as MuleSoft and Dell Boomi.

In Italy, the specific shape of the corporate technology landscape matters: many SMEs run Italian vertical management software such as TeamSystem, Zucchetti, Ad Hoc Revolution and similar products. Yellow Tech has developed connectors and integration patterns specific to these systems, building experience across 300+ automation projects with Italian clients.

The simplest pattern is integration through REST APIs: the AI agent calls the business system’s APIs to read or write data. It works perfectly with modern, well-documented systems. The second pattern is integration middleware: a platform such as MuleSoft, n8n or Azure Integration Services acts as a hub, translating data formats and orchestrating the flows between AI and business systems.

The third pattern, needed for systems without APIs, is database-level integration: the AI agent reads and writes directly in the company database tables through SQL connectors. It calls for care with security and transaction handling, but it works with any system. The fourth pattern, emerging in 2025-2026, is AI native to the system: SAP has built AI (Joule) into S/4HANA, Salesforce has introduced Einstein Copilot, and Microsoft has added Copilot across the whole Dynamics 365 suite.

The choice of pattern depends on API availability, latency requirements, data volumes and security requirements. For systems that handle personal data (GDPR) or sensitive financial data, the integration must be designed with encryption in transit and at rest, a complete audit log and role-based access control (RBAC).

SAP launched Joule in 2023, its AI copilot built into S/4HANA. Joule lets users query the ERP in natural language, automate recurring tasks and get insights from company data. For SAP users, the fastest AI integration is to enable Joule. For more customized automation, SAP exposes REST APIs through the SAP API Business Hub, with more than 2,000 documented endpoints.

Oracle has introduced Oracle AI Services (now Oracle Fusion AI), built into the Oracle Cloud applications. Microsoft Dynamics 365 is perhaps the most advanced in native AI integration: Copilot is available in every module (Sales, Finance, Supply Chain, Customer Service), with automatic email drafting, conversation summaries and suggested actions.

For companies running on-premise ERPs (common among Italian mid-sized companies), AI integration typically requires a middleware layer. Yellow Tech has developed connectors for SAP ECC, Oracle E-Business Suite and Microsoft Dynamics NAV/Business Central, handling the data transformation between the ERP’s proprietary formats and the LLM APIs.

Legacy systems (applications built decades ago, with no APIs and proprietary databases) are the biggest challenge in AI integration. There are three main approaches. The first is the RPA-AI bridge: an RPA bot interacts with the legacy system’s interface, passes the data to the AI for processing, and writes the results back into the system. Slow, but it works with almost any system.

The second approach is the database connector: if you have access to the underlying database (typically SQL Server, Oracle DB, MySQL), the AI can read and write data directly through SQL queries. It requires an analysis of the database schema and care over data consistency. The third approach is gradual modernization: placing a modern API layer alongside the legacy system that exposes its data in a standard format, without replacing the underlying system. This is the route often taken for critical systems that cannot be touched.

A public example of this kind of work is Nital, the official Italian distributor of Nikon and DJI and a Yellow Tech client. In its customer story, the IT Manager describes how reconciling courier invoices, previously handled by hand in accounting, moved to an automated flow: a document integration on a management system already in production, without replacing the underlying system.

One of the most expensive mistakes is underestimating the complexity of legacy integration. Yellow Tech typically allocates 30-40% of an automation project’s budget to systems integration, a share that surprises clients who assume the expensive part is the AI model. In reality, the AI model is a commodity; the value is created in the integration with the company’s specific context.

Step one: inventory the systems. Start with a survey of what the agent will need to read and write: legacy management systems, ERPs, CRMs, databases and departmental applications. For each one, record API availability, the version in production, the internal system owner and the category of data processed. This inventory is also the documentary basis for the activity log that the AI Act requires for high-risk systems.

Step two: technical and security assessment. For management systems without modern APIs, consider an abstraction gateway that exposes the data in a standard format while leaving the underlying system untouched. On access, the agent’s credentials must be kept separate from human accounts through IAM, with permissions limited to the data the task needs (data minimization, Article 5 GDPR). This is also the phase where you decide where the models run: for companies with data residency requirements, the route is an on-premise or EU-hosted deployment.

On the most widespread systems, the connection work has already been done and can be reused. Yellow Tech has developed connectors for SAP ECC and S/4HANA, Oracle E-Business Suite, Microsoft Dynamics NAV and Business Central on the ERP side, and for the Italian vertical management systems most common in mid-sized companies, including TeamSystem, Zucchetti and Ad Hoc Revolution. Starting from an existing connector shifts the work from building the integration to validating it on the client’s data.

Step three: implementation and testing. Writes to management systems are validated first in a sandbox, on a copy of the data, because a mapping error on a production ERP is expensive to fix. Then come the connectors to the channels where people work (Teams, Slack) and the logging of every run with input, output and timestamp. Yellow Tech typically allocates 30-40% of the project budget to this phase, and across 300+ automation projects with Italian clients it is the item that decides whether the integration succeeds.

Every AI integration that accesses company data raises security and compliance questions that must be addressed at the design stage. The first requirement is data minimization (a GDPR Article 5 principle): the AI agent must access only the data strictly necessary for its task, not the entire company database.

The second requirement is data localization: for companies subject to data residency requirements (common in Finance and Healthcare), data cannot be sent to the cloud APIs of non-European providers. In these cases, Yellow Tech configures on-premise or EU-hosted deployments of open source models (Mistral, LLaMA) that keep the data in Europe.

The third element is the AI activity log: required by the AI Act for high-risk systems, it is best practice for every system that makes automated decisions about people or company assets. Every run must be logged with input, output, timestamp and, where relevant, an explanation of the reasoning. For more on the regulatory side, see our guide to the AI Act for Italian companies.