The short answer. The leading platforms for connecting AI to ERP and CRM are MuleSoft Anypoint, Boomi, Workato, Azure Logic Apps, SAP Integration Suite, Oracle Integration, and IBM App Connect. Each now exposes governed connectors or APIs to AI agents, often through the Model Context Protocol (MCP). The best choice usually follows your dominant ERP, cloud, and security model, not a feature checklist.
Why AI needs an integration layer
AI agents are only useful when they can safely read and act on live orders, invoices, and customer records. Most of that data sits in ERP systems such as SAP S/4HANA, Oracle Fusion, and Dynamics 365, and in CRMs such as Salesforce.
Wiring a model straight into those systems is risky. An integration platform sits between them and handles authentication, mapping, rate limits, approvals, and audit trails. The model assists with reasoning and language; the platform controls which data and actions are available.
This guide compares seven platforms on ERP and CRM coverage, AI features, security, deployment, and fit, using vendor documentation checked in October 2026. Preview and roadmap features change quickly, so confirm status before you commit.
Key terms in plain English
iPaaS (integration platform as a service): cloud software that connects applications, data, and APIs through prebuilt connectors and workflows.
AI agent: software that uses a large language model (LLM) to plan steps and call tools to finish a task.
MCP (Model Context Protocol): an open standard that lets AI models discover and call external tools, such as an ERP API, in a consistent way.
Native vs external AI: native features ship inside the platform; external ones need your own LLM account or custom code.
How AI agents securely interact with ERP and CRM
A secure design never gives an AI model direct database access. The agent calls a small set of approved tools, each a narrow business action such as "get customer credit status" or "submit quote for approval", and the platform enforces identity, permissions, and logging on every call.
Trigger: a user request or business event, such as an overdue invoice, starts the agent.
Tool discovery: the agent sees only approved tools, often published through an MCP server.
Authentication: OAuth or a managed identity verifies the agent, ideally acting for the signed-in user.
Authorization: scopes, rate limits, and data masking limit what each tool can read or change.
Human approval: risky writes, such as refunds or vendor changes, pause until a person approves.
Execution: connectors call ERP and CRM APIs, map data, and handle retries.
Audit: every prompt, tool call, and result is logged for review.
Example: a customer asks an agent to change a delivery date. The agent calls an approved order-lookup tool; the platform checks access and fetches current ERP data, validates the change, waits for any required approval, then commits it. The agent reports the verified result from the ERP rather than assuming the change worked. MCP standardizes tool discovery, but it does not replace authorization, data classification, or approval. For one example of these controls, see LuMay's security and governance approach.
1. MuleSoft Anypoint Platform
MuleSoft, part of Salesforce, combines integration with full API lifecycle management. It fits large enterprises that want reusable, governed APIs, especially where Salesforce is strategic.
ERP and CRM: Official connectors cover SAP (including an S/4HANA OData connector for cloud and on-premise), Oracle E-Business Suite, NetSuite, Workday, Dynamics 365, and Salesforce.
AI and agents: Agent Fabric is a native control plane for governing agents, with an MCP Bridge that turns existing APIs into MCP servers, MCP and agent-to-agent (A2A) connectors, an LLM proxy, and an agent kill switch. OpenAI and Einstein models use separate connectors.
APIs and sync: Mule apps handle orchestration and transformation; Anypoint MQ adds asynchronous queues.
Security and deployment: SSO, roles, audit logs, and a secrets manager are built in. Run on CloudHub 2.0, your own Runtime Fabric (AWS, Azure, VMs), hybrid, or Private Cloud Edition. MuleSoft Government Cloud is FedRAMP Moderate.
Complexity and cost: High; plan for skilled Mule and DataWeave developers. Licensing is consumption-based on Mule Credits.
Main trade-off: strong governance and reuse come with architecture and operating complexity. Best for Salesforce-centric enterprises with a central integration team.
Official source: MuleSoft Anypoint Platform
2. Boomi Enterprise Platform
Boomi is a low-code iPaaS that builds integrations from visual shapes and prebuilt connectors. It suits organizations needing broad application, API, data, and B2B/EDI coverage without heavy coding.
ERP and CRM: Connectors cover SAP (through the Boomi for SAP add-on on NetWeaver ABAP systems), NetSuite, Dynamics 365 Finance and Operations, Dynamics 365 Customer Engagement, Workday, and Salesforce.
AI and agents: Native Agentstudio builds agents with guardrails (Agent Designer), deploys them (Agent Garden), and governs Boomi and Amazon Bedrock agents (Agent Control Tower). Agents act as MCP clients, and managed APIs can be exposed as MCP servers. The MCP Server connector is a tech preview; OpenAI and Bedrock connectors need your own accounts.
APIs and sync: Event Streams and SAP change capture support near-real-time sync, alongside batch and API-led patterns.
Security and deployment: Runtimes run in Boomi's cloud or on your own Windows, Linux, or Docker hosts, with secrets in AWS, Azure, or Google Cloud vaults. Boomi lists SOC 2, ISO 27001, and HIPAA among its attestations.
Complexity and cost: Moderate. Pricing is an edition-based subscription with a pay-as-you-go entry plan.
Main trade-off: SAP access needs an ABAP add-on, and complex transformations still need experienced architects. Best for teams that want low-code speed with enterprise breadth.
Official source: Boomi Enterprise Platform documentation
3. Workato
Workato is a cloud automation platform built around "recipes," which are event-driven workflows. It fits IT and business teams building automations together, including AI agents inside Slack or Microsoft Teams.
ERP and CRM: Prebuilt connectors include SAP OData (S/4HANA Cloud public and private editions), SAP RFC, Oracle Fusion Cloud, Oracle E-Business Suite, NetSuite, Workday, Dynamics 365, Business Central, Salesforce, and HubSpot.
AI and agents: Agent Studio builds "genies," agents that combine an LLM, a knowledge base, and recipes as skills. Genies use Anthropic Claude by default, with OpenAI or your own model as options. Enterprise MCP hosts MCP servers over your API collections, secured with OAuth and verified user access.
APIs and sync: Event Streams offers pub/sub with guaranteed delivery; high-volume recipes handle bulk jobs.
Security and deployment: RBAC, SAML SSO, encryption, and audit logs are standard. Compliance includes SOC 2, ISO 27001, ISO 42001, HIPAA, and PCI DSS; FedRAMP is not claimed. It is cloud-only, with an on-premises agent tunnelling out to firewalled systems.
Complexity and cost: Low build effort. Usage-based pricing is quoted.
Main trade-off: no self-hosted runtime, and heavy agent activity needs close monitoring of usage and permissions. Best for fast-moving business automation.
Official source: Workato documentation
4. Microsoft Azure Logic Apps
Logic Apps is Microsoft's low-code workflow service on Azure, with more than 1,400 connectors. It suits organizations standardized on Azure, Microsoft 365, and Dynamics 365.
ERP and CRM: The SAP connector supports S/4HANA, ECC, and R/3 through IDoc, BAPI, and RFC. Microsoft connectors cover Dynamics 365 Finance and Operations, Business Central, Dataverse (for Dynamics 365 Sales), and Salesforce; HubSpot relies on a community preview connector.
AI and agents: Agent workflows run an LLM "agent loop" natively, and any connector action can become an agent tool. Models come from Azure OpenAI or Microsoft Foundry. Standard logic apps can also act as MCP servers (preview).
APIs and sync: Integration accounts add B2B and EDI; Service Bus, Event Grid, and Event Hubs enable event-driven sync. Delivery is at-least-once, so design idempotent workflows.
Security and deployment: Microsoft Entra ID secures agent and MCP endpoints. Standard plans add VNet integration and private endpoints. Hosting spans multi-tenant Consumption, single-tenant Standard, and hybrid on your own infrastructure.
Complexity and cost: Low on Consumption; Standard networking and SAP setup add work. Pricing is per execution or hosting plan, plus model usage.
Main trade-off: agent model data is not guaranteed to stay in-region, and some agent and MCP features are still in preview. Best for Azure-first organizations.
Official source: Azure Logic Apps
5. SAP Integration Suite
SAP Integration Suite is SAP's iPaaS on the Business Technology Platform (BTP). It is the default choice for SAP-centric enterprises, including those migrating from SAP PI/PO.
ERP and CRM: SAP-to-SAP integration is native, with prebuilt content for S/4HANA and other SAP applications. Adapters for non-SAP systems include Salesforce, Workday, NetSuite, and HubSpot.
AI and agents: Integration flows can be generated from plain-language prompts. An AI adapter calls SAP's generative AI hub. MCP Gateway (Enhanced edition) creates and governs MCP servers from APIs and flows, with OAuth and rate limiting. SAP lists an AI gateway and agentic orchestration as roadmap items, not released features.
APIs and sync: API Management, business events, queues, and B2B libraries are included; Enhanced adds Advanced Event Mesh.
Security and deployment: Tenants run on SAP BTP. Edge Integration Cell is a hybrid runtime on your own Kubernetes, so sensitive data can stay inside your landscape.
Complexity and cost: Moderate for SAP teams, steeper for mostly non-SAP estates. Pricing is per tenant by edition (Starter, Standard, Enhanced) with a monthly message allowance.
Main trade-off: AI features depend on edition, and roadmap items should not be treated as available. Best for organizations whose core ERP is SAP.
Official source: SAP Integration Suite
6. Oracle Integration
Oracle Integration 3 is a fully managed service on Oracle Cloud Infrastructure (OCI). It covers integration, process automation, and B2B, and suits Oracle Fusion Cloud and NetSuite customers.
ERP and CRM: Adapters cover Oracle ERP Cloud, NetSuite, Salesforce, Microsoft Dynamics 365 CRM, HubSpot, and SAP S/4HANA Cloud (outbound OData calls).
AI and agents: An AI assistant helps build integrations and fix errors. AI Agent Designer builds agents with memory and human approvals, first released through an early-access program. Integrations can be registered as agent tools and published as an OAuth-secured MCP server. OCI Generative AI works natively; OpenAI and Anthropic models need an LLM connection.
APIs and sync: Event-driven and scheduled integrations, B2B partner management, and embedded SFTP cover most patterns.
Security and deployment: Private endpoints keep traffic on private networks, and a connectivity agent reaches on-premises systems. Oracle also offers the service in its EU Sovereign Cloud and a government edition.
Complexity and cost: Lower for Oracle-skilled teams. Licensing uses Universal Credits as hourly message packs.
Main trade-off: agent features are new, and LLM calls stay in-region. Best for Oracle-first ERP estates.
Official source: Oracle Integration documentation
7. IBM App Connect
IBM App Connect is hybrid integration software that runs on-premises, on any cloud, in containers, or as a managed service. It fits regulated enterprises with IBM MQ, mainframe, or strict hybrid requirements.
ERP and CRM: App Connect Enterprise 13 covers Dynamics 365 Finance and Operations, Dynamics 365 for Sales, Workday, Salesforce, and SAP OData; version 13.0.7 adds SAP S/4HANA nodes.
AI and agents: An App Connect MCP server can expose connector actions and REST API flows as MCP tools. An Azure OpenAI node handles chat completions and embeddings, and a watsonx.ai connector is available. App Connect skills can also be used inside IBM watsonx Orchestrate.
APIs and sync: Event-driven SAP nodes, IBM MQ, and REST flows cover real-time and asynchronous patterns.
Security and deployment: Deploy on-premises, in containers, on z/OS, or as a service on AWS. IBM also offers an IBM-managed SaaS option through webMethods Hybrid Integration.
Complexity and cost: Higher; toolkit development and container operations need specialists. The SaaS edition is usage-based; enterprise licences are quoted.
Main trade-off: more specialist effort, and not every connector action is available as an MCP tool. Best for organizations already invested in IBM middleware.
Official source: IBM App Connect
Comparison: AI-to-ERP and CRM integration platforms
All seven connect AI to ERP and CRM; they differ most in agent maturity, deployment options, and pricing.
Platform | Best fit | Native agent features | MCP support | Deployment | Pricing model |
|---|---|---|---|---|---|
MuleSoft Anypoint | Salesforce-centric, API-led enterprises | Agent Fabric, LLM proxy, kill switch | MCP Bridge; MCP connector | CloudHub 2.0, Runtime Fabric, hybrid, private cloud | Consumption (Mule Credits) |
Boomi | Low-code breadth, including B2B/EDI | Agentstudio: Designer, Garden, Control Tower | MCP client; MCP Server connector in tech preview | Boomi cloud or self-hosted runtimes | Edition subscription; pay-as-you-go entry |
Workato | Business-led automation | Agent Studio genies | Hosted Enterprise MCP servers | Cloud only, plus on-prem agent | Usage-based, quoted |
Azure Logic Apps | Azure and Microsoft 365 estates | Agent workflows (agent loop) | MCP server on Standard (preview) | Consumption, Standard, ASEv3, hybrid | Per execution or hosting plan |
SAP Integration Suite | SAP-first ERP landscapes | AI-generated flows; generative AI hub adapter | MCP Gateway (Enhanced edition) | SAP BTP plus Edge Integration Cell | Per tenant by edition, message allowance |
Oracle Integration | Oracle Fusion and NetSuite customers | AI assistant; AI Agent Designer | Integrations published as an MCP server | OCI managed, private endpoints | Universal Credits, hourly message packs |
IBM App Connect | IBM middleware and regulated hybrid estates | watsonx Orchestrate skills; Azure OpenAI node | App Connect MCP server | On-premises, containers, z/OS, SaaS | Usage-based SaaS; enterprise quote |
How to choose: seven selection criteria
Start from your system of record, then test agent readiness and governance.
ERP gravity: favor the platform closest to your core ERP; native content cuts mapping work.
Agent standards: check MCP and agent-to-agent support, and whether each feature is generally available or still in preview.
Identity model: prefer platforms that let agents act on behalf of the signed-in user, not through a shared admin account.
Deployment and residency: confirm hybrid or private runtimes if ERP data must stay on-premises or in a specific region.
Governance: look for audit logs, approval steps, rate limits, and a kill switch.
Skills and speed: low-code tools ship faster; pro-code platforms give more control over complex APIs.
Cost model: agents call tools far more often than people, so model how that changes credits, messages, or tasks.
Implementation checklist for AI-to-ERP integration
Ship a narrow, read-only use case first, then add approval-gated writes.
☐ Pick one or two high-value, low-risk use cases, such as order-status or invoice lookups.
☐ Inventory the ERP and CRM APIs, objects, and data owners involved.
☐ Define tools narrowly: one business action per tool, with typed inputs and outputs.
☐ Start read-only, then add writes that pause for human approval.
☐ Use OAuth with least-privilege scopes; never share admin credentials with an agent.
☐ Mask personal and financial fields the model does not need.
☐ Set rate limits, timeouts, retry rules, and idempotency for every tool.
☐ Log prompts, tool calls, and results to your SIEM.
☐ Test in a sandbox with real-world requests, edge cases, and prompt-injection attempts.
☐ Agree on a rollback plan and a kill switch before go-live.
☐ Track accuracy, latency, and cost per completed task after launch.
Common integration challenges and how to handle them
Most AI-to-ERP problems come from data, permissions, and cost, not the model.
Challenge | Why it happens | How to reduce it |
|---|---|---|
Messy master data | Duplicate customers or inconsistent codes confuse the model | Clean key records; validate before acting |
Over-privileged access | Agents inherit a broad service account | Least-privilege scopes and on-behalf-of user identity |
Preview features | MCP and agent capabilities are moving fast | Confirm general availability in writing |
Latency | Multi-step agent loops chain several API calls | Cache reference data and keep tools coarse-grained |
Cost spikes | Agent retries multiply credits, messages, or tasks | Budgets, rate limits, and per-task cost tracking |
Unclear ownership | Integration, AI, and ERP teams each own a piece | Name one owner per tool and a shared runbook |
If your team lacks capacity to design this layer, LuMay AI integration services help connect AI to existing applications and workflows without replacing them.
Recommendations: which platform fits which enterprise
Match the platform to your dominant ERP and cloud, then validate it in a pilot.
Salesforce plus a mature API program: MuleSoft Anypoint, for governed APIs and Agent Fabric controls.
SAP S/4HANA at the core: SAP Integration Suite, adding another iPaaS only for real gaps.
Azure and Microsoft 365 standard: Azure Logic Apps, for native agent workflows and Entra ID security.
Oracle Fusion or NetSuite first: Oracle Integration, for managed adapters and MCP publishing.
Business-led speed: Workato, for fast recipes and agents inside collaboration tools.
Low-code breadth with B2B/EDI: Boomi, for wide connector coverage and Agentstudio governance.
IBM middleware or regulated hybrid: IBM App Connect, for deployment flexibility and MQ strength.
Pilot one read-only and one approval-gated use case, then compare accuracy, latency, and cost. For many enterprises the answer is not one "best" iPaaS but a primary platform plus clear coexistence boundaries for existing middleware and departmental tools. Agents also need well-scoped tools, prompts, and guardrails, which is where AI agent development and workflow automation agents come in.
Conclusion
Connecting AI to ERP and CRM is now more an integration problem than a model problem. Choose the platform that matches your system of record, prove it with a narrow pilot, and scale only once security and cost are under control.
