How to use Claude with Dynamics 365 Finance and Supply Chain
Short answer
Connect Claude to FSCM through a Model Context Protocol server that wraps OData and custom services, then choose where Claude runs: the Anthropic apps, the API, or Claude on AWS Bedrock or Google Vertex AI. If data cannot leave your network at all, use an open-weight model with the same tools.
Teams on FSCM who try Claude usually begin with documents: a long supplier contract, a quality procedure, a stack of change requests. Claude handles long text well, and that is useful. The step change comes when it can also look things up in the ERP, so a question about a late job is answered from the production order, not from memory.
Model Context Protocol (MCP) is the cleanest way to do that. You expose a small set of FSCM operations as tools, and any MCP-capable client, including Claude, can call them. The tools, not the model, define what is possible and what is logged.
The second decision is hosting. Claude is available through Anthropic's own apps and API and through cloud marketplaces such as AWS Bedrock and Google Vertex AI, which matters if your security team prefers a model reached inside an existing cloud contract. Microsoft also offers Claude models in some of its own AI offerings, but availability changes, so verify the current state before planning around it.
What teams use it for
Production order status with context
Ask which released production orders are at risk this week; Claude calls tools for orders, BOM shortages and routing status, then explains the ranking.
Contract and PO term comparison
Upload a supplier agreement and compare its payment terms and lead times with what is configured on the vendor and open PO lines.
Engineering change impact
Given an ECO description, Claude looks up where-used for the affected released product and lists open orders and inventory that the change touches.
Customer complaint investigation
Paste a complaint and have Claude pull the sales order, shipment, lot trace and quality order history into one timeline.
Close checklist assistant
Finance asks Claude to summarise unposted journals and open vendor invoice approvals by legal entity as a pre-close briefing.
Developer support
Use Claude Code or the chat app to reason about X++ extensions, data entity design and OData filters, with no ERP data involved.
Microsoft Dynamics 365 Finance and Supply Chain integration surface
- OData v4 data entities under the environment /data endpoint, the main read surface behind MCP tools.
- Custom X++ services exposed under /api/services for operations that should run application logic.
- Microsoft Entra ID app registration and OAuth client credentials or delegated flows to authenticate the MCP server to FSCM.
- Business events, optionally used to push notifications into a workflow that then asks Claude for a summary.
- Microsoft's own Dynamics 365 ERP MCP server capabilities, which Microsoft has been developing; check current availability and scope for your release and compare them with a server you control.
How it works
- 1
Decide the tool list
Choose five to ten read tools first: find released product, get production order, get purchase order lines, check on-hand by site and warehouse, vendor lookup.
- 2
Implement or adopt an MCP server
Either use a Microsoft-provided server where it fits, or write your own that maps each tool to a bounded OData query with fixed fields and a row limit.
- 3
Authenticate with least privilege
Use an Entra app mapped to an FSCM user whose roles match the tool list. For per-user answers, prefer a delegated flow so security follows the person.
- 4
Choose the Claude surface
Use a Claude team or enterprise plan with connector support for interactive use, or call the API, Bedrock or Vertex AI from your own application for embedded use.
- 5
Pin data handling terms
Review retention, training and regional terms for the plan or cloud service you chose and record them with your security review.
- 6
Test with real questions
Collect fifty questions people actually ask, run them, and check each answer against the screen in FSCM.
Example flows
At-risk jobs briefing
Trigger: A production manager asks which jobs will miss ship date this week.
- Claude calls list_released_production_orders with a date window.
- For each, it calls get_bom_shortages.
- It joins results to promised dates from sales order lines.
- It answers with a ranked list and the specific missing components.
Result: A briefing built from live data with the tool calls visible for checking.
ECO impact check
Trigger: An engineer pastes an engineering change request.
- Claude extracts the item number and revision.
- It calls where-used and open order tools.
- It lists affected orders, WIP and inventory on hand.
Result: A first-pass impact list for the change board, to be verified by an engineer.
Keeping data on-prem
If FSCM runs in the Microsoft cloud, Claude through Bedrock or Vertex AI keeps traffic in your chosen cloud account and region, which may satisfy many policies. If data must not leave your premises, such as defence work or an on-premises (local business data) FSCM deployment, a hosted Claude is not an option and you should run an open-weight model (Llama, Qwen, Mistral, gpt-oss class) on your own GPUs with vLLM or Ollama. Because MCP is an open protocol, the same server and tools work with that private model, so the integration investment carries over.
Security checklist
- Expose tools, not the OData endpoint; each tool has fixed fields, a row cap and a clear purpose.
- Run read-only first; add write tools later behind an approval step and a separate credential.
- Authenticate per user where answers depend on security roles, and always pass legal entity explicitly.
- Record tool name, arguments, keys returned and caller for every call.
- Treat tool output as untrusted text: a vendor note field can contain instructions, so keep tools from taking actions based on free text alone.
Ways to get it done
Claude connectors with a Microsoft or community MCP server
Effort: Days to configure and review.
Quickest for exploration if an existing server covers your entities and your security team accepts it.
Your own MCP server plus Claude via API, Bedrock or Vertex AI
Effort: Two to six weeks including security review.
Best when you want specific tools, auditing, and the model reached under an existing cloud agreement.
Private open-weight model with the same MCP tools, with a partner such as Netray
Effort: A pilot on one workflow, with evaluation against Claude-level answers to confirm quality is sufficient.
For data that cannot leave your network, or on-premises FSCM deployments.
Pitfalls
- !Pointing the model at the raw OData surface with a free-form query tool, which invites expensive, slow or over-broad requests.
- !Forgetting company context, so answers silently mix or omit legal entities.
- !Assuming a Microsoft-provided connector matches your customisations; extension fields may not be exposed.
- !Not testing with production-sized data, where unfiltered queries on large tables run into throttling.
- !Skipping the review of cloud region and retention terms until after rollout.
Frequently asked questions
Does Claude have a native Dynamics 365 FSCM connector?
Treat availability as something to verify. Microsoft has been building MCP access for Dynamics 365 ERP, and community servers exist, but scope and release support vary. Many teams write a small MCP server of their own so the tool list, fields and logging match their security policy.
Can Claude run through our existing Azure or AWS agreement?
Claude models are offered through AWS Bedrock and Google Vertex AI, which lets you use them under an existing cloud contract and region. Availability inside Microsoft services has changed over time, so confirm current options. Either way, the model access path is separate from how it reaches FSCM.
Is it safe to let Claude write to FSCM?
Not at first. Start with read tools, then add narrow write tools such as creating a draft note or a proposed PO change, requiring human approval. Avoid generic update tools. Writes should call application-level services so validation and workflow still run.
How is this different from the ChatGPT approach?
The pattern is similar: a governed layer between model and ERP. With Claude, MCP is the natural layer and works across clients. The same MCP server can serve other models too, which keeps you from being locked to one vendor while you decide.
What if our FSCM is on-premises?
Reach it from an MCP server inside your network, authenticating through your identity provider such as AD FS. The model then has to be inside that boundary too if data cannot leave, which means an open-weight model. Check which data entities your on-premises build supports.
More for Microsoft Dynamics 365 Finance and Supply Chain
Want Claude working on your Microsoft Dynamics 365 Finance and Supply Chain?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment