Which supply chain tasks should run automatically, and when can a language model help with analysis?
In procurement, some order proposals are checked every day and then approved without changes. Other questions have no fixed workflow: Why has inventory increased? Which suppliers need attention? What should management know this week?
numi provides an approach for both types of task: Native numi agents automate recurring workflows based on numi's decisions. Through the numi MCP server, ChatGPT or Claude use the available data and functions for targeted analysis, flexible evaluations and summaries. Both approaches build on numi's supply chain intelligence.
What are AI agents in the supply chain?
AI agents in the supply chain are software systems that handle specific tasks in planning, procurement or inventory management. They process information, identify where action is needed and carry out agreed steps. This might mean creating a purchase order in the ERP, updating an inventory parameter or notifying the responsible person.
The degree of independence is defined in advance: an agent can monitor a process, request approval or act automatically within agreed limits. The term agentic AI is often used for AI systems that work through tasks in multiple steps using tools. The implementation can vary by application.
In numi, native agents are workflows configured directly in numi, combining a trigger, checks and an action. A routine workflow governed by clear rules does not require an additional call to a large language model, also known as an LLM.
Why numi offers native agents and an MCP server
Both approaches share numi's forecasting, optimisation and decision logic. This connects factors such as demand, available stock, open purchase orders and lead times with the company's objectives and constraints. The result is specific recommendations for planning and procurement.
The intelligence is already in the supply chain calculations. An order proposal, for example, considers required quantities, lead times, minimum order quantities and target availability. Here, “optimised” means suited to the data, objectives and constraints used. Even a sound calculation needs to be reassessed when those inputs change.
On this foundation, numi supports two ways of working:
- Recurring execution: A native agent implements calculated recommendations according to defined rules and sends exceptions for review.
- Flexible investigation: A language model accesses the numi data and functions made available through MCP, investigates a specific question and presents the findings clearly.
The same supply chain intelligence can therefore support both daily routines and individual questions.
When do we recommend native numi agents?
We recommend native numi agents for recurring tasks with clearly defined workflows and approval limits. They are particularly useful for processes that employees regularly review and almost always approve without changes.
One example is purchase order approval in procurement:
- numi calculates an order proposal using forecasts, inventory, open purchase orders and lead times.
- The agent checks the proposal against agreed rules, such as quantity, order value and approved suppliers.
- If the proposal is within those limits, the purchase order can be created in the ERP through the configured integration. All other cases are sent for human approval, together with an explanation.
The workflow can start on a schedule or in response to an event. Other applications include the controlled application of calculated inventory parameters and passing identified risks to the responsible teams.
For execution governed by clear rules, a language model does not need to interpret each transaction again. This routine path therefore requires no additional LLM calls and no additional LLM tokens per execution. The supply chain calculations and software operation are still required. If further functions are added, such as interpreting free-form supplier replies, a language model may also be used for those tasks.
When do we recommend MCP with ChatGPT or Claude?
We recommend the numi MCP server for targeted analysis, flexible evaluations or management summaries. The question is described in natural language and can be refined during the conversation.
MCP stands for Model Context Protocol. This open standard allows AI applications to access external data and tools. An MCP server makes those connections available; the AI application calls them. The MCP project explains the interface and its tools in detail.
The numi MCP server can help address questions such as:
- “Which items contribute most to the increase in inventory? Investigate the available inventory and demand data.”
- “Compare the delivery performance of the available supplier groups and highlight notable trends.”
- “Summarise the main inventory, delivery and supply risks for management. Include the underlying KPIs.”
The advantage is flexibility: an initial overview can lead to a follow-up question about an item, period or supplier. The language model uses the available tools and information to help interpret and explain results. The scope of analysis depends on the data, functions and permissions made available. Statements and KPIs in a summary should be traceable to the underlying data.
MCP also supports actions, provided that suitable tools and permissions are available. This is documented by both OpenAI for MCP tools and Anthropic for Claude connectors. Our recommendation for analysis describes a suitable use within numi; it does not mean MCP is technically limited to read-only access.
Native numi agents and MCP compared
The choice depends on the task. The following table describes how we recommend using both approaches with numi.
| Criterion | Native numi agents | Language model through the numi MCP server |
|---|---|---|
| Typical task | Execute recurring recommendations according to agreed rules | Targeted analysis, flexible evaluations and management summaries |
| Example | Transfer order proposals to the ERP within approval limits | Investigate inventory trends and explain them to management |
| Workflow trigger | Configured schedule or defined event | A question or analysis request in the use described here |
| Language model | No additional LLM call required for routine execution governed by clear rules | LLM for natural language interaction, tool use and presenting results |
| Control | Defined rules, approval limits and exception review | Permitted tools, permissions and required confirmations |
| Supply chain intelligence | Data, forecasts, optimisation results and functions made available by numi | |
MCP is a connection standard, not a separate type of agent. An agent using a language model can also use MCP. A broader workflow can also start automatically when configured accordingly. For numi, the practical distinction is therefore: which recurring steps can be configured in advance, and where does flexible work with a language model add value?
How both approaches work together in procurement
A possible workflow shows how routine execution and analysis complement each other: numi calculates new order proposals. A native agent checks the configured approval rules, transfers suitable proposals to the ERP and passes exceptions on for a decision.
Later, the head of procurement wants to understand why purchasing volume has changed. Through the MCP connection, they ask for the available numi data to be investigated and summarised. During the conversation, they can explore individual product groups or periods in more detail and check the KPIs.
The recurring process runs within the agreed limits. The individual analysis helps explain changes and, where appropriate, informs adjustments to planning objectives or rules. An analysis in a chat does not automatically change those rules; that requires an explicitly authorised action.
How do you get started with supply chain agents?
A useful starting point is a manageable process built on sound supply chain data and logic:
- Choose the task: Which recurring transactions are frequently approved without changes today? Which analyses need changing questions and explanations instead?
- Clarify data and objectives: Inventory, demand, lead times and master data must be usable for the chosen process. Objectives such as availability and inventory value need to align with the constraints.
- Define permissions and responsibilities: Who can read data, trigger purchase orders or change parameters? Which cases need approval, and who handles missing data or a failed action?
- Automate gradually: Start by monitoring what the agent would do. Then work with human approval. Only after review by the responsible experts should suitable cases run automatically within defined limits.
- Review results: Monitor processing time, approval rates, corrections and effects on supply. Adjust rules and data where needed.
These limits are often called guardrails. They include quantity and value limits, approved suppliers and escalation paths. Such controls, logging and suitable system integration matter for both approaches. The MCP standard alone does not define the company's business rules.
Frequently asked questions about AI agents and MCP in the supply chain
Do AI agents always need a language model?
No. A native numi agent can execute calculated recommendations according to defined rules without an additional LLM call. A language model is useful when questions need to be phrased flexibly, information needs interpretation in natural language or results need to be prepared as a report. Both approaches can be combined according to the task.
Does an MCP server consume tokens?
MCP is an interface standard. LLM tokens are used when the language model processes inputs, tool descriptions, results and responses. How this translates into costs depends on the product and plan used. The native routine path described here does not need an additional LLM call per execution.
Can ChatGPT or Claude also execute actions through MCP?
Yes, if the MCP server provides suitable tools and the necessary permissions and approvals are in place. These can include actions that change data. When choosing between native numi agents and a language model using MCP, we recommend considering repeatability, analysis needs and the desired level of workflow control.
Do AI agents replace procurement and inventory planners?
Agents handle specific tasks. The responsible teams set objectives and limits, review exceptions and contribute their knowledge of customers, suppliers and operational requirements. At numi, automation is intended primarily to handle recurring execution, while analysis supports people with specific questions.
Which numi approach fits your process?
If your team repeatedly approves calculated proposals without changes, native agents are worth exploring. If the team wants to investigate developments and prepare information for different questions, the MCP connection with a language model is a suitable starting point. Many teams need both ways of working.
Visit AI agents and supply chain automation with numi to explore the available functions. In an intro call, we can look together at a specific routine process and a typical analysis question.
