A single supplier for a component, open invoices, an announced delay, a change of managing director – each signal on its own is harmless.
The question
You ask Mara – in one sentence.
Instead of building reports, you simply ask. Mara shows what gets checked – step by step, across all sources.
The evidenced answer
Every finding comes with its citation.
Every statement names the source it comes from. Whatever can't be evidenced, the agent marks as open.
The next step
Mara prepares – you approve.
Mara prepares the next step. It is only carried out once you approve it.
Hi, I'm Mara – your AI agent for procurement. I keep an eye on your systems and speak up when something stands out.
I noticed four signals today – each inconspicuous on its own:
Signals · today
ERPComponent 4711
Lexware2 invoices open
OutlookDelivery delay announced
c:node GraphCommercial register
@Mara, who is our biggest default risk?
I'm checking all sources:
Read supplier master data from the ERP
Reconciled open items
Checked framework agreements
Checked commercial register and insolvency notices
Weber Metall is your no. 1 procurement risk:
Sole supplier for component 4711 per framework agreement
Two invoices open for more than 60 days
Change of managing director per commercial register
SharePoint · Framework agreement § 3Lexware · Open itemsCommercial register
I've prepared the next step:
Draft · Request to second supplier
Request for quotation for component 4711 to Stahlbau Süd – with quantities and delivery dates from the ERP.
ApproveEditNothing goes out without approval
Example case with sample data.
c:node Research · Procurement & Sourcing
Why procurement risks show up too late — and how traceable AI changes that.
Reading time ~5 minBasis 6 sources, verifiedTopic Supplier default riskApproach source-backed · on-prem possible
Fig. 0 · Procurement & warehouse: signals of supplier failure arise here — but they often become visible too late.
Summary
Supplier failures announce themselves — but the signals are scattered across ERP, accounting and supplier files, and general AI guesses more than it knows. This article shows, with sources, why that is expensive and why the obvious solutions fail — and how c:node makes risks visible early: it connects ERP, accounting and supplier files with the commercial register and backs every statement with its source.
1 The problem: the signals are there — just not connected.
A critical supplier rarely fails without warning. Overdue invoices, declining delivery reliability, dependence on a single source — the clues already exist, but they sit in separate systems and are not brought together. Peer-reviewed research on supply chain risk management shows that the dangerous risks in particular are multi-tier and hidden: they arise from chains across companies, products and sites — exactly the connections that tabular analyses systematically miss.1
2 Why the obvious solutions fail.
Two approaches dominate today — both with a built-in weakness. Manual analysis in Excel and individual systems is slow, point-in-time and breaks down with chains. And generic AI chatbots? The US standards institute NIST states that fabricated statements are inherent to the design of such models — they predict the statistically most likely word, not the traceable one.2 Worse still: they supply fabricated justifications and sources along with it, tempting you to trust what is wrong — especially risky for decisions with consequences.2 Even when you give the AI the real data to look up, the error rate only drops — it does not disappear.3
3 The approach: look it up first, then think along traceable connections.
c:node connects ERP, accounting and supplier files and doesn't let c:node narrate freely — instead it reasons over a knowledge graph: your knowledge as a network of nodes and edges. Research calls this combination of pattern recognition and firmly checked rules neuro-symbolic AI; it delivers answers that can be justified and traced.4 Via the graph, c:node thinks several steps ahead — overdue invoice → supplier → sole critical part → affected line — which simple search demonstrably cannot do.5
:nodeGraph— Procurement / default risklive
Source (API)EntityRiskRule / evidence
Nodes active · edges fire in real timeisolated per tenant · grows with every piece of evidence
Fig. 1 · The knowledge graph connects sources, suppliers, invoices and rules. c:node fires the path to the risk — every step stays bound to its source.
4 In practice: scattered signals become a traceable recommendation.
The same question that today costs days of manual research, c:node answers in seconds — and discloses every statement:
What is our biggest default risk in procurement?
Supplier B — invoice #4471 is 34 days overdue, and it is the sole source for a critical part (assembly X). If it fails, the line stops in ~2 weeks. Recommendation: qualify alternative supplier A, prioritise payment.
ERPAccountingSupplier fileFramework agreement §4
5 Why this also holds up legally and organisationally.
For sensitive AI, the EU AI Act requires exactly what this architecture does anyway: complete logs, traceable documentation and the ability for a human to intervene.6 Because c:node backs every step with source and history, the decision can be audited afterwards — without you having to disclose the inner workings of the model. And because c:node runs on-prem or in the EU cloud, your procurement data stays in-house.
In short: The problem isn't a lack of data, but a lack of traceable connection. c:node establishes it — and turns scattered signals into a decision that comes early, is evidenced and can be audited.
Sources
Brintrup et al., “Towards knowledge graph reasoning for supply chain risk management using graph neural networks”, International Journal of Production Research, 2022. tandfonline.com…
NIST, Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1), 2024 — fabricated statements are inherent to the design. nvlpubs.nist.gov…
MEGA-RAG (Multi-Evidence RAG) — reduces fabricated statements by more than 40 %, but does not eliminate them. PMC12540348, 2025. ncbi.nlm.nih.gov…
Hitzler et al. (eds.), Handbook on Neuro-symbolic AI and Knowledge Graphs, IOS Press, 2025. iospress.nl…
GraphRAG-Bench — graph-based AI outperforms plain text search in multi-step reasoning. arXiv:2506.02404, 2025. arxiv.org…
European Union — EU AI Act, High-level Summary: logging, documentation, human oversight for high-risk AI. artificialintelligenceact.eu…
The sources support the principles (supply chain risk in the knowledge graph, limits of generative AI, neuro-symbolic approach, EU AI Act requirements). Names and figures in the live graph and the sample dialogue are simplified illustrations, not real cases.
Get started
Three ways to c:node.
Same product, same evidence – you choose where it runs.
Lexware2 invoices open
OutlookDelivery delay announced