Applications, guidelines and supporting documents are brought together by hand.
The application arrives by mailbox, the supporting documents as attachments, the guideline sits on the intranet and state-aid law in Brussels. Every review starts with gathering it all.
The question
You ask Nora – in one sentence.
Instead of building reports, you simply ask. Nora 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
Nora prepares – you approve.
Nora prepares the next step. It is only carried out once you approve it.
Hi, I'm Nora – your AI agent for the public sector. 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
OutlookApplication with 6 attachments
SharePointDigitalisation funding guideline
c:node GraphCommercial register
c:node GraphDe minimis Regulation
@Nora, is the application from Weber Metall GmbH eligible for funding?
I'm checking all sources:
Read the application and 6 attachments
Checked the guideline's funding requirements
Reconciled company size with the commercial register
Checked state-aid ceilings (de minimis)
Eligible – one document is missing.
Applicant is an SME per commercial register and annual accounts
Project fits the guideline's funding purpose
De minimis declaration missing – required before approval
Application p. 2Funding guideline no. 4Commercial registerRegulation (EU) 2023/2831
I've prepared the next step:
Draft · Request for de minimis declaration
Letter to Weber Metall GmbH with a 14-day deadline, justified with citation.
ApproveEditNothing goes out without approval
Example case with sample data.
c:node Research · Public administration
Why funding decisions take too long — and how traceable AI speeds them up.
Reading time ~5 minBasis 6 sources, verifiedTopic Funding & decisionsApproach source-backed · on-prem possible
Fig. 0 · Public administration decides on the basis of law, budget and the case file — every decision must remain traceable.
Summary
Funding and administrative decisions are slow because the relevant facts are scattered across registers, budget plan, guidelines and case files — and because every decision must remain auditable. General AI speeds things up, but it is not permissible in a high-risk context, because it cannot make its statements traceable. This article shows, with sources, why that is — and how c:node speeds up decisions: it connects case file, registers and law and discloses every citation.
1 The problem: deciding here means making it traceable — and that takes time.
Whether a funding programme is extended, an application approved or a condition imposed depends on many facts: target achievement, budget coverage, the applicable guideline, the previous year's history, running deadlines. These facts sit in separate systems and files. Before a decision may be taken, every point has to be looked up and made traceable with its legal or file basis — otherwise the decision won't withstand review. It's exactly this gathering that is the bottleneck, not the decision itself.
2 Why generic AI is ruled out here.
An AI chatbot would be fast — but it guesses. 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.3 For sovereign decisions, that is disqualifying: the EU AI Act classifies AI in public administration as sensitive and requires complete logs, documentation and human oversight.1 An AI that can't bind its statement to a citation doesn't meet that — no matter how fast it is.
3 The approach: decide faster without losing auditability.
This very conflict of goals can be resolved. Current research shows that AI-assisted grant evaluation can be made auditable by keeping every partial statement traceable to its origin (provenance) — the decision becomes faster and still remains traceable afterwards.2 c:node implements this: the c:node Graph reasons over a knowledge graph built from your real sources — the combination of pattern recognition and firmly checked rules that research calls neuro-symbolic AI.4 The result is not a quick guess, but a traceable recommendation.
:nodeGraph— Funding / extensionlive
Source (API)EntityRecommendationRule / evidence
Nodes active · edges fire in real timeisolated per authority · grows with every piece of evidence
Fig. 1 · Register, budget, guideline and case file connect into a recommendation. c:node fires the path to the decision — every step stays bound to its citation.
4 In practice: a scattered case file becomes a traceable proposal.
The same question that today costs days of file work, c:node answers in seconds — and discloses every statement:
Should we extend programme P-12?
Yes — with a condition. Target achievement is 87 % (above the guideline's threshold), funding is covered in the budget plan. However, review rule §7 requires an updated proof of use by the deadline of 30 Sep. Recommendation: extend, add the condition “proof Q3”.
RegisterBudget planGuideline §7Case file / previous year
5 Why this also holds up legally and organisationally.
Because c:node backs every step with citation and history, the decision can be audited afterwards — by the court of auditors, supervisory authority or court — without the inner workings of the model having to be disclosed.1 And because c:node runs on-prem or in the EU cloud, sensitive citizen and case data stays in the EU or in your own data centre. Data protection is built in, not bolted on.
In short: The bottleneck isn't the decision, but the traceability. c:node establishes it automatically — and makes public administration faster without losing traceability.
Sources
European Union — EU AI Act, High-level Summary: AI in public administration as high-risk — logging, documentation, human oversight mandatory. artificialintelligenceact.eu…
Auditable AI-assisted grant evaluation via provenance tracking — decisions remain auditable without disclosing the model. arXiv:2604.25200. arxiv.org…
NIST, Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1), 2024 — fabricated statements are inherent to the design. nvlpubs.nist.gov…
Hitzler et al. (eds.), Handbook on Neuro-symbolic AI and Knowledge Graphs, IOS Press, 2025. iospress.nl…
MEGA-RAG (Multi-Evidence RAG) — retrieval reduces fabricated statements by more than 40 %, but does not eliminate them. PMC12540348, 2025. ncbi.nlm.nih.gov…
GraphRAG-Bench — graph-based AI outperforms plain text search in multi-step reasoning. arXiv:2506.02404, 2025. arxiv.org…
The sources support the principles (EU AI Act requirements for high-risk AI, auditable AI evaluation via provenance, limits of generative AI, neuro-symbolic approach). 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.
OutlookApplication with 6 attachments