01Knowledge
The Knowledge Graph
Every entity. Every relationship. One canonical truth.
Knowledge Graph + LAMThe substrate
Astral. The knowledge graph powering the world's finance OS.
Every supplier, every contract, every invoice, every payment, every entity across every system you run — unified into one canonical graph. AI agents reason on it. Large Action Models execute through it. This is what makes Flowie a finance OS, not a finance assistant.
600M+
Entities indexed
Every registered company on Earth + every internal entity across customer ERPs, normalized into one graph.
<5s
Sync latency
From source system event to canonical entity update in the graph.
30+
ERPs unified
SAP S/4HANA, Sage X3, Oracle NetSuite, Microsoft Dynamics, Odoo, Pennylane, Cegid — all into one substrate.
0
Schema migrations
Astral evolves with your data. No DBA project to add a new entity type.
The substrate
Astral in 90 seconds.
Modern finance and procurement run on dozens of systems. Five ERPs across business units. Three banking interfaces. A treasury platform. A spend management tool. Two CRMs. A dozen partner portals. Each one holds a slice of the truth. None of them talks to the others. Reconciliation is a quarterly archaeology project.
Astral is the substrate that fixes this. A typed knowledge graph that ingests from every system you run, resolves entities across systems (one supplier, one ID, one truth), exposes the canonical graph to AI agents through a query layer, and lets those agents take real-world actions through a typed action grammar.
Three layers stack on each other: Knowledge (the graph itself), Reasoning (LLMs grounded in the graph), and Action (Large Action Models that execute typed operations with governance, preview, and audit). Together they make Flowie a finance OS — not a chatbot, not an assistant, but the operating system on which finance and procurement workflows actually run.
Architecture
Three layers. One platform.
Knowledge → Reasoning → Action. Each layer builds on the one beneath. Together they turn AI from a probabilistic chatter into a deterministic, auditable operator on your finance and procurement stack.
01Knowledge
The Knowledge Graph
Every entity. Every relationship. One canonical truth.
Astral's foundation is a typed knowledge graph — entities (companies, persons, contracts, invoices, payments, products, accounts, regulations) connected by typed relationships (signed_by, issued_to, paid_via, supplies, owned_by, reports_to, governs). Built automatically from every connected system. Resolved canonically across systems. Queried by every agent and every workflow.
Capabilities
- Entity resolution: "ACME Corp" / "Acme Corporation" / "Acme Inc." → one canonical Astral ID
- Real-time bidirectional sync from 30+ ERPs, banks, partner portals
- Versioned, audit-trailed — every state change traceable to source event
- Self-evolving schema: new entity types absorbed without migrations
Concrete example
Your books say "Acme Corp · supplier · €450K outstanding" in SAP, "Acme Corporation · client · €1.2M signed" in Salesforce, "AcmeCorp Ltd · counterparty · 3 active POs" in your old Sage. Astral resolves all three to a single entity with both supplier and client roles, total exposure €1.65M, 3 active POs, full transaction history.
02Reasoning
The Reasoning Layer
LLMs ground their reasoning in your real, structured data.
03Action
The Action Layer (LAM)
Beyond reading. Beyond reasoning. Astral takes action.
Layer 01 deep-dive
How the knowledge graph actually works.
Astral is not an ontology stored in RDF. It is not a SPARQL endpoint. It is an engineered, AI-native graph designed for real-time queries by autonomous agents in production finance workflows. Six properties make this work.
01
Typed entities
Every node carries a type (Company / Person / Contract / Invoice / Payment / Product / Account / Regulation / Tax_Code) with a strongly-typed schema. No raw blobs, no untyped JSON. Reasoning stays grounded in semantics.
02
Typed relationships
Edges are typed with cardinality and direction (signed_by:1→1, issued_to:1→1, supplies:N→N, governs:1→N). Graph traversal is deterministic. Path queries ("all invoices from suppliers governed by FR-PA") run sub-100ms.
03
Entity resolution
Cross-system identity is the hardest problem. Astral combines deterministic keys (VAT ID, DUNS, IBAN), fuzzy matching (Jaro-Winkler on names, Levenshtein on addresses), and ML-assisted deduplication. Confidence-scored. Human-in-loop for low-confidence merges.
04
Versioned state
Every entity carries a version chain. Yesterday's PO, today's amended PO, tomorrow's cancelled PO — all queryable as point-in-time snapshots. Audit trails are first-class. Compliance forensics is one query, not a 6-day archaeology project.
05
Self-evolving schema
New entity type detected (ESG_Score, Carbon_Footprint, Anti_Bribery_Certification)? Astral adds it to the schema and back-fills without a migration. Your business adds a dimension; the graph absorbs it. No DBA project, no downtime.
06
Provenance + lineage
Every fact in Astral links back to its source: which system, which event, which timestamp, which user. "Why does Astral think Acme Corp's address is 12 rue de Rivoli?" → traceable to ERP event ID, source field, sync timestamp, confidence score.
Layer 03 deep-dive
Beyond LLMs. The shift to Large Action Models.
For two years the AI conversation has been about LLMs predicting text. The next two are about LAMs predicting and executing actions. This shift matters most where actions have real consequences — payments, approvals, regulatory filings. Finance and procurement are the canonical LAM domain.
| Dimension | LLM (text predictor) | LAM (action predictor) |
|---|---|---|
| Output | Predicts the next text token | Predicts and executes the next action |
| Modality | Text in → text out | Context in → action committed (with side-effects) |
| Grounding | Trained on internet text | Bound to a typed action grammar + state |
| Reversibility | Output is text — no real-world consequence | Actions have consequences — reversibility is engineered |
| Governance | Prompt-level guardrails | Policy engine, dry-run preview, autonomy thresholds, audit trail |
| Examples | GPT-4, Claude, Gemini, Mistral, Llama | Adept ACT-1, Anthropic Computer Use, Salesforce xLAM, Astral Action Layer |
Why LAMs matter for finance + procurement
A finance team's day is full of decisions that translate to actions: approve this PO, post this journal entry, dispatch this payment, register this supplier, file this VAT return, dispute this invoice. Every one of these is an action with real-world consequences and traceability requirements. An LLM that says "you should approve this PO" is not enough — you still have to click. A LAM that can actually approve it (with policy, preview, and audit) closes the loop.
Astral's Action Layer is a domain-specific LAM. Bound to a typed action grammar (~40 verbs covering the full P2P / O2C lifecycle). Wrapped in a policy engine (autonomy thresholds per amount, per category, per role). Backed by dry-run preview. Logged in append-only audit. Reversible where business logic permits.
The result: agents can take action on behalf of finance and procurement teams with the same level of trust as a junior team member — but at machine speed and with full traceability.
The frontier
Toward a world model for finance.
World models are the next frontier beyond LAMs. The leading research direction — Yann LeCun's JEPA (Joint Embedding Predictive Architecture, 2023+) at Meta — proposes AI systems that build a comprehensive internal representation of how a domain works, and use it to simulate consequences before they happen.
For finance, this matters more than for almost any other domain. A world model would not just say “approve this PO.” It would simulate: “If you approve this PO at this stage, here is the impact on budget reservation, supplier ledger, period close, audit trail, cash forecast, and compliance posture — all 18 hours into the future, across all 5 of your ERPs.” Decisions become forecasted, not reactive.
Astral’s typed graph + dry-run action layer is the foundation of a finance-domain world model. The work ahead: predictive simulation across the full state, multi-step rollout previews, counterfactual analysis (“what would have happened if we’d not approved this PO last quarter?”), and policy reasoning (“would this action have triggered our compliance rules?”). We are advancing in this direction. Astral makes the path possible.
We are not claiming to be a world model today. We are claiming the substrate we have built is the foundation that finance world models will run on — both ours and others'. Open MCP. Open APIs. Typed graph + typed actions. The infrastructure shifts; the moat is the data.
Why Flowie
Why Flowie is uniquely positioned to build the finance OS.
A knowledge graph is only as good as the data feeding it. A LAM is only as useful as its action authority. A world model is only as accurate as the substrate it predicts on. On all three dimensions, Flowie is positioned where no other vendor is.
Multi-tenant network density
Astral spans 600M+ companies. A single-tenant CRM or ERP knowledge graph sees only its customer's data. Astral sees the supplier-of-the-supplier, the upstream parent, the cross-tenant fraud signal. The graph is the moat.
Real-time vs warehouse latency
Data warehouses snapshot at midnight. Astral syncs in under 5 seconds. When an agent reasons about "current outstanding balance with supplier X", the answer is live — not 18 hours stale.
Certified action authority
Flowie holds PA/PDP certification (France), Peppol Access Point status, ISO 27001:2022, payment-issuer authorizations. Astral's actions are not simulated — they are legally authorized, regulator-recognized executions.
Network data improves entity resolution
When 50,000 customers are observing the same supplier across their stacks, Astral's resolution accuracy converges on near-perfect. Single-tenant graphs have no chance: they only see one slice of one entity.
Built EN 16931 native, not retrofitted
Every transaction enters Astral as structured data — not OCR'd PDFs, not unstructured emails. The graph never has to guess. Reasoning stays deterministic.
MCP-compatible from day one
Bring your Claude. Bring your ChatGPT. Bring your internal copilot. Astral exposes itself as a first-class MCP data source. The graph is open infrastructure, not a walled garden.
Asked & answered
Architects ask us this.
References: Yann LeCun et al., “A Path Towards Autonomous Machine Intelligence” (2022) · Adept ACT-1 (2023) · Anthropic Computer Use (Oct 2024) · OpenAI Operator (Jan 2025) · Microsoft Research LAM-V · Salesforce xLAM family.
See Astral on your stack.
A 30-min demo. Bring your ERPs, your CRMs, your banking systems, your partner portals. Watch them unify into one canonical graph in real time. Then watch our agents reason and act on it.
ISO 27001
GDPR
CyberVadis
PA Certified
Peppol