PROCUREMENT AGENT
Your procurement never sleeps.
340 suppliers watched. Continuously.
Ari monitors your entire supplier base in real time — detecting risk 44 hours before it becomes public, raising purchase orders autonomously, and negotiating price without being asked. Your procurement team focuses on strategy. Ari handles everything else.
Every capability below runs continuously, without anyone asking, day and night.
Most procurement teams find out about supplier problems after they happen. Ari reads the signals 44 hours before the problem reaches your inbox.
Five steps. Ari handles the first four autonomously. Only the genuinely uncertain decisions reach your team.
These are the documents and actions Ari produces — published directly into your ERP, sent to suppliers, and shared with your procurement team.
A pre-recorded simulation of Ari detecting a TSMC yield anomaly 44 hours before it became public and activating an alternative sourcing strategy autonomously. Book a session to see it run on your supplier network.
30 minutes. Manish runs Ari live using your supplier base and spend data. You see exactly which suppliers are at risk, what Ari would do, and the financial impact at your scale.
Configure your organisation for responses specific to your supplier network and spend profile.
340 suppliers across 47 categories and 23 countries in real time.3. POs in progress: 8. Next supplier sweep: 4 minutes.This paper presents Ari, the autonomous procurement agent within the VYUH Supply Chain Neural Network. Ari employs a Graph Neural Network (GNN) for multi-tier supplier risk intelligence, a game-theoretic autonomous negotiation engine, and a confidence-gated purchase order execution system to manage procurement across 340 suppliers in 47 spend categories. Operating at 91.5% autonomous decision rate, Ari achieves an average cost saving of 23.4% against budget and detects supply chain disruptions 44 hours before they appear in public news. This paper describes the complete technical architecture including the GNN supplier risk model, cascade risk propagation mechanism, autonomous negotiation protocol, confidence gating system, and integration with the VYUH agent network.
Enterprise procurement encompasses three structural challenges that conventional tools address inadequately. First, supplier risk visibility: multi-tier supply chains create opaque risk dependencies where a tier-3 raw material shortage can halt tier-1 production weeks later — yet conventional procurement systems monitor only direct (tier-1) suppliers. Second, decision latency: the cycle from demand signal to purchase order typically spans 48 to 72 hours through manual approval workflows, creating unnecessary risk exposure. Third, negotiation suboptimality: human negotiators, constrained by time and cognitive bandwidth, consistently leave value on the table relative to what systematic, data-driven negotiation approaches can achieve.
Ari addresses all three challenges through continuous multi-tier supplier monitoring via a Graph Neural Network, confidence-gated autonomous PO execution, and a game-theoretic negotiation engine that operates without human involvement at 91.5% of interactions.
Ari's architecture comprises five layers: (1) a multi-source intelligence ingestion pipeline, (2) a Graph Neural Network for supplier risk scoring and cascade propagation, (3) a supplier selection optimiser using risk-adjusted total cost of ownership, (4) a confidence-gated autonomous execution layer, and (5) a game-theoretic negotiation engine. The agent operates on a continuous real-time basis for risk monitoring, with procurement execution triggered by demand signals from Sara Jr (Demand Planning Agent) or autonomous threshold breaches.
The diagram below illustrates Ari's supplier risk classification framework — the two-dimensional matrix used to prioritise procurement actions across the supplier portfolio.
Ari models the supplier network as a directed heterogeneous graph G = (V, E) where nodes V represent suppliers, sub-components, raw materials, and geographic regions, and edges E represent supply relationships, material flows, and shared infrastructure dependencies. The graph encompasses tier-1, tier-2, and tier-3 supplier relationships — providing visibility into cascade risk that is invisible to conventional tier-1-only monitoring approaches.
Node features for each supplier include: financial health score (composite of credit rating, Z-score, payment behaviour), capacity utilisation rate, delivery performance (on-time-in-full over trailing 52 weeks), geographic risk index (geopolitical stability, natural disaster frequency, infrastructure reliability), and commodity price exposure (share of COGS attributable to monitored commodity indices).
The GNN uses a message-passing mechanism where risk signals propagate through the supply graph according to relationship weights. A stress event at a tier-3 raw material supplier propagates upstream through tier-2 processors to tier-1 manufacturers with a time-decay function calibrated to historical lead time data. This enables Ari to detect the precursor signals of a tier-1 disruption 44 hours before it surfaces — sufficient lead time to activate pre-qualified alternates before production impact.
Ari's negotiation engine models procurement negotiations as a two-player incomplete information game between Ari (buyer) and the supplier. The framework employs a Bayesian Nash Equilibrium approach to identify the optimal negotiation strategy given uncertainty about the supplier's cost structure and reservation price.
The supplier's reservation price is estimated using a proprietary model combining: publicly observable commodity input costs, peer supplier pricing from ERP purchase history, market benchmark data from industry pricing APIs, and financial distress signals that indicate a supplier's willingness to accept below-market pricing to maintain cash flow.
Ari generates counter-offers using a concession schedule calibrated to the estimated negotiation zone (the range between Ari's target price and the supplier's estimated reservation price). The opening position is set at 85th percentile of the estimated negotiation zone, with concessions structured to maximise expected value whilst maintaining supplier relationship quality as a secondary objective. In production deployment, this approach has achieved an average 23.4% cost reduction versus initial supplier quotes.
Ari applies a three-tier confidence gate analogous to Sara Jr's forecasting gate. The confidence score for a PO execution decision incorporates: supplier risk score (higher risk → lower confidence), demand signal confidence from Sara Jr, market pricing certainty (narrower commodity price range → higher confidence), and delivery commitment reliability (supplier's historical OTIF performance).
Decisions above 85% confidence execute autonomously. Between 65% and 85%, the recommendation is presented for human review with a 4-hour default approval window. Below 65%, a full escalation briefing is produced and procurement team approval is required before the PO is raised.
Ari receives procurement trigger signals from Sara Jr (demand-driven procurement needs) and risk signals from Tom (logistics lane disruptions affecting supplier access). Ari publishes PO confirmations to Cho (Inventory Agent) for expected receipt planning and delivery risk alerts to Tom for freight pre-booking. This bidirectional integration ensures procurement decisions are made with full visibility of logistics constraints and inventory positions.