What is agentic AI for procurement?

Agentic AI for procurement means specialist AI agents perform research, cross-referencing, and verification autonomously across sanctions, commodity, FX, shipping, carbon, supplier, sentiment, air-freight, and compliance data sources, on documented cadences, with a full audit trail. A buyer submits a natural-language requirement and the system returns one ranked, evidence-backed recommendation in seconds — not a chat session, not a portal, but a defensible decision. WYRM Sentinel runs a roster of nine specialist agents; a minimum of seven run on every query, with the remaining two activating when the query pattern makes them material.

How is agentic AI different from a chatbot or copilot for procurement?

A chatbot answers one user prompt at a time from a static knowledge base; it cannot price live LME, FX, or freight at decision time. A copilot summarises the screen the buyer is currently on and drafts text in that single application context. Agentic AI runs specialist agents in parallel against external live data sources, fuses their verdicts through a regime-adaptive confidence ensemble, and produces a structured recommendation with source-linked evidence. Procurement decisions are multi-source, time-sensitive, and legally defensible — they need the third pattern.

How does WYRM Sentinel compute CBAM liability at decision time?

Sentinel's Carbon Agent computes CBAM liability synchronously inside every procurement query for in-scope goods (cement, iron, steel, aluminium, fertilisers, electricity, hydrogen). It pulls the current LME or reference price for the commodity, applies an embedded carbon factor appropriate to the supplier's production route (verified supplier-specific data where available, default factors otherwise — the choice is disclosed), credits any deemed carbon price already paid at origin, and applies the current UK CBAM certificate price to the net liability. The result appears as a line item in the landed-cost breakdown alongside product, freight, FX, and tariff. Embedded carbon factors come from Climatiq lifecycle data and country grid-carbon intensity.

Which sanctions lists does WYRM Sentinel screen against?

Sentinel's Sanctions Agent screens every counterparty and declared parent ownership against UK OFSI, US OFAC (SDN and sectoral), the EU consolidated list, and UN Security Council designations, sourced via OpenSanctions which consolidates 40+ global lists with daily updates. The Supplier Agent extends the screen up the parent-company chain via OpenCorporates' 200M+ legal-entity dataset to detect indirect exposure that direct counterparty screening would miss. Advisory-only findings are surfaced explicitly rather than silently blocking, so compliance teams can review before a hard gate is applied.

Is the WYRM Sentinel API publicly accessible?

Yes. The REST API is live at https://api.wyrm.ai/v1 with the OpenAPI specification at https://api.wyrm.ai/openapi.json and full reference documentation at https://wyrm.ai/docs/api. Authentication accepts either an X-API-Key header (issued from the Sentinel dashboard on the Pro tier or above) or a Supabase-issued Authorization: Bearer JWT. A Model Context Protocol (MCP) server is also available, exposing the same agentic engine to Claude, ChatGPT, Cursor, Cline, and any compliant MCP client. Rate limits and monthly decision quotas are tier-based: Starter 5,000 decisions/month, Pro 25,000, Enterprise unlimited.

How does WYRM Sentinel produce an audit trail for each decision?

Every procurement decision persists the agent outputs, their individual confidence signals, the ensemble weights used, and the final composite score into an immutable decisions table. The evidence trail is the primary deliverable — not a separate reporting step. A buyer asked to defend a recommendation six months later can open the original decision record and see which agent surfaced which signal, what source was queried, when it was queried, and how the verdicts were weighted. Audit log retention is seven years (aligned with UK Companies Act and HMRC requirements), extendable on Enterprise. This is the artefact format that satisfies Procurement Act 2023 defensibility requirements.

Where is data stored and processed?

United Kingdom by default for all customers. EU data residency is available on Enterprise plans for buyers with European operations. A UK-only processing path is offered for public-sector and regulated-industry workloads. Encryption is TLS 1.3 in transit and AES-256 at rest with 90-day key rotation. SOC 2 Type II, ISO 27001, and Cyber Essentials Plus are in progress; NHS DSPT is on the roadmap. The trust centre at https://wyrm.ai/trust publishes the current state of each certification as a living document.

Technical Overview

Agentic AI for Procurement

Sentinel deploys specialist AI agents that run parallel research across procurement data sources, normalise their findings, and return one ranked recommendation per query. Each agent is bounded to its domain; the orchestrator fuses outputs with explicit confidence weights, and every step is logged for audit.

Definition

What Is Agentic Procurement?

A system that performs a full procurement analysis autonomously, against a documented cadence, with human-reviewable output.

Most AI products in the procurement market today are reactive. A buyer types a prompt, the tool returns a generic answer. Chatbots wait for input. Copilots suggest completions. They are assistants that respond when prompted.

Agentic procurement is different. A specialist agent operates within defined polling and scoring schedules, produces outputs on a documented cadence, and logs every action for audit. The sanctions agent re-checks designations on every revision. The commodity agent re-prices on every market tick. The FX agent recomputes exposure on every ECB update.

The distinction matters because procurement decisions require continuous awareness of sanctions updates, FX movement, commodity prices, carbon factors, and supplier status. A question-answering interface is insufficient — the data changes between questions, and the audit trail depends on recording every intermediate state.

Comparison

Chatbot vs Copilot vs Agent

Attribute	Chatbot	Copilot	Agent
Initiation	User prompt	User action	Scheduled + event-driven
Persistence	Session-based	Session-based	Continuous, audit-logged
Planning	None	Suggestion only	Goal-directed, bounded
Multi-step execution	No	Limited	Yes, with checkpoints
Cross-domain reasoning	No	No	Via orchestrator ensemble
Proactive updates	No	No	On material change only

Architecture

The Specialist Agents

A roster of nine bounded domain agents. A minimum of seven run on every procurement query; the remaining two activate when the query pattern makes them material.

Sanctions Agent

Screens every counterparty and declared parent ownership against UK OFSI, US OFAC, EU, and UN consolidated lists via OpenSanctions. Runs on every list revision. Surfaces advisory-only findings alongside hard blocks so compliance teams can review rather than being silently gated.

Commodity Agent

Prices HS-coded commodities against LME, Pink Sheet, and reference sources. Produces forward-looking point estimates with volatility bands, not a single number. Regime classifier (stable / volatile / crisis) re-weights forecasts to prevailing conditions.

FX Agent

Uses ECB reference rates and forward curves to price FX exposure on multi-month procurement commitments. Tranche planner for committed base plus capped overage. Highlights pairs where the forward differential materially changes the decision.

Shipping Agent

Scores every candidate route against AIS lane telemetry and the status of the six global chokepoints (Hormuz, Red Sea, Malacca, English Channel, Panama, Gibraltar). Dark-ship watch surfaces vessels that deactivate AIS near sanctioned ports as a sanctions-evasion signal.

Air Freight Agent

Complements the Shipping Agent for time-critical cargo. Models airport throughput, lane capacity utilisation, and IATA spot rates by regional hub (PVG, DXB, FRA, MEM, LHR) to quantify when air freight is the economically correct mode versus sea.

Carbon Agent

Computes embedded lifecycle emissions using Climatiq factors and country grid-carbon intensity. Returns CBAM liability at the decision point, not as a spreadsheet afterwards. Scope-3 exposure reported at HS-code granularity.

Supplier Agent

Verifies supplier existence and traces ownership via OpenCorporates across 200M+ entities. Detects common ownership across declared-independent suppliers and capacity discrepancies against COMTRADE reported trade flows.

Sentiment Agent

Scores commodity-class sentiment using Claude Haiku NLP across GDELT event streams and curated RSS (Reuters, BBC, Al Jazeera). Class-aware clustering prevents an aluminium-specific headline from distorting an energy trade. Stale signals are attenuated toward neutral.

Compliance Agent

Applies CBAM, CPTPP, UK-EU TCA, and jurisdiction-specific procurement rules. Resolves HS codes to duty rates including preferential origin treatment. Flags dual-use export control exposure where relevant.

In Practice

Cross-Agent Correlation

A sanctions revision changes three baskets in parallel.

Sanctions Agent

OFAC adds the parent entity of a Tier-2 aluminium supplier to the SDN list. Revision detected within minutes of publication.

Supplier Agent

Cross-references parent ownership across the buyer's active supplier network. Flags three additional suppliers owned by the same parent, previously unlinked in buyer records.

Commodity + Carbon Agents

Re-score each affected basket assuming loss of the flagged origin. Landed-cost delta computed using alternative origins with current LME prices, FX forwards, and CBAM liability for the swap.

Orchestrator

Issues composite update: 'OFAC SDN addition invalidates 3 of 8 active basket origins. Recomputed ranking shifts Canada from 2nd to 1st; Turkey removed. Estimated landed-cost delta: +4.2%. CO₂ grade maintained (B). Full evidence trail at /decisions/{id}.'

Each agent saw one piece. The orchestrator produced the defensible procurement answer. Every step is logged in the audit trail.

Impact

Why This Matters

OpenSanctions consolidates 40+ global sanctions lists, updated daily. UN COMTRADE exposes over three billion bilateral trade records. The London Metal Exchange publishes hourly commodity prices. The ECB publishes reference rates and forward curves daily. Every CBAM-in- scope origin has an evolving lifecycle carbon profile. No human analyst can maintain live awareness across all of this for every active basket.

The viable approach is specialist agents that do the research and initial analysis, with procurement officers applying judgement to the flagged material changes. Not replacing buyers — augmenting them. An agent can re-check every sanctions update, FX tick, and supplier registry change. A procurement officer decides what to do about the ones that matter.

For the supply-chain-specific application of this architecture — risks covered, comparison against SCM suites and RPA, defensibility under CBAM and the Procurement Act — see the agentic AI for supply chain guide. For a deeper look at the terms used here, see the procurement glossary. For specific applications, see the use cases.

Defensible procurement decisions, in seconds.

Run your first agentic procurement decision today. Starter from £199 / month, Pro from £499 / month — every tier includes the full specialist-agent suite.