AI procurement
negotiation at machine scale.

Buyer Team is an autonomous AI system that transforms every Purchase Requisition into a negotiated Purchase Order — covering classification, strategy, bidding, evaluation, and award with consistent governance.

Request early access See how it works
0%
Spend coverage
0%
Faster cycle time
0%
Cost savings per item
0×+
Steady-state ROI

The problem

Your procurement team can't negotiate everything.

Enterprise procurement organizations negotiate fewer than 20% of transactions. The other 80% — tail spend — gets zero negotiation effort, defaults to catalog pricing, and creates compliance gaps your team never sees.

80%
of transactions receive zero negotiation

Tail spend flows through without competitive sourcing, leaving savings on the table and governance gaps undetected. Manual cycles take weeks; policy enforcement is people-dependent and inconsistent.

100%
coverage — every requisition, negotiated

Buyer Team targets ≥ 80% autonomous coverage from day one, scaling to every transaction with consistent strategy, governance, and a complete audit trail per negotiation.

Weeks of sourcing cycles — invitations, follow-ups, clarifications, evaluations, and award decisions require sustained human bandwidth your team doesn't have for low-value items. Hours lost to admin, not strategy.

Spot bids close in < 24 hours. Competitive auctions close in ≤ 5 business days. Buyer Team handles supplier comms, bid evaluation, and PO assembly autonomously — for every item, at any volume.

Workflow

Orchestration before intelligence.

A deterministic DAG on AWS Step Functions provides the governance backbone. LLM-powered Strands A2A agents supply adaptive intelligence at each decision point — within guardrails.

01

Ingest & validate

Purchase Requisition received, items validated against catalog, delivery constraints applied.

Graph node
02

Kraljic classification

Items scored on profit impact x supply risk. Quadrant assigned automatically. Semantic cache eliminates redundant classification calls for known categories.

Graph nodeSemantic cache
03

Strategy routing

Quadrant maps to strategy: SPOT_BID, COMPETITIVE_AUCTION, PARTNERSHIP_RISK, or PARTNERSHIP_VALUE.

Graph router
04

Autonomous negotiation

Specialized A2A agent executes the strategy — sends invitations, manages rounds, requests clarifications, enforces supplier delivery gates.

A2A agent
05

Bid evaluation & award

Multi-constraint scoring: price, delivery, quality, ESG, supplier history. Auto-award below threshold; human approval gate above.

A2A agentHITL gate
06

PO assembly & comms

One Purchase Order per awarded supplier, with full OTEL audit trail. Suppliers notified. KPI metrics emitted to CloudWatch.

Graph node

Kraljic matrix — strategy routing

NON_CRITICAL

Automated spot bidding, single-shot, up to 200 concurrent bids. Optimized for high-volume throughput.

SLA: < 24 hrs · Auto-approved

LEVERAGE

Competitive multi-round auction. Top 5-8 suppliers by KPI. Convergence detection. Cost weight 40%.

SLA: 3-5 days · HITL above threshold

BOTTLENECK

Risk-managed negotiation. ESG & reliability weighted 40%, cost secondary. Volume guarantees + backup.

SLA: 1-2 weeks · HITL always

STRATEGIC

Value-based partnership engagement. Multi-round AI analysis. Relationship + ESG primary.

SLA: weeks · HITL always

 Step Functions orchestrator — execution trace
▶ PR -> ingest_validate OK 42ms +-> kraljic_classify OK cache_hit +-> strategy_route -> SPOT_BID +-> spot_bid_agent [A2A] +-> send_invitations 3 suppliers +-> collect_bids 3/3 received +-> bid_evaluation [A2A] +-> award_decision AUTO_APPROVED +-> issue_po PO-2026-0042 +-> DONE 18m 34s total

A typical day

While you were in meetings today...

Buyer Team runs in the background, handling every routine negotiation autonomously and queuing your strategic decisions — fully analysed, recommendation-ready.

47
Negotiations completed
$28,400
Savings captured
3
Awaiting your review
100%
Policy compliant

Real-world scenario — IT hardware refresh

IT Hardware Refresh — $280K budget LEVERAGE quadrant
Round 1 — 7 suppliers, prices $232K-$278KDay 1
Round 2 — Rank feedback only, no price leakageDay 2
Round 3 — Convergence detected, auction closesDay 3
Best bid: $218,000 $37,000 saved (14.5%)
8:02
AUTO-AWARDED

Janitorial supplies — 3 suppliers, best bid $1,840, $210 saved

9:15
ROUND 2

Packaging materials — 6 suppliers, avg price down 8%

10:30
AWAITS YOU

Steel components (critical) — AI recommendation ready, 2 bids evaluated

11:45
AUTO-AWARDED

Office furniture — 4 suppliers, best bid $8,920, $1,430 saved

13:00
FOLLOW-UP

Printer cartridges — 2 non-responsive suppliers nudged automatically

14:20
AWAITS YOU

Maintenance contract renewal — 3-year term, AI flagged escalation clause

15:50
AUTO-AWARDED

Cleaning products — 5 suppliers, $3,200 awarded, all ESG certified

17:00
AWAITS YOU

New strategic supplier — first order, approval required per policy

Financial impact

$1.5-3.8M in run-rate savings for a mid-market procurement org.

$1.5-3.8M

Annual run-rate savings

Conservative ($1.5M), base case ($2.6M), and optimistic ($3.8M) scenarios for a mid-market organisation with $50M annual spend. Year 1 trends ~30% below run-rate during deployment ramp ($1.05M conservative, $1.82M base case); full run-rate from Year 2.

$800K - $2.4M
Tail Spend Uplift
8-15% cost savings on the ~60% of spend that sits in the long tail — the high-volume, low-value transactions that account for 80%+ of your PRs but were never competitively sourced. Every tail item now goes through a competitive process that previously received zero negotiation effort.
$400K - $800K
Cycle Time Reduction
40%+ cycle time reduction (KR2.1). Approximately 2 FTE of sourcing overhead recovered as strategic capacity. Spot bids close in < 24 hours, freeing your team for high-value work.
$200K - $400K
Policy Compliance
100% governance compliance (KR4.1). Reduced policy exceptions, audit exposure, and remediation costs. Every negotiation passes through all policy guardrails before award, with OTEL audit trail completeness per run.
$100K - $200K
Supplier Data Quality
Unified supplier performance scores — quality, ESG, on-time delivery — drive better decisions. Reduced off-contract and maverick spend through consistent supplier screening.
85×+
Steady-state ROI on ongoing costs
< 1 mo
Payback from run-rate operation
Month 3
Full KPI baseline live

See the full financial model methodology →

Pricing

Transactional pricing. No seats, no minimums.

You pay for the work the platform processes — per PR line — plus a one-time onboarding engagement scoped to your integration. Usage and value scale together by construction.

Two-part model

Per PR line
$0.50
Minimum per PR
$1.00
One-time onboarding
$25K – $100K
Deployment time
3 – 6 weeks

Mid-market example · $50M spend · representative deployment

Year 1 total
≈ $55K
Year 2+ usage
$6K – $17K / yr
5-year investment
≈ $85K
5-year savings
$7M – $12M+
NO
Seat licenses
Unlimited users per tenant. Procurement, finance, requestors, approvers — everyone uses the platform, no per-seat charge.
NO
Platform-tier minimums
No starter / pro / enterprise floors. The per-PR rate applies from your first transaction, at any volume.
NO
Percent-of-spend backloading
We don't take a cut of negotiated savings. Catalog price, awarded price, and the delta are all yours to keep.
NO
Capacity commits
No buying capacity you'll need later, no monthly minimum, no platform floor beyond the per-PR rate itself.

Business outcomes

Measurable results, not aspirational claims.

Every KPI is instrumented end-to-end via OTEL spans and domain metrics — governance-compliant, tenant-scoped, audit-ready from day one.

80%+
Spend coverage rate — transaction volume processed autonomously, up from ~20%
40%
Sourcing cycle reduction vs. manual baseline (KR2.1)
8-15%
Cost savings per negotiated item vs. estimated or catalog price (KR3.1)
100%
Policy guardrail compliance — zero violations across all negotiations (KR4.1)
90%+
Award rate — negotiations completing with a valid, evaluated award
< 1 mo
Payback period from run-rate operation — onboarding investment recovered against transactional usage and realised savings

Architecture

Hybrid topology: deterministic graph + autonomous agents.

Level 1 is a governed DAG on AWS Step Functions. Level 2 deploys each negotiation agent as an independent AgentCore Runtime — Strands A2A agents that never call each other directly; every hop is mediated through DynamoDB state, dispatched by the orchestrator.

AgentCore services

01
Runtime
Each A2A agent as independent Runtime; Graph Orchestrator as primary entry point
02
Memory
Short-term: active negotiation session. Long-term: supplier history, past bid patterns (Mem0)
03
Gateway
Tenant Skills (MCP servers) for ERP/P2P, supplier portals, dataset ingestion, PO export
04
Identity
Inbound: JWT via Cognito. Outbound: OAuth 2.0 for supplier MCP servers
05
Observability
ADOT instrumentation, CloudWatch Transaction Search, custom procurement.* metrics
06
Evaluations
LLM-as-Judge, Ground Truth, code-based evaluators — CI/CD integrated
Purchase Requisition GRAPH ORCHESTRATOR Ingest & Validate Kraljic Classify Strategy Router Spot Bid A2A Agent Leverage A2A Agent Bottleneck A2A Agent Strategic A2A Agent Bid Evaluation Agent Human-in-the-Loop Gate Award + PO Assembly Purchase Orders

Integration

One Skill per tenant. Connects to everything you already use.

Buyer Team integrates with SAP, Oracle, and Coupa via AI Plugins. Each tenant gets exactly one isolated Skill as the single integration hub for all external interactions.

ERP
SAP S/4HANA · Oracle Fusion · Coupa
Master data flows inbound (categories, items, suppliers, PRs). Negotiated POs flow back out in JSON, XML, EDI, or CSV. Configurable field mappings per tenant.
MCP
Supplier comms — MCP-first, email fallback
Invitations, clarifications, award notifications all go via supplier MCP servers (OAuth 2.0, credentials in AgentCore Token Vault). Circuit breaker auto-falls back to SES email when unreachable. Full CommunicationLog audit trail.
ISO
Tenant isolation — zero cross-tenant data access
Defense-in-depth, no single point of trust. Cognito Pre-Token-Generation V3 Lambda normalises tenantId from two trusted bindings — federated IdP mapping for human users, authenticated per-tenant App Client for M2M. Gateway Request Interceptor overwrites tenant_id on every tools/call from the JWT claim. Per-request ABAC credentials via sts:TagSession + ${aws:PrincipalTag/tenant_id} scope IAM at the resource layer. Cedar policies and DynamoDB PK isolation complete the stack — a failure in any one layer cannot produce cross-tenant exposure.
6x
6 MCP servers inside each Skill
erp-connector, supplier-comm, s3-reader, s3-writer, dynamodb-domain, email. Transport: MCP synchronous, Kafka async, REST/Webhook, SDK (AWS) — declared in the Plugin manifest, resolved at runtime.

Transport channels

-- MCPSynchronous tool invocation via AgentCore Gateway
KafkaHigh-volume async events — AWS MSK, Confluent Cloud, or self-managed
RESTOutbound REST clients and inbound webhooks at AgentCore Gateway
SDKIAM-bound boto3 calls — bypasses Gateway; IAM role IS the registration
Buyer Team — Integration Architecture Structural diagram showing the Buyer Team AI Skill at the centre, surrounded by AI Plugins, connected to external systems (SAP, Oracle, Coupa, Suppliers, S3) via four transport types: MCP, Kafka, REST/Webhook API, and SDK. Integration Architecture — Skills, Plugins & Transport Buyer Team · External systems integration via AI Plugins · MCP · Kafka · REST/Webhook · SDK Buyer Team AI PROCUREMENT NEGOTIATION SAP S/4HANA ERP · SAP Joule / BTP MCP · Kafka · API Oracle Fusion ERP · OIC adapter API · Kafka Coupa P2P · Connect extension API · Kafka Suppliers MCP servers · SES email Bids · notifications Amazon S3 Dataset storage SDK (AWS) ▲ Plugin layer — transport declaration boundary ERP / P2P AI Plugins Declare transport · expose Skill tools · no logic · one Plugin per external system per tenant Supplier Plugin MCP primary · SES fallback S3 Plugin AWS SDK adapter Buyer Team — AI Skill Integration logic · schema mappings · tenant config · tool & event handlers · one Skill per tenant MCP servers (transport) erp-connector · supplier-comm dynamodb-domain · email s3-reader · s3-writer Kafka handlers (transport) Consumers · producers Schema validation · DLQ on failure API handlers (transport) REST clients · webhook receivers AgentCore Gateway inbound SDK handlers (transport) boto3 clients · IAM-bound Bypasses Gateway · no MCP server Domain model & DynamoDB (20 tables) Categories · Items · Suppliers · Purchase Requisitions · Negotiations · Bids · Purchase Orders · CommunicationLog Never directly accessible to external systems — all access through Skill tools only ▼ Async Kafka transport (all brokers) Kafka broker (tenant-declared) AWS MSK · Confluent Cloud · Self-managed Kafka Broker implementation resolved from tenant config at runtime Transport mechanisms ▪ MCP (Model Context Protocol) Synchronous tool invocation via AgentCore Gateway ▪ Kafka topics (async) Plugin declares topics + broker; Skill validates schema ▪ API / REST / Webhooks Outbound via erp-connector; inbound via Gateway ▪ SDK (AWS Service) IAM-bound boto3 calls; bypasses Gateway entirely Plugin delivery by ERP/P2P SAP S/4HANA SAP AI Plugin · SAP Build Joule Studio Registered in SAP BTP · bound to SAP Joule Oracle Fusion Oracle Integration Cloud adapter Bound to OIC AI Agent runtime Coupa Coupa Connect extension or webhook adapter Bound to Coupa Orchestration Kafka broker options AWS MSK Native IAM auth · VPC-native Glue Schema Registry Confluent Cloud API key auth · multi-cloud Confluent Schema Registry Self-managed Kafka SASL/TLS auth · tenant-operated Tenant schema registry ERP / P2P system External endpoint AI Plugin layer AI Skill (Buyer Team) AWS infrastructure

Security

Defense-in-depth. Six layers, no gaps.

Mapped to OWASP Agentic 2026 and MITRE ATLAS v5.4. Every negotiation runs inside a fully audited, tenant-isolated execution context. Immutable S3 audit trail with configurable retention (S3 Object Lock), enforced via OTEL trace completeness per negotiation.

01
Infrastructure Security
IAM least-privilege, VPC isolation, KMS encryption at rest, ECR image scanning
02
Identity & Authentication
AgentCore Identity — Cognito JWT inbound, OAuth 2.0 outbound to supplier MCP servers
03
Tool Access Control
Cedar + Gateway Request Interceptor — per-agent tool permissions enforced at the boundary; deterministic tenant_id injection on every tools/call request (REQ-S700)
04
Content Filtering
Bedrock Guardrails — prompt injection defense, PII detection, topic restriction
05
Behavioral Guardrails
8 Strands steering hooks distributed across agents, enforced at every tool-call boundary. BidConfidentialityGuard · BudgetCeilingGuard · AuctionIntegrityGuard · RiskAssessmentEnforcement · TCOEnforcementGuard · RelationshipHistoryEnforcement · EvaluationCompletenessGuard · WinnerDisclosureGuard
06
Application Security
Input sanitization, idempotency keys, immutable S3 audit trail (configurable retention, S3 Object Lock), OTEL trace completeness per negotiation

Governance

Your rules. Enforced on every purchase, automatically.

Set your policies once. Buyer Team applies them consistently — whether it's Monday morning or Friday at 4:55pm. No exceptions, no matter who is on shift.

$
Budget limits

Every negotiation runs within your approved budget ceiling. The AI cannot invite a supplier or accept a bid that exceeds your limit — no matter what the supplier proposes.

"No award above $50K without manager sign-off" → automatically enforced on every negotiation
✓
Approved supplier list

Only pre-approved, compliant suppliers receive invitations. Suspended or blocked vendors are excluded automatically before a single message is sent.

"Only invite suppliers with ESG score > 0.6" → enforced on every negotiation
ESG
ESG requirements

Your sustainability commitments aren't advisory. The AI screens every supplier and every bid against your ESG criteria before evaluation — a low-ESG bid is disadvantaged in scoring.

"No award to suppliers with pending compliance issues" → enforced on routine and strategic bids alike
🔒
Confidentiality guardrails

Competitor prices, your budget ceiling, and your negotiation strategy are never disclosed to suppliers — even in AI-drafted communications. Bid information is sanitized from every outgoing message automatically.

AuctionIntegrityGuard + BidConfidentialityGuard + WinnerDisclosureGuard active on every outbound supplier communication

Getting started

Three steps. Three to six weeks. Fully operational.

No coding required. No lengthy IT project. Start with your approved supplier list, configure your policies, and turn on Buyer Team for tail spend first.

01
Week 1
Connect your supplier list

Upload your approved supplier list. Buyer Team immediately knows which vendors to invite for each category and applies your existing supplier tiers.

Takes ~30 minutes
02
Weeks 1–2
Set your policies

Configure budget limits, ESG requirements, approval thresholds, and supplier preferences. All point-and-click — no coding needed. Your existing governance rules in minutes.

Point-and-click, no coding
03
Weeks 2–4
Start with tail spend

Turn on Buyer Team for your routine low-value purchases first. See savings appear in your dashboard. Expand to competitive and strategic categories on your own schedule.

Expand at your own pace
200+
Auto-negotiations Month 1
8–15%
Avg savings on Quick Bids
< 4h
Avg routine cycle time
100%
Policy compliance, day one

What changes

Less admin. More strategy.

Buyer Team handles everything a computer can do — freeing your team for decisions only people can make.

Before — Manual Procurement
  • Manually drafting RFQ emails to each supplier
  • Chasing non-respondents with follow-up emails
  • Building bid comparison spreadsheets from scratch
  • Manually checking bids for budget and policy compliance
  • Writing award notifications and rejection emails
  • Maintaining supplier data across multiple spreadsheets
  • Explaining why 80% of tail spend wasn't negotiated
  • Spending Fridays on admin instead of strategy
After — With Buyer Team
  • Review AI award recommendations (avg. 3 min each)
  • Approve or redirect — AI has done the full analysis
  • Monitor savings dashboard in real time
  • Focus on strategic supplier relationships
  • Set and adjust negotiation policies in minutes
  • Spend more time on market intelligence and planning
  • Confidently report 80%+ spend coverage to leadership
  • Leave on time on Fridays

Blog

Thinking in public.

Technical deep-dives from the engineers building Buyer Team, and AI procurement strategy perspectives.

About the Architect

Conceptualized, Built and Architected by one engineer.

Buyer Team is a production-grade demonstration of what a senior architect can ship end-to-end — from business concept to deployed multi-agent system on Amazon Bedrock AgentCore.

Gustavo Peixoto de Azevedo

Gustavo Peixoto de Azevedo

Hands-on AI Solution Architect — production-grade, event-driven agentic systems, from business concept to production and evolution.

25+ years architecting distributed systems. Designed and built Buyer Team end-to-end — agent orchestration on Amazon Bedrock AgentCore and the Strands Agents SDK, multi-tenant runtime, MCP integrations (SAP S/4HANA, Oracle Fusion, Coupa), plus the observability, evals, security mapping, IaC, and GitOps needed for production multi-tenant operation at scale.

Previously: Staff Engineer on streaming data platforms at Grupo SBF, Platform Engineer at Zwift (300K concurrent users on real-time leaderboards), Founding Team engineer at Grepr. M.Sc. Computer & Systems Engineering, UFRJ. Open-source contributor to awslabs/fullstack-solution-template-for-agentcore, aws-samples/sample-strands-agent-with-agentcore, spring-ai-community/spring-ai-agentcore, and SYSTRAN/faster-whisper among others.

inLinkedIn </>GitHub @Email Full bio →

Ready to negotiate
every requisition?

Join the early access program. We are onboarding enterprise procurement teams and technology partners now. Expected Month 1: 200+ automatic negotiations, 8-15% average savings, 100% policy compliance from day one.

Request access

No spam. We will reach out within 2 business days.