North star
Decision OS, not chatbot
Enterprise Cognitive OS
Introduction · next-generation decision operating system
bp Sphere Next
Enterprise Decision Runtime
2026-2030 roadmap
bp Sphere Next: Enterprise Cognitive Operating System
bp Sphere should evolve beyond enterprise memory into a model-independent Decision OS: persistent context, knowledge graph, policy-as-code, decision memory, skill marketplace, multi-agent orchestration, evidence, simulation, learning, and controlled execution. The strategic asset is BP's decision intelligence, not any single LLM or hyperscaler.
Strategic asset
BP policies, evidence, memory, outcomes
Architecture stance
Model-independent and sovereign
Roadmap shape
16 pillars, 12 executive PRDs
North Star: from AI application to Enterprise Cognitive Operating System
The next 5-10 years are not only about better models. The enterprise advantage comes from persistent, learning decision systems that retain proprietary context, policies, evidence, approvals, outcomes, and expertise across model generations.
Enterprise inputs
ERP, SAP, Ariba, Databricks, Foundry, documents, email, Teams, IoT, market data
SAP and Ariba
Databricks and UDP
Foundry and OneData
Documents and SOPs
Teams and Outlook
Market and IoT data
↓
Enterprise Cognitive OS
bp Sphere Decision Operating System
Context layer
Knowledge graph
Policy runtime
Decision runtime
Skill fabric
Agent fabric
Learning runtime
Evidence runtime
Simulation runtime
Memory runtime
Human collaboration
Controlled execution
↓
Model-independent execution
Model router chooses the right intelligence for the decision
OpenAI
Claude
Gemini
Azure AI
AWS Bedrock
Ollama / local
Design rule: BP should be able to replace models, clouds, and infrastructure without losing its ontology, policies, evidence, replay, decision memory, human expertise, or learning history.
16 pillars for the future-proof platform
These pillars turn memory architecture into a broader enterprise decision runtime: context, graph, policies, agents, evidence, simulation, learning, collaboration, optimization, and execution.
1. Become Model Independent
Sphere routes the business problem through a model gateway so GPT, Claude, Gemini, Bedrock, Azure, Ollama, and future models are replaceable.
model router
2. Decision Memory, Not Chat Memory
The durable record is the decision: evidence, policy, assumptions, approvals, confidence, outcome, value, and learning.
decision IP
3. Continuous Organizational Learning
Human corrections, overrides, approvals, and outcomes become governed learning signals rather than lost case history.
enterprise RL
4. Enterprise Knowledge Graph
People, processes, suppliers, assets, invoices, contracts, controls, risks, agents, evidence, and policies are connected.
semantic foundation
5. Skill Marketplace
Reusable skills replace monolithic agents: duplicate detection, tax validation, FX analysis, journal validation, supplier risk, and more.
composable skills
6. Multi-Agent Collaboration
Coordinator, domain agents, compliance, risk, tax, treasury, optimization, and human reviewers operate under explicit contracts.
governed fabric
7. Cognitive Digital Twin
Finance becomes simulatable: payment terms, cash, FX, supplier risk, working capital, discounts, controls, and outcomes.
decision twin
8. Event Native Architecture
Invoice, payment, journal, policy, and evidence events wake the runtime; polling and batch jobs become fallback paths.
event first
9. Sovereign AI
Sphere can run on Azure, AWS, private cloud, local compute, edge, or air-gapped environments without losing enterprise intelligence.
portable runtime
10. Policy-as-Code
Policies move from PDFs to executable DSL, tests, evidence binding, versioning, replay, and audit.
policy compiler
11. Autonomous Improvement
The runtime measures accuracy, latency, cost, overrides, hallucination risk, policy conflicts, and SAP calls, then tunes routing.
self optimization
12. Universal Evidence Layer
Every recommendation carries source, confidence, reasoning summary, policy, replay, lineage, and authority boundary.
trust layer
13. Token Economy
Optimization is cost per business decision: cache, route to smaller models, batch, use local inference, and prefer deterministic execution.
decision FinOps
14. Human Expertise Capture
Analyst rationale, corrections, exceptions, and patterns become enterprise memory before expert knowledge leaves the organization.
judgment capture
15. Cross-Enterprise Learning
Learning from P2P can update supplier risk, treasury forecasts, FP&A cash views, procurement actions, and tax risk.
cross-domain memory
16. Decision Operating System
Sphere becomes the governed context, graph, policy, decision, skill, agent, learning, evidence, simulation, memory, and execution layer.
Decision OS
What changes compared with a memory-centered slide
Graph memory, replay, reinforcement learning, and privacy-preserving memory are important, but they are subsystems. The durable platform vision is a Decision OS where memory is one part of a governed runtime.
Memory architecture alone
- Graph memory
- Replay records
- Privacy-preserving storage
- Learning history
Enterprise Cognitive OS
- Enterprise context and ontology
- Knowledge graph and decision memory
- Policy intelligence and authority runtime
- Multi-agent orchestration and skill fabric
- Learning, simulation, evidence, replay, model routing, collaboration, and optimization
12 executive-grade PRDs to define bp Sphere Next
Each PRD should be independently implementable while sharing platform contracts for identity, ontology, events, policy, evidence, replay, observability, FinOps, security, and human-in-the-loop execution.
| PRD | Purpose | Roadmap | Core deliverables |
|---|---|---|---|
| 1. Enterprise Knowledge Graph | Semantic backbone across SAP, Ariba, data platforms, documents, policies, suppliers, assets, controls, risks, and users. | 2026 foundation | Unified ontology, graph APIs, identity resolution, semantic context assembly. |
| 2. Enterprise Decision Memory | Persistent organizational memory beyond chat history. | 2026 foundation | Episodic, semantic, and procedural decision memory with aging, confidence, replay, and retrieval. |
| 3. Continuous Learning Runtime | Every analyst action improves future recommendations under governance. | 2026-2027 foundation | Override capture, rationale capture, pattern extraction, supervised promotion, evaluation, and feedback loops. |
| 4. Policy Intelligence Runtime 2.0 | Policies become executable decision logic. | 2026 foundation | Policy compiler, DSL, simulation, conflict detection, test suites, versioning, replay, and regulatory mapping. |
| 5. Skill Marketplace & Skill Fabric | Reusable enterprise skills replace one-off agents. | 2026-2027 agent platform | Skill lifecycle, certification, versioning, dependency management, marketplace, testing, and governance. |
| 6. Multi-Agent Collaboration Fabric | Specialized agents collaborate safely under explicit delegation and conflict rules. | 2027 agent platform | Planner, coordinator, specialists, negotiation, escalation, execution contracts, and human boundary. |
| 7. Enterprise Decision Twin & Simulation Engine | Finance decisions can be simulated before execution. | 2027 intelligence | Scenario simulation, Monte Carlo, policy impact, supplier impact, cash impact, FX impact, and forecast impact. |
| 8. Model Orchestration & AI Gateway | Vendor-independent AI runtime. | 2027 agent platform | Model routing, cost and latency optimization, fallback, prompt adaptation, local inference, Azure, AWS, OpenAI, Anthropic, Ollama. |
| 9. Evidence Intelligence Fabric | Universal trust, lineage, and explainability. | 2027 intelligence | Lineage, confidence, source ranking, provenance, contradiction detection, evidence scoring, replay, and audit packages. |
| 10. Autonomous Optimization Runtime | Sphere improves its own prompts, routing, caching, tools, memory, token usage, latency, and GPU scheduling. | 2028 optimization | Continuous measurement, change proposals, guarded deployment, champion/challenger, and rollback. |
| 11. Enterprise Cognitive Operating System | Master architecture connecting runtime, memory, graph, policies, agents, events, security, governance, UX, APIs, and deployment. | 2028 platform | System-wide contracts, reference architecture, operating model, governance model, and platform roadmap. |
| 12. Enterprise Decision Marketplace | Reusable decision services across finance, supply chain, HR, legal, operations, trading, retail, and sustainability. | 2029-2030 enterprise | Decision APIs, certification, usage economics, ownership, deployment, and marketplace governance. |
Recommended implementation sequence
The sequence starts with semantic and governance foundations, then scales into composable agents, simulation, optimization, and marketplace-based decision services.
| Phase | Capabilities | Outcome |
|---|---|---|
| Phase 1 - Enterprise Foundation | Enterprise Knowledge Graph, Enterprise Decision Memory, Continuous Learning Runtime, Policy Intelligence Runtime 2.0. | Establish the semantic, memory, learning, and policy substrate. |
| Phase 2 - Agent Platform | Skill Marketplace, Multi-Agent Collaboration Fabric, Model Orchestration Gateway. | Make agents composable, governable, measurable, and model independent. |
| Phase 3 - Intelligence | Decision Twin, Evidence Intelligence Fabric, Autonomous Optimization Runtime. | Move from recommendation to simulation, proof, and continuous runtime improvement. |
| Phase 4 - Enterprise Platform | Enterprise Cognitive Operating System and Enterprise Decision Marketplace. | Expose reusable decision services across the enterprise. |
2026
Enterprise foundation
Knowledge graph, decision memory, learning runtime, policy compiler.
2027
Agent platform
Skill marketplace, multi-agent fabric, model gateway, decision twin.
2028-2030
Decision OS
Autonomous optimization, enterprise memory fabric, decision marketplace, cognitive operating system.
Cross-cutting requirements for every PRD
These requirements keep the roadmap credible for architects: not a collection of slides, but a platform contract that can be engineered, tested, governed, and operated.
| Requirement | How it applies | Reason |
|---|---|---|
| Event-driven architecture | All PRDs define events, subscriptions, replay, retention, and source freshness. | Prevents batch-only AI and enables real-time decisioning. |
| Zero Trust execution | Identity, role, policy, source, action, and data boundary are checked at runtime. | Keeps agentic execution safe. |
| Human approval boundaries | Every PRD declares inform, recommend, prepare, execute-with-approval, and autonomous authority levels. | Prevents uncontrolled financial action. |
| Decision replay | Inputs, policies, evidence, prompts, model/tool calls, human action, and outcome are reconstructable. | Creates auditability and recoverability. |
| Cost per decision | FinOps tracks model, token, tool, SAP call, GPU, and latency cost by business decision. | Optimizes value rather than raw token use. |
| Multi-cloud and sovereign deployment | Azure, AWS, local, private, and air-gapped deployment paths are treated as first-class. | Avoids hyperscaler and model lock-in. |
| Deterministic fallback | Policy gates, matching, validation, and controls have non-LLM execution paths where required. | Reduces hallucination and runtime fragility. |
| Continuous evaluation | Prompt, skill, model, policy, and agent changes are evaluated against BP outcomes. | Makes improvement measurable and governable. |
| Data residency and governance | Memory, graph, evidence, learning, and replay artifacts carry classification, retention, and access policy. | Keeps enterprise memory usable and safe. |
Executable runtime APIs
The architecture is backed by an executable contract layer that can be called by mission services, tests, and platform validation scripts.
Contract and audit
GET /api/enterprise-cognitive-os/contractGET /api/enterprise-cognitive-os/readinessGET /api/enterprise-cognitive-os/audit-scorecardGET /api/enterprise-cognitive-os/architecture-maturity-matrixGET /api/enterprise-cognitive-os/master-audit-v1
Decision and learning
POST /api/enterprise-cognitive-os/decision-memoryPOST /api/enterprise-cognitive-os/decision-memory/persistPOST /api/enterprise-cognitive-os/learning-signal
Routing and simulation
POST /api/enterprise-cognitive-os/model-routePOST /api/enterprise-cognitive-os/multi-agent-planPOST /api/enterprise-cognitive-os/policy-as-code/compilePOST /api/enterprise-cognitive-os/policy-as-code/packPOST /api/enterprise-cognitive-os/decision-twin/simulate
Context and optimization
POST /api/enterprise-cognitive-os/graph-contextPOST /api/enterprise-cognitive-os/context/unifiedPOST /api/enterprise-cognitive-os/evidence-envelopePOST /api/enterprise-cognitive-os/event-fabric/validatePOST /api/enterprise-cognitive-os/optimization-telemetry
Memory and prompt governance
POST /api/enterprise-cognitive-os/decision-memory/retrievePOST /api/enterprise-cognitive-os/learning/proposalsPOST /api/enterprise-cognitive-os/prompt-registry/packagePOST /api/enterprise-cognitive-os/prompt-registry/evaluate
Strategic positioning
Not an AI platform
bp Sphere is not defined by the model. It is defined by BP's decision context, policies, evidence, controls, approvals, outcomes, and learning loops.
Not a chatbot
Chat is one interface. The durable record is the enterprise decision: why it happened, what evidence supported it, who approved it, and what outcome followed.
Not a hyperscaler architecture
Azure, AWS, OpenAI, Bedrock, and local models are execution options. BP-owned intelligence remains portable and sovereign.
Positioning line. bp Sphere Next is a governed, source-grounded, model-independent Enterprise Cognitive Operating System that turns BP's policies, evidence, expertise, outcomes, and decisions into a compounding strategic asset.