Full Overlap
2 / 8
delegate to the agent platform
Partial / Minimal
2 / 8
Cockpit grounds the answer
No Overlap
4 / 8
Cockpit-unique capability
Recommendation
Delegate
do not duplicate agent strengths
Capability Overlap Matrix
CPM AI Cockpit vs. Claude Enterprise vs. Coworker.ai
| Cockpit capability | Claude Enterprise | Coworker.ai | Verdict | Assessment |
|---|---|---|---|---|
| Chat / Q&A over documents | Yes — native strength | Partial (via OM1 context) | Some overlap | Cockpit's chat layer is grounded in live spend and supplier data, not general documents |
| Cross-module data joins (GraphQL, REST) | No — no ERP/finance connectors natively | No — connectors are Jira/Slack/CRM, not SAP/Ariba | No overlap | This is the core gap the Cockpit closes |
| Read-only adapters to 30+ finance / register systems | None | None | No overlap | Neither platform has finance-grade system-of-record connectors (Ariba, S/4HANA, DED, D&B) |
| Monte Carlo, VaR/CVaR, FX and commodity shock simulation | No — can compute if manually fed data, no simulation engine | No — task/workflow orchestrator, not a quant engine | No overlap | Proprietary CPM AI engine is unique |
| 19-chain causal / scenario pathway modelling | No — can narrate a scenario, no structured causal engine | No | No overlap | No driver library outside the Cockpit |
| Governance: KPI dictionary, source lineage, data-vintage stamping | No | Partial — OM1 context memory, no KPI governance | Minimal overlap | Cockpit remains the governed source of truth |
| Task automation / drafting emails / following up | Yes | Yes — core strength | Full overlap | Exactly what Coworker.ai already does well — delegate, do not rebuild |
| General business writing, summarisation, coding help | Yes | Partial | Full overlap | Claude Enterprise's bread and butter — delegate, do not rebuild |
Coverage by Capability
2 = native capability · 1 = partial · 0 = none
Where Full Overlap Exists
Two capabilities — delegate rather than rebuild
Full overlap · Coworker.ai
Task automation, email drafting, deal & issue follow-up
Do not build a workflow engine in the Cockpit. Emit the existing webhook events — Maverick Spend Detected, Reconciliation Mismatch, Financial Health Deterioration, ESG Assurance Expiring, Contract Compliance Breach — and let Coworker.ai run the follow-up sequence with the supplier or category owner.
Cockpit role · Signal producer — detects, scores and stamps the event
Full overlap · Claude Enterprise
General writing, summarisation, narrative reporting, coding help
Do not build a general chat product. Expose a read-only, RBAC-scoped connector over /api/v1/spend/*, /api/v1/suppliers/* and /api/v1/metrics/* so Claude answers are grounded in live Cockpit data with source lineage and data-vintage stamping.
Cockpit role · Grounded data provider — governed, versioned, auditable
Recommended Integration Surface
How each agent platform plugs into the existing API layer
| Tool | Protocol | Auth | Scopes | Data returned | Limits | Audit |
|---|---|---|---|---|---|---|
| Claude Enterprise | MCP server / REST (read-only) | OIDC service identity via Azure AD, short-lived token | spend:read, suppliers:read, metrics:read, esg:read | Classified spend lines, supplier master + financial health, SUM / maverick / compliance metrics | 600 req/min, 50k rows per query, no raw registry payloads | Every call written to audit.event.v1 (WORM) with actor, prompt hash, resource |
| Coworker.ai | HTTPS webhook subscriber + task write-back | HMAC-signed webhooks, scoped API key for status write-back | events:subscribe, tasks:write | Alert payloads only (IDs, category, severity, thresholds) — no supplier financial raw data | Event-driven, at-least-once delivery with DLQ replay | Delivery receipts and task outcomes appended to audit.event.v1 |
Guardrails
Applies to every external AI agent connected to the Cockpit
- ▍Read-only by default — agents never write to spend, classification or supplier master records
- ▍RBAC-scoped service identity mapped to the existing role matrix; agents inherit the requesting user's scope
- ▍Every agent request and response hash written to the append-only WORM audit log (audit.event.v1)
- ▍No agent access to raw commercial register or credit-bureau payloads — only derived, licensed-safe scores
- ▍Answers always carry source lineage and data-vintage stamps so agent output stays defensible