01 · The thesis
Do not ask the customer to start over.
The initial sale is a high-value AI layer that orchestrates the enterprise systems already in place. Replacement becomes an option later, after the capability boundary has earned trust.
“Try the layer that makes every rep dramatically more effective—not the migration that asks the enterprise to rip out its commercial core.
Avoid
CRM with AI attached
- Tools mirror database objects
- Every channel rebuilds business logic
- Agents inherit vendor and schema coupling
- Modernization begins with migration risk
Recommended
AI commercial operating system
- Tools express business outcomes
- Every experience reuses the same capabilities
- Policies and provenance travel end to end
- Modernization begins with measurable value
Capabilities, not tables
Agents and applications ask for business outcomes. They never inherit the shape of an underlying CRM schema.
Stable boundary, replaceable core
Systems, models, storage, brokers, and channels can change behind versioned capability contracts.
Policy before execution
Identity, tenant, purpose, consent, risk, and approval context travel with every invocation.
Events as the nervous system
Services publish durable facts. Subscribers react without turning the platform into a distributed tangle.
02 · Reference architecture
Seven layers. One governed nervous system.
The event backbone cuts across the stack. Synchronous requests still serve immediate user needs; durable events decouple everything that can react independently.
Mobile · Web · Teams · Voice · API · Ambient
Sales · Manager · Compliance · Analytics · Coordinator
Discovery · Contracts · Authorization · Audit
Briefings · Engagements · Compliance · Next best action
Adapters · Entity resolution · Source-of-truth routing
Domain events · Schema registry · Replay · Subscribers
Veeva · Salesforce · Microsoft · SAP · Reltio · IQVIA
Business capabilities form the stable contract. MCP makes them model-usable; agents coordinate them; channels present them. No layer is allowed to bypass policy, tenant boundaries, source-of-truth rules, or audit.
The durable boundary
Capabilities survive the technology underneath them.
Replace a CRM, model provider, event broker, vector store, or mobile client without changing what prepare_customer_briefing means to the business.
03 · Reference request
“Prepare me for Dr. Jones.”
One request proves the whole architecture: identity resolution, territory access, governed retrieval, compliance, recommendations, provenance, graceful degradation, and an auditable result.
Normal
Resolve the HCP, gather current evidence, cite every fact, emit BriefingPrepared.
Ambiguous
Do not guess which Dr. Jones. Ask the rep to disambiguate with authorized context.
Degraded
Show source outages and freshness explicitly. Return a partial briefing only when policy permits.
Denied
Refuse restricted data and record the policy decision without leaking the protected reason.
04 · Capability registry
The business contract every model and app can trust.
Each capability has a purpose, typed inputs and outputs, eligible actors, policy requirements, system dependencies, events, side-effect semantics, SLO hypotheses, and a versioned lifecycle.
| ID | Business capability | Domain | Effect | Primary event |
|---|---|---|---|---|
CAP-BRF-001 | Prepare customer briefing | Meeting prep | Read | BriefingPrepared |
CAP-ENG-004 | Capture engagement | Engagement | Write | EngagementCaptured |
CAP-CMP-003 | Check interaction compliance | Compliance | Read | ComplianceChecked |
CAP-NBA-002 | Recommend next best action | Decisioning | Read | ActionRecommended |
CAP-CNT-006 | Find approved content | Content | Read | ContentSelected |
CAP-CNS-002 | Resolve communication eligibility | Consent | Read | EligibilityResolved |
CAP-MDM-005 | Resolve customer identity | Master data | Mixed | IdentityResolved |
CAP-ANL-008 | Summarize territory performance | Analytics | Read | SummaryGenerated |
Formal MCP layer
A secure tool catalog—not a second business-services layer.
MCP adapters describe, authorize, invoke, observe, and version capabilities. They do not recreate compliance rules or query source databases directly.
- Delegated user + tenant context
- Read, write, destructive, and approval annotations
- Stable errors, cancellation, pagination, and job handles
- Correlation, causation, idempotency, and provenance
- Model-independent conformance tests
{
"name": "briefing.prepare_customer",
"capability_id": "CAP-BRF-001",
"version": "1.0.0",
"effect": "read",
"risk": "moderate",
"requires": {
"identity": true,
"purpose": "meeting_preparation",
"policies": ["POL-TERR-001", "POL-CNS-002"]
},
"emits": ["BriefingPrepared"],
"audit": "full_provenance"
}05 · Agent architecture
Specialize the judgment. Centralize the guardrails.
Bounded agents bring domain context. A coordinator handles routing and durable workflow state, but never owns the business logic or the user's standing permissions.
Sales agent
Prepares briefings, finds approved content, captures engagements, and coordinates follow-up.
Manager agent
Surfaces team signals, prepares coaching context, and tracks decisions without becoming a surveillance layer.
Compliance agent
Explains applicable rules, evaluates content and interaction context, and escalates uncertainty.
Analytics agent
Answers governed questions across commercial data with freshness, confidence, and source provenance.
Memory is not a system of record
Six tiers. Six explicit policies.
Every tier defines ownership, permitted content, provenance, tenant isolation, TTL, correction, deletion, encryption, and user visibility.
SecondsHoursPolicy-definedWorkflow-definedSource-ownedRecord policy06 · Security & trust
Identity and purpose travel with the request.
The architecture is deny-by-default. Agents do not hold source-system credentials, the LLM is never a policy engine, and consequential writes require deterministic authorization plus appropriate confirmation.
Confused deputy
Re-evaluate delegated identity, tenant, purpose, and policy at the capability boundary.
Prompt injection
Treat retrieved content as data; constrain tool selection and isolate untrusted instructions.
Excessive agency
Classify consequences, cap permissions, require confirmation, and stop recursive loops.
Unsafe retries
Attach idempotency keys, concurrency guards, and explicit commit states to every mutation.
Tenant escape
Enforce isolation in policy, data access, caches, memory, telemetry, and event partitions.
Opaque evidence
Persist source, freshness, model, prompt, tool, policy, and human-decision provenance.
HIPAA, FDA, GxP, 21 CFR Part 11, GDPR, state privacy, promotional practice, records retention, and pharmacovigilance are applicability questions—not marketing badges. Each requires a specific trigger, evidence, owner, and qualified review.
07 · Ecosystem strategy
A marketplace is an outcome, not a launch feature.
The near-term product is a governed catalog. Broader exchange only makes sense after the platform can prove demand, certification, runtime isolation, support boundaries, and trustworthy economics.
First-party catalog
Own the contracts, reference connectors, policy packs, and support model.
Certified design partners
Prove publisher identity, conformance, signing, review, and revocation.
Private enterprise exchange
Let customers distribute approved internal and partner packages.
Curated external marketplace
Open only when trust, discovery, demand, billing, and liability are ready.
08 · Adoption roadmap
Prove value before asking for replacement.
Start with one read-only, high-frequency workflow. Add writes only after identity, policy, audit, connector health, and evaluation gates are strong enough to carry consequence.
Prove the overlay
Prepare me for Dr. Jones
- Read-only aggregation
- Identity + territory controls
- Cited briefing
- Shadow-mode evaluation
Add governed action
From insight to follow-through
- Engagement capture
- Human approvals
- Idempotent writes
- Connector health
Scale the operating model
Agents across workflows
- Manager + compliance agents
- Event subscriptions
- Developer portal
- Private catalog
Open the ecosystem carefully
Progressive replacement
- Certified partners
- Vertical packs
- Optional component replacement
- Marketplace gate
Architecture decision register
Decisions that hold the thesis together.
Earn replacement rights after proven value.
Keep CRM schemas and vendor APIs behind adapters.
Agents invoke tools; tools invoke authorized services.
Route and persist workflows without creating a new monolith.
Avoid unbounded exactly-once claims.
Prove trust, demand, certification, and support first.
The first proof
One request. Every architectural promise.
If NXTCRM can prepare a reliable, cited, policy-safe briefing from fragmented enterprise systems—and show exactly how it did it—the platform has proved the wedge. Everything else can expand from that stable capability boundary.