Why This Framework Exists
Most enterprise AI architectures published in the last two years fit on one of two spectrums: too academic (great research, no operating model) or too vendor-glossy (great brochures, no implementation detail). The LuMay Enterprise AI Framework is the operating-model version we wished we had when we started shipping enterprise agents.
It's opinionated. We've made specific calls about how the architecture sits on existing systems of record, how compliance posture flows through the stack, and what the engagement model looks like for strategy → implementation → adoption. Some of those calls may be wrong for your context - but at least they're explicit.
The Six-Layer Agentic Core
We model every enterprise AI deployment on a six-layer stack, top to bottom:
L6 · Analytics & ROI
Real-time measurement of the business outcomes the agents are driving - cost savings, capacity unlocked, CSAT delta, audit findings reduced. Without this layer the agents become science projects.
L5 · Connectors & Adapters
Native integrations to your existing systems of record - Salesforce, Dynamics, ServiceNow, SAP, your custom DMS/ERP. The agent layer reads from and writes to your canonical data; it does not create a parallel data store.
L4 · Core Engine
Multi-agent orchestration. We use LangGraph, Autogen, CrewAI, and Semantic Kernel underneath - the framework is agnostic about which orchestrator runs which workflow.
L3 · Voice OS
For voice deployments specifically, a telephony-grade voice layer with sub-second turn-taking. We provide Voxentis; you could substitute another voice OS if it meets the latency and compliance bar.
L2 · Security & Governance
SOC 2, ISO 27001, HIPAA, GDPR, India DPDP 2023, and industry-specific frameworks (FDA, ISO 13485, GMP for medical-device; OCG for legal billing). Posture flows from L1 fabric choices up through every layer.
L1 · Deployment Fabric
SaaS, public cloud, private cloud, or fully on-prem. The fabric is the constraint that defines what L2 compliance posture is possible.
Operating Model
Capability layers without an operating model are blueprints without buildings. The framework specifies four operating-model pieces:
- Reusable platform vs. customer-specific configuration- roughly ~80% of any deployment is platform; you bring the ~20% that's yours (glossary, playbook, compliance posture).
- Strategy → Implementation → Adoption sequencing- never skip adoption. Where engineering has shipped but adoption has lagged, we've had to come back to add it.
- Outcome-priced engagement - the engagement model that aligns incentives is fixed-fee, outcome-priced, not time-and-materials.
- Single accountable owner - one named person on the partner side is accountable for outcomes. Not a rotating account team.
Compliance Posture Flows Top-Down
A common mistake: choosing the deployment fabric (L1) without thinking through the compliance posture (L2) that the workload requires. Once you've made an L1 choice like "SaaS shared tenancy," certain L2 outcomes become impossible or expensive to retrofit.
The framework recommends mapping every candidate workflow against required compliance posture first, then choosing the deployment fabric that satisfies the strictest workload. Workloads can step down from the strictest fabric (e.g., a HIPAA-aligned on-prem can also run a SOC 2 workload), but they cannot step up without redeployment.
Engagement Model
We use a four-phase engagement cadence that maps to roughly four quarters:
- Discovery (Weeks 1–4)- Strategy & Advisory engagement producing the signed deployment plan. This is the only phase that's safe to parallelize across vendors.
- First agent live (Weeks 5–8) - Pick the highest-confidence workflow, ship it. ROI-positive within this phase or the project is wrong.
- Adoption (Months 3–4) - Academy cohort, runbooks documented, named accountable owners. The agent goes from working to used.
- Scale (Months 5+) - Add agents per quarter; per-agent cost drops as shared fabric amortizes.
What This Framework Deliberately Omits
- Specific model selection guidance - model selection is workload- and time-dependent. Hardcoding model names into a framework dates badly.
- Build-vs-buy economics- we're biased; this isn't the right doc for that analysis. We recommend Gartner-style frameworks for the build-vs-buy decision.
- Cost forecasting - the ROI calculator is the right tool. Strategy & Advisory delivers the custom version.
- Model governance details - covered in the separate Enterprise Agentic AI Guide.
The framework is opinionated and open. We've made specific calls so the architecture is debatable rather than vague. Disagree with a call and we'll learn from it.