Enterprise AI Blueprints & Reference Architectures

Every blueprint carries a planning horizon — Now: Decisions in flight this year · Next: 2027 planning horizon · Beyond: 2028–2029 structural shifts.

Decision briefs for leaders, reference architectures for the teams who build

Agentic operations

The control plane for work that runs itself — orchestration, tools, memory and the supervision layer that keeps autonomous work inside policy.

Blueprint · Next

The agentic control plane

Whether to fund a control plane the enterprise owns before scaling agents, or let each team ship its own agent stack and consolidate later.

Stage 3 · 10 min read + diagramsRead the decision brief →

Conversational service

Voice and digital service delivered by AI first, with humans held for the contacts that need judgement. The largest near-term cost and experience lever in most enterprises.

Blueprint · Now

Agent Assist on Amazon Connect + Bedrock

Whether to put a real-time assistant beside every agent now, or wait for the platform vendor's roadmap to close the gap.

Stage 1 · 3 min readRead the decision brief →

Blueprint · Next

Voice AI on NICE CXone + Enlighten

Whether to retire the IVR estate on the incumbent platform, or run voice AI as a layer above it and decouple from the platform decision.

Stage 2 · 3 min readRead the decision brief →

Shared vocabulary

Four stages, one scale — used in every blueprint on this shelf.

Most disagreement about AI progress is really disagreement about definitions. Each blueprint states the stage it takes you to and the measure that proves you got there.

Stage 1
Assisted
AI suggests, a person decides and acts. Value shows up as speed and consistency, not headcount.
Proof: Suggestion accept rate, handle time
Stage 2
Automated
Defined tasks complete end-to-end inside fixed rules. Exceptions route to people by design.
Proof: Containment, straight-through rate
Stage 3
Agentic
Software plans across several steps and tools toward an outcome, inside policy and budget limits, with supervision.
Proof: Outcome completion, intervention rate
Stage 4
Autonomous
Whole workstreams run without a queue. People set policy, review exceptions and own the result.
Proof: Cost per outcome, incident rate

Quality & assurance

Scoring every interaction rather than a sample — the evidence base regulators, boards and operating committees increasingly ask for.

Blueprint · Now

AI QA on Genesys Cloud CX

Whether quality moves from a sampled, human-scored process to full-coverage scoring with analysts moved onto calibration and coaching.

Stage 2 · 3 min readRead the decision brief →

The 2027 reference stack

What a 2027 enterprise AI estate looks like — and who owns each layer.

Every blueprint below plugs into one of these five layers. Read top-down for the experience your customers and employees see; bottom-up for the decisions that determine whether it holds.

Experience surfacesVoice, chat, agent desktop, employee portal, embedded assistantsCX & workplace leaders
Agentic control planePlanning, orchestration, tool contracts, memory, human-in-the-loopChief AI Officer
Knowledge & data foundationRetrieval, entitlement, identity resolution, systems of recordCIO & data leadership
Model & routing layerMulti-provider routing, caching, failover, cost and quality gatesPlatform engineering
Governance, assurance & FinOpsEvaluation, audit, agent identity, unit economics, incident responseCIO, CFO & risk

Publishing roadmap

What joins this shelf next.

One new blueprint a quarter, every published blueprint refreshed twice a year, each tied to the current Annual Outlook.

Governance & controlsNext
Agent accountability & controls
Agent identity, audit trails, evaluation gates in CI and the escalation model that answers the question a board asks after the first incident.
Expected Q4 2026
Data & knowledge foundationNow
The enterprise knowledge layer
Retrieval, entitlement and identity as one funded asset shared across service, employee support and automation.
Expected Q1 2027
Employee & IT serviceNow
Employee & IT service agents
IT, HR and workplace requests resolved without a ticket queue — the same architecture as customer service, different systems of record.
Expected Q1 2027
Agentic operationsNext
Back-office document automation
Intake, extraction, adjudication and exception handling across claims, onboarding and collections work.
Expected Q2 2027
Governance & controlsBeyond
AI cost & FinOps architecture
Cost attribution, cheap-first routing and budget guardrails designed in, so unit economics hold as volume scales.
Expected Q3 2027