You Have the Map. Now You Need the Route.
In Part 1, we mapped the terrain: seven distinct AI paradigms for enterprise systems, each with its own data requirements, risk profile, and value trajectory. From Knowledge Retrieval (RAG) at the accessible end to fully autonomous Agentic AI at the frontier.
The response was telling. The most common question wasn’t “Which paradigm should we start with?” — it was “What’s the sequence? How do we actually execute this?”
Fair question. Knowing what to build is necessary. Knowing when to build it — and why that order matters — is what separates a strategy from a slide deck.
Missed Part 1? Read it here.
The 24-month playbook isn’t about doing everything at once. It’s a crawl-walk-run approach where each phase funds the next, and early infrastructure becomes the foundation for later ambition.
Why Sequence Matters
Enterprise AI isn’t a single deployment decision. It’s a portfolio that unfolds over time. Deploying all seven paradigms simultaneously would be like building seven houses on the same foundation before the concrete sets.
Three principles drive the sequencing:
1. Data pipelines compound. RAG forces you to build document ingestion and embedding infrastructure. That same pipeline powers Document Intelligence. The clean data feeds Process Mining. Each paradigm inherits infrastructure from its predecessor.
2. Organizational trust is earned. An enterprise that hasn’t seen AI deliver accurate document classification won’t trust AI to make autonomous purchasing decisions. Trust scales with demonstrated accuracy — and that takes time.
3. Early ROI funds later ambition. Phase 1 wins (faster ticket resolution, automated invoice processing) generate measurable savings. Those savings become the business case for Phase 2 investment. Self-funding beats budget fights.
The Four Phases
The 24-month journey breaks into four overlapping phases. Note the overlaps — this isn’t a strict waterfall. As Phase 1 paradigms stabilize, Phase 2 work begins. The key is that each phase’s prerequisites are met before its paradigms go live.
Phase 1: Foundation
The goal: Build the core data infrastructure, prove AI value with visible quick wins, and establish governance patterns that every later phase will inherit.
RAG goes live first because it’s read-only — no writes to enterprise systems, no transaction risk. Deploy it against system documentation, configuration guides, and process manuals. The help desk sees immediate relief. The business sees a tangible AI win within weeks.
Document Intelligence follows because it leverages the same document ingestion pipeline RAG established. Start with the highest-volume, most standardized document type — typically invoices. A confidence-threshold framework (auto-post above 95%, review at 80-95%, manual below 80%) means humans stay in the loop while straight-through rates climb toward 60-80%.
Phase 2: Intelligence
The goal: Move from automating what you already know to discovering what you don’t — hidden process inefficiencies, data-driven decision points, and task-level automation opportunities.
Process Mining is the intelligence multiplier. Enterprise systems already generate massive event logs — change documents, workflow logs, audit trails. Process Mining reconstructs how processes actually execute versus how they were designed. The insight density per dollar is among the highest of any paradigm.
Decision Augmentation deploys in shadow mode first: the AI makes recommendations silently alongside human decisions for 4–8 weeks. Where it outperforms, you surface it. Where it doesn’t, you retrain. Trust is built with data, not faith.
Narrow Agents enter here — task-level agents handling well-defined, single-step actions: routing tickets, validating data, triggering notifications. Low autonomy, high volume, immediate ROI.
Copilot launches in Phase 2 because it doesn’t need a complete backend to deliver value — it needs a growing one. Connected to the RAG layer from Phase 1 and progressively enriched by process mining insights and decision models as they come online, the copilot becomes a real-time window into every capability you’ve built so far.
Phase 3: Transformation
The goal: Deploy the paradigm that requires the most data depth and the longest runway — and delivers the largest per-unit ROI when the foundation is right.
Deep ML / Forecasting requires the most historical data: 24+ months of clean, cross-functional records spanning promotions, seasonality, and supplier constraints. By month 8, Phase 1’s data pipelines have been running long enough to provide the foundation, and Phase 2’s process mining has identified the highest-impact prediction targets. Start with demand forecasting — a 10% accuracy improvement can reduce inventory carrying costs by 15–25%.
With the copilot already live from Phase 2, forecasting outputs can be surfaced to planners through natural language queries from day one — no new UI required.
Phase 4: Autonomy
The goal: Graduate from agents that assist to agents that act — multi-step, cross-system autonomous workflows with human oversight at the governance layer, not the transaction layer.
By now, the enterprise has 12–18 months of AI governance in production. Model accuracy is tracked. Audit trails are mature. Organizational trust has been earned through demonstrated performance across three prior phases.
Autonomous agents scale from the narrow agents deployed in Phase 2. A ticket-routing agent becomes an exception-handling agent. A data-validation agent becomes a PO-creation agent. The scope widens, but the governance framework — human-in-the-loop escalation, confidence thresholds, rollback mechanisms — was established in Phase 1.
Making It Real
The framework is deliberate. Four phases, overlapping timelines, self-funding economics. But the specific entry point, pace, and paradigm priority depend on where your enterprise is today.
A company with mature data pipelines and an existing analytics practice might compress Phases 1 and 2 into a single push and reach Phase 3 by month 10. A company still running manual AP processing with inconsistent master data starts at Phase 1 and invests the first three months in data engineering before the first RAG model goes live.
The playbook is the same. The entry point is different. The key is starting.
