The 95% Problem
The number has become an industry cliche, but that doesn't make it less true: the vast majority of enterprise AI initiatives fail to reach production. Not because the models don't work in the lab. Not because the data scientists aren't talented. Not because leadership lacks ambition. They fail because there's a fundamental mismatch between how enterprises approach AI and what AI actually needs to deliver value in operational environments.
The typical enterprise AI playbook goes something like this: hire a data science team, pick a use case, collect data, train a model, build a proof of concept, present impressive accuracy metrics to the steering committee, then watch the project stall when it encounters the messy reality of production systems, imperfect data, and organizational resistance.
This pattern repeats across industries, but it's particularly acute in supply chain and logistics, where the gap between AI's theoretical promise and operational reality is wider than almost anywhere else.
The Three Failure Modes
Enterprise AI failures in operations cluster around three root causes. Understanding them is the first step toward an approach that actually works.
1. Insufficient Domain Data
General-purpose large language models are trained on internet-scale data. They know a lot about a lot of things. But they know almost nothing about your specific operation: your carrier performance patterns, your seasonal demand curves, your facility-specific labor productivity benchmarks, your vendor lead time reliability, your SKU velocity distributions.
This domain data is what separates a useful operational AI from a sophisticated chatbot. Without it, the model can talk intelligently about supply chain concepts but can't answer the question that matters: "Should I move labor from receiving to picking right now, in this facility, given what's happening today?"
Most enterprise AI initiatives attempt to solve this by connecting the model to internal data. But "connecting to data" and "understanding operational context" are very different things. A model that can query your WMS doesn't automatically understand that a 94% on-time ship rate in your business is actually a crisis because your SLA commitment is 98%.
A model that can query your data doesn't automatically understand your business. Context isn't a data connection. It's a knowledge architecture.
2. Lack of Operational Context
Operational decisions don't happen in isolation. They happen in the context of everything else happening simultaneously. The decision to authorize overtime on second shift depends not just on the current workload, but on tomorrow's inbound volume, next week's carrier schedules, the labor budget year-to-date, the availability of cross-trained workers, and the fact that half the team just worked overtime yesterday.
General-purpose AI models can't reason across this operational context because they don't have access to it. The data lives in eight different systems, each with its own data model, its own update frequency, and its own definition of terms. Building the cross-system context layer is the hard engineering work that most AI initiatives underestimate by an order of magnitude.
3. Over-Reliance on General-Purpose Models
The recent explosion of large language model capabilities has created a tempting illusion: that a sufficiently powerful general-purpose model can be pointed at any problem and produce useful results. This works for tasks where the knowledge is broadly available (writing, summarization, code generation) but breaks down in domains where the critical knowledge is proprietary, contextual, and operational.
A general-purpose model can write a reasonable-sounding analysis of warehouse productivity drivers. But it can't tell you that the productivity drop in Zone 3 last Tuesday was caused by a combination of a new temporary worker who wasn't cross-trained on the voice pick system, an inbound receipt that overflowed into the pick aisle, and a wave release that was 45 minutes late because the planner was in a meeting. That level of specificity requires a system designed from the ground up to ingest, correlate, and reason across operational data.
The Hybrid Model
The enterprises that are succeeding with AI in operations have converged on what we call the hybrid model: a combination of general-purpose AI capabilities with domain-specific data platforms purpose-built for the operational environment.
The hybrid model has three layers:
- Domain data layer: A unified operational data fabric that ingests, normalizes, and correlates data from all operational systems (WMS, TMS, ERP, LMS, YMS) in real time. This isn't a data warehouse. It's a live, queryable representation of the current state of the operation.
- Domain intelligence layer: Pre-built analytical models tuned to operational patterns: demand forecasting, labor productivity analysis, carrier reliability scoring, inventory optimization. These models understand the domain because they were built for it, not adapted to it.
- General AI layer: Large language models and generative AI capabilities that sit on top of the domain layers, providing natural language interfaces, anomaly explanation, scenario generation, and decision support. The general AI makes the domain intelligence accessible. The domain intelligence makes the general AI useful.
Why This Works in Supply Chain
Supply chain and logistics operations are uniquely suited to the hybrid model for three reasons.
First, the data is structured and abundant. Unlike industries where AI struggles with unstructured data (legal documents, medical images, creative content), logistics operations generate massive volumes of structured transactional data: shipments, orders, inventory movements, labor transactions, dock events. This data is the raw material that domain-specific models thrive on.
Second, the decisions are high-frequency and high-impact. A warehouse makes thousands of operational decisions every day: what to pick next, where to put incoming inventory, how to staff the next shift, which dock to assign to an arriving truck. Each decision is individually small but collectively determines whether the operation hits its service levels and cost targets. AI that improves these decisions by even a few percentage points delivers measurable ROI.
Third, the domain knowledge is deep but bounded. Supply chain operations follow known patterns and principles. Velocity-based slotting, wave planning optimization, labor productivity benchmarking, carrier performance scoring: these are well-understood problems with established best practices. A domain-specific platform can encode these practices and adapt them to each operation's specific context, rather than trying to learn everything from scratch.
The future of enterprise AI isn't bigger models. It's smarter architecture: the right model, with the right data, in the right context.
What Changes
The hybrid model changes the enterprise AI conversation in three fundamental ways.
First, time-to-value collapses. Instead of spending 12-18 months building a custom AI solution, organizations deploy a domain-specific platform in weeks, with pre-built models that start delivering insights from day one. The general AI layer provides the interface; the domain layer provides the intelligence.
Second, the build-versus-buy decision resolves. Most enterprises shouldn't be building operational AI from scratch, just as most enterprises shouldn't be building their own WMS. The hybrid model lets organizations buy the domain platform and use their internal AI talent for the last mile: customizing insights, building organization-specific models, and integrating AI outputs into existing workflows.
Third, the feedback loop closes. Domain-specific platforms get better with use because they accumulate domain data. Every operational decision, every outcome, every anomaly feeds back into the models. General-purpose models don't have this feedback loop because they aren't connected to operational outcomes. The hybrid model creates a compounding advantage where the AI gets more accurate and more useful every week.
The 95% failure rate isn't inevitable. It's the result of a specific approach: trying to make general-purpose tools do domain-specific work. The hybrid model is the alternative, and it's working.
