Why Ontology, Not Speed, Is the Real Bottleneck After Rapid AI App Deployment

01 From Requirement Docs to On‑Site Knowledge Direct Compilation

The AIPCon 9 presentation highlighted a battery‑factory case where the team recorded on‑site discussions and directly transformed process knowledge, exception patterns, and role requirements into applications and workflows within the same day. Although generative AI can quickly produce interfaces and basic logic, production still requires mapping spoken descriptions to stable objects, fields, rules, and actions—for example, linking "line downtime" to equipment, work orders, spare parts, personnel, and capacity impact.

02 Production‑Grade AI Apps Need Ontology, Tools, and Governance

A usable agent application consists of three layers. The first layer, Ontology, models customers, orders, equipment, inventory, and personnel as shared business objects. The second layer, Tools, wraps queries, predictions, optimizations, and transactional interfaces into callable actions. The third layer, Governance, defines data and action permissions, approval workflows, logging, and evaluation. Large language models handle intent understanding and context organization, but they cannot replace these deterministic foundations.

03 From Prediction to Optimization to Execution, Agents Must Cross Three Boundaries

The fleet‑management case illustrates a typical upgrade path: stage 1 builds a prediction model to assess asset availability; stage 2 adds an agent that continuously monitors input streams, detects model drift, and flags invalid assumptions; stage 3 integrates inventory and sales forecasts into an optimizer that generates tens of thousands of asset‑reallocation recommendations. Each stage introduces new engineering requirements: prediction needs features and evaluation, monitoring needs event streams and drift detection, optimization needs objective functions and constraints, and execution demands permission, approval, and rollback mechanisms. Many projects stall at stage 2 because they lack reliable write‑back interfaces and responsibility mechanisms.

04 Value Extends Fractally Along the Up‑ and Down‑Stream

Even if a production‑line app goes live in a day, it may merely shift congestion to procurement, quality, or customer service. Teams typically expand the solution the next day to upstream or downstream departments, and continue adding new object relationships, constraints, and actions until the entire value chain is covered. To support this fractal expansion, technical teams must pre‑define module boundaries: a shared ontology and master data, event‑driven state synchronization, composable agent tools, and role‑based permission policies. Without these, rapid delivery devolves into the rapid creation of isolated silos.

05 AI Engineering Teams Will Evolve From Project Groups to Value‑Chain Squads

When prototypes can be edited the same day, keeping business experts outside the "requirement review" loop is no longer sensible. Experts who understand factory cadence, underwriting rules, or exception handling should collaborate continuously with data engineers, application engineers, and governance specialists to define objects, rules, and evaluation metrics. The talk proposes an aggressive cadence: one week to form a value‑chain squad, 30 days to replicate key links, and a year to integrate the whole chain. This speed relies heavily on reusable business semantics, reliable action interfaces, expert time, and cross‑department decision authority.