Enterprise AI proof-of-concept deployment: tracking which projects actually reach production

Over eighteen months, we tracked 128 enterprise generative AI projects across Greater China, centered on Taiwan. Of those, 18% deployed to production and remained in active use three months later.

What happened to the other 80%? Forty-one percent stalled at the PoC and demo stage, never reaching production. Twenty-two percent launched but were shelved within six months, with adoption below 10%. Another 19% remained in perpetual pilot mode, timelines stretching to twice the original plan, no one dared shut them down, and no one pushed them live.

Few organizations publicly share how many AI projects actually reach production. Within Chinese-speaking markets, this data is especially scarce. We compiled these numbers from 128 projects we tracked directly, interviewed, or audited.

Defining go-live

To make sense of these numbers, we apply a strict definition. A project counts as live only when all three conditions are met: it entered a production environment (not a test server or sandbox); real business users rely on it daily (not internal demos); and it sustained that use for at least 90 days with at least one tracked business metric, ticket deflection rate, processing time, conversion rate, or similar. All three must hold.

Most organizations call something live once they've launched it. This definition counts it only when someone actually depends on it daily. The 18% figure reflects that stricter boundary.

The sample comprises 128 projects directly managed, audited, or studied between Q4 2024 and Q1 2026, spanning financial services, healthcare, manufacturing, retail, logistics, and automotive. Sixty implementation leaders participated in anonymous surveys. This is a pragmatic sample with inherent selection bias, consulting firms typically engage with organizations that are stuck. The numbers reflect directional patterns rather than statistical precision.

Deployment rates by industry

The overall rate is 18%, but significant variation exists across industries.

Financial services leads not through superior technical capability but because compliance requirements force clear process documentation, creating defined integration points for AI. Healthcare faces obstacles primarily in data governance and accountability. A single model error carries high cost, making teams reluctant to deploy.

Project type shows sharper variation: internal knowledge Q&A and Copilot-style applications reached 12% deployment, the lowest category. Targeted agentic process automation reached 21%. General-purpose tools designed for everyone lack a non-negotiable use case; they more easily become unused browser tabs.

Three variables linked to deployment success

Three variables correlate strongly with deployment success across the 128 projects.

The first is whether a quantifiable business metric was set from day one. Projects with defined targets, deflecting 30% of frontline customer service tickets, for instance, showed three times the deployment rate of projects without metrics. Projects lacking clear targets struggle to define success and naturally extend as endless pilots.

The second is senior executive sponsorship. Projects with C-level ownership reached roughly double the deployment rate. The final push to production requires cross-functional coordination and IT/security sign-off. Without executive authority to make decisions, projects stall in coordination cycles.

The third is sustained engineer involvement through production. Consulting-only engagements, providing a report and roadmap, showed 9% deployment. On-site engineers addressing edge cases and coaching adoption metrics upward reached 34%. That represents a significant gap.

Most projects stall not because of model performance but because production deployment requires ongoing engineering support. Field deployment engineering, having engineers on-site to address edge cases and track adoption through the first 90 days, appears critical to success. We update this data quarterly. Organizations with stalled projects can analyze these patterns to identify specific obstacles.