McKinsey published a figure at the end of 2024: among companies that claim to use AI, fewer than 10% have seen "meaningful cost reductions." Models keep getting stronger, tools keep multiplying, but at most firms AI still means employees drafting copy or summarizing documents — a long way from real business workflows.
An engineer who shared his hands-on experience on Juejin calls this gap "deployment engineering." Between off-the-shelf AI tools and actual enterprise work, four barriers stand in the way: source credibility, system permissions, fallback handling, and employee habits. He proposes a framework called AKA — A for Agent (handling tasks), K for Knowledge (feeding trustworthy data), A for Automation (wiring steps into automated flows) — along with a new role called FDE (Forward Deployed Engineer: an engineer embedded on-site to integrate AI into business processes).
What this is
Tools like Codex and WorkBuddy are essentially "general-purpose AI workbenches": they can read files, write code, and call APIs, but they don't know your company's customer data, refund policies, or approval permissions. To get AI actually handling support tickets, letting it read chat logs isn't enough — you also need knowledge bases, business system APIs, permission controls, and human fallbacks. The FDE's job is to assemble, test, train, and maintain these generic tools around specific scenarios until employees can use them independently on real tasks.
Industry view
Supporters argue the bottleneck for AI entering the enterprise has indeed shifted from "can the model do it?" to "can engineering deploy it?" The "on-site + customization + hand-holding" FDE model has been validated overseas by OpenAI and Anthropic, and started appearing at China's big tech firms and consulting circles last year.
Critics are equally sharp: so-called FDEs are often just system integrators or outsourcing consultants rebranded with an AI label. If a company's own processes and data governance are broken, more FDEs only automate the chaos. There is also the data-exfiltration risk — sensitive business data fed into third-party tool foundations leaves compliance boundaries fuzzy.
Impact on regular people
For enterprise IT: Buying an AI workbench is no longer the finish line. Three things come after: knowledge-base governance, API integration, and permission auditing. IT departments will shift from "installing systems" to "managing context."
For individual careers: Over the next one to two years, hybrid roles that "understand both the business and AI deployment" will outvalue pure prompt-engineering roles. Frontline employees willing to pilot are often the hidden beneficiaries of process redesign.
For the consumer market: In the short term, enterprise AI services will push SaaS subscription fees higher, but improvements in consumer-facing product experience will lag — because the bottleneck sits inside the company, not in the model.