AI bridges the technology gap but sparks fiercer debates — how product managers and developers can actually collaborate

AI closes information gaps — but creates a new “half‑understanding” wall

Generative AI has moved from novelty to workplace staple in China: product managers use it to draft requirements and mockups, while engineers lean on tools like Copilot and Tongyi Lingma (通义灵码) from Baidu (百度) to spin up demo code. The expectation was simple: fewer misunderstandings, smoother handoffs. Reality is the opposite. It has been reported that AI’s ability to make people “look like they understand” often produces a brittle middle ground — enough knowledge to argue, not enough to reach shared technical or business judgment. The result: more coordination friction and, in some teams, “each‑side‑for‑itself” dynamics.

A practical cost reckoning — when demos and 20‑page specs collide

A frontline case illustrates the gap. A developer delivered a working demo that scraped data and auto‑generated weekly reports; a product manager presented a 20‑page, AI‑studded requirements pack promising department‑level customisation. Engineers flagged compute and architecture limits; product managers countered that the demo lacked real user adaptability. The deeper problem was a “half‑bucket” cognition mismatch: PMs could read architecture diagrams but underestimated hidden costs; engineers could sketch features but missed user pain and prioritisation. The team did the math — at ~100,000 comments/day and an estimated per‑call inference cost of ¥0.00011, annual inference alone would be ~¥400,000, rising to ¥500,000 with tuning and ops — versus a projected revenue uptick of only ~¥180,000. Those numbers forced a rule change: no proposal without three quantified ledgers — business benefit, inference/training compute cost, and data labelling/cleaning cost. Note: it has been reported that broader geopolitical factors, such as export controls on advanced chips, are tightening access to high‑end accelerators and can push cloud GPU costs higher, making such cost calculations more consequential.

Three practical dimensions to rebuild collaboration

Teams that have moved from finger‑pointing to co‑creation focus on three dimensions: value density (is this worth the compute and ops spend?), data readiness (do we actually have the coverage and quality for the model to work?), and shared metrics (joint ROI and operational thresholds). Practically, that means three prechecks before any AI feature moves forward: (1) joint ROI modeling by PM and engineering; (2) a data‑asset pre‑audit with the data team; (3) a scaled‑down technical proof that includes compute and latency estimates. Change the rituals and incentives, and AI becomes a tool again — not an argument starter. Will these rules scale across China’s fast‑moving product shops? The better answer is not a policy but a habit: insist on numbers, surface unknowns early, and make accountability shared.