The true value of AI to enterprises lies not in enhancing the operational speed of individual employees, but in rewriting the way entire organizations move forward.
The delivery efficiency of a project depends not only on capacity and models, but also critically on the gaps between various stages. Feishu Projects integrates AI into core R&D workflows, transforming individual acceleration into measurable, replicable, and sustainable organizational delivery capabilities.
On September 15, 2024, Feishu hosted the "2024 Feishu Future Unlimited Conference & Doubao Office Launch Event" in Beijing. During the "AI + R&D" session, Feishu Projects brought AI deep into organizational R&D management. Four speakers from NIO, Insta360, and Feishu shared practical insights, exploring together: as AI coding becomes engineers' daily routine, how can we truly bring AI into organizations and transform individual speed into enterprise delivery capability?
As AI evolves from a tool to an organizational force, R&D management is undergoing a fundamental rewrite. In the AI era, competition is no longer about who adopts models first, but who integrates AI into processes, context, and collaboration methods fastest. At this pivotal transformation point, the question enterprises must answer is no longer "whether to use AI," but "how to turn AI into genuine organizational capability." The four speakers provided concrete answers to this very question.
Event scene
Define What's “Right”
Before Agents Can Go Far
Based on nearly a year of team practice, Feishu Senior Architect He Zhe shared several data points: a 35% increase in demand delivery volume, a 300% rise in effective code per engineer, and nearly 100 MRs independently completed by designers and product managers through the Harness platform. Behind these numbers lies a paradigm shift in R&D—from AI coding to Agent-First.
Regarding this new paradigm, He Zhe proposed a core insight: the prerequisite for Agent-First isn't whether the model is smart enough, but whether the enterprise can clearly define "what is correct." The more measurable the task and the more verifiable the outcome, the broader the scope of AI authorization. Otherwise, AI will remain merely a "personal productivity tool."
Event scene
AI compresses individual execution time but cannot simultaneously shorten the end-to-end distance. The internal Feishu team’s solution was to embed AI into core R&D workflows via an engineered Harness platform, embedding security, quality, and operational constraints into default workflows so that outputs at each stage automatically trigger the next, with humans intervening only at critical gates.
What problems can the Agent-first R&D paradigm solve? He Zhe explained using two high-difficulty engineering practices.
The first involved a cross-layer architectural refactoring of a multi-dimensional table formula engine. A frontend team with almost no C++ knowledge used AI to complete approximately one million lines of core code within three months, shortening the overall cycle by over 50%, saving more than 1,800 person-days cumulatively, improving performance fivefold. Formula Engine 3.0 is now fully online, stably supporting complex calculations for single tables at the tens of millions scale.
The second practice went further—through Loop Engineering, under conditions where validation could be closed-loop, shifting from "humans find problems, test solutions" to "AI finds problems, tests solutions, learns from experience," enabling AI to autonomously achieve goals. By encapsulating Base’s core capabilities into AI-friendly SDKs and Skills, Agents could directly build and manage multi-dimensional tables via code. Based on a benchmark system built from over 1,700 real tasks, Skill effectiveness improved by 30%, total time decreased by 57%, and weekly experiment cycles increased from 20 to 1,000.
Event scene
While enterprises remain stuck at the stage of "each engineer configuring their own AI plugin," AI will forever remain a "personal productivity tool" rather than evolving into "organizational capability." This integrated work system combining Harness Engineering and the Agent-first R&D paradigm has already been productized and embedded into Feishu Projects. What it brings to organizations is a work system that integrates Harness, AI Workflow, and shared context into core R&D processes.
He Zhe pointed out that in the AI era, engineers won’t be simply replaced—they’ll be exponentially amplified. What gets amplified could be either capability or problems. The key to determining the outcome lies in whether enterprises are willing to clearly define "what is correct" upfront. This is exactly the question Feishu Projects aims to help every enterprise answer.
NIO: Connecting “Smart Islands”
into a Collaborative “Lake”
NIO presented issues much closer to real-world industrial scenarios.
As an automaker responsible for delivering new vehicle models, maintaining existing ones, and managing continuously rising software complexity, enabling AI to autonomously act within R&D processes has become a strategic solution for NIO to address internal organizational demands. Kong Yunhui, head of NIO’s R&D toolchain and AI+R&D transformation, stated: The challenge organizations face isn't isolated efficiency improvement, but how R&D teams can continue operating effectively amid rising complexity.
Reducing a bug resolution cycle from 3 days to 2 hours reflects not just a change in processing time, but demonstrates that NIO’s R&D mechanism has achieved full flow—"from output to output."
Event scene
By 2027, NIO expects significant growth in both software baselines and vehicle models requiring maintenance, indicating robust business expansion. If critical decisions still rely on temporary interventions by a few experts, organizational scale and collaboration costs will rise alongside complexity. Therefore, NIO’s first step was to codify years of accumulated digital tools, data, and expert experience into Skills and specialized Agents—building individual "islands" of capability.
But they soon discovered that even if individual capabilities are strong, without smooth handoffs between them, the overall cycle wouldn’t shorten.
Feishu Projects provides precisely the work system needed to connect these "islands" into a "lake." Using CLI + MCP along with AI Workflow / AAMP, Agents can be plugged into specific nodes, reading context, writing results, and progressing through workflows around the same work item. This process resembles a flowchart with branches and loops, allowing specialized capabilities across the Thousand Island Lake to enter the main R&D stream: during development, Agents participate in solution design and code implementation, while humans focus on reviewing designs and code. When analysis proves insufficient or real-world issues arise, the process can directly revert to earlier design or coding stages.
In more complex whole-vehicle R&D scenarios, engineers previously had to constantly switch between vehicle-cloud logs, software versions, historical issue databases, and problem tickets. Now, multiple Agents can relay along a unified workflow, consolidating analysis, solutions, modifications, and verification into a single work item. With Feishu Projects, people are no longer mere movers within the process, but decision-makers at critical junctures.
Event scene
"From islands to a lake"—what NIO is doing is enabling expert wisdom to travel along workflows to deliver results.
Kong Yunhui frankly admitted: "What truly transforms technological dividends into organizational gains is a work system capable of both codifying expertise and carrying core processes." Feishu Projects is becoming the key connector within this system.
Insta360: Reorganizing “Complexity”
by Restructuring It
When a product itself is inherently complex, how can R&D management avoid being bogged down by complexity? This is the challenge facing Insta360 in the consumer electronics industry. Liu Wenzhu, head of software R&D and project management at Insta360, began with the story of a sports camera—a device whose development involves collaboration across hardware, structure, chips, algorithms, and other specialties.
Facing the high-complexity hardware-software R&D typical of consumer electronics, Liu Wenzhu offered a core insight: "Complexity doesn't disappear, but it can be reorganized." For Insta360, the first step in AI-driven R&D efficiency isn't deploying agents immediately, but making complex R&D manageable.
Event scene
Hardware must freeze while software remains fluid; a single requirement splits into dozens of tasks, layered with versions, product lines, and defects—turning R&D complexity from "addition" into "multiplication." Previously, although processes were documented, actual collaboration happened interpersonally: hardware progress relied on asking around, software dependencies on memory, with requirements, defects, and code managed in silos. This was the contradiction Insta360 faced.
The first role Feishu Projects played was bringing software and hardware into the same space. Requirements, versions, hardware milestones, software workflows, and member permissions were unified within a single project system—not just reducing multi-process switching, but enabling true "cross-domain correlation": software engineers can now see hardware progress within their own workflows, while key hardware milestones reflect software version data, making delivery transparent, processes visible, and pathways seamless.
With unified workflows and shared context, AI could finally be integrated. For example, in a release cycle—from requirement breakdown to review—what once relied on a project manager’s experience can now be supported by AI at each node: checking whether all reviews are complete, descriptions are clear, tracking points match UI designs, etc. AI can proactively lock requirements, complete missing reviews, and automatically generate and assign follow-up tasks.
During the live demo, Liu Wenzhu showcased how AI enters workflow nodes—from requirement submission to task assignment—demonstrating Insta360’s real-world validation of Feishu Projects’ value: integrating AI into key workflow steps, replacing "people hunting for risks" with "risks finding people," which sparked enthusiastic responses from the audience.
Insta360 live demo
This wasn’t just a demo—it reflected actual practices underway at Insta360. Measurable changes include over 460,000 cumulative calls to Feishu Projects’ CLI, MCP, and AI-like services, and a defect analysis coverage rate reaching 65% (statistics blurred for confidentiality). These figures result from AI continuously invoking standardized processes within Insta360’s organization.
Event scene
Let one R&D cycle become the starting point for the next innovation. Insta360’s lesson is clear: standardize first, then intelligentize; organize complexity first, then let AI amplify organizational capability. By preserving the experience gained from each delivery, the next innovation won’t have to start from zero.
Turning 10x Individual Speed
into Organizational Capability
"Individuals have already gotten faster—why haven’t projects sped up accordingly?" This is a problem every enterprise faces. Ding Peng, Pre-sales Solutions Director at Feishu Projects, presented a contrast between individual and organizational efficiency: AI can accelerate personal tasks like coding, documentation, and analysis by up to 10x, yet at the organizational delivery level, improvements reach only 1.6x.
What truly matters isn’t the 10x or the 1.6x, but the gap in between—the "not equal to." His core judgment: "Models bring speed; work systems turn speed into organizational capability." The most costly time in projects is often spent on repetitive explanations, waiting for confirmation, context translation, and rework.
Event scene
Ding Peng shared three common pitfalls enterprises encounter: lack of clear exit criteria, turning development into QA for AI; long processes becoming black boxes, forcing restarts upon interruption; and multiple Agents lacking shared context, forcing humans to become "message buses." The problem isn’t with the models—it’s that work methods haven’t been redesigned.
Feishu Projects breaks down organizational efficiency into three components.
Harness acts like an engineering gate, converting explicit entry rules at each workflow step into reliable engineering rules executable by Agents. Workflow supports integration of existing enterprise-specific Agents, with Feishu Projects handling node triggering, state scheduling, exception recovery, and human intervention, while Agents perform specific executions. Shared context connects enterprises and necessary states, enabling collaboration across a delivery chain.
Ding Peng demonstrated common enterprise development scenarios to introduce the quick use of AI assistants and AI Workflows—how tasks are triggered, how workflows advance, and how results are written back into projects.
Live demo
Once workflows run smoothly, how we measure AI also changes. Managers should no longer focus solely on "how much AI is used" or "code generation rate," but on whether delivery cycles have shortened, human intervention reduced, and first-time success rates improved. When processes become measurable, managers can identify "super individuals," capture their best practices, and embed them into workflows—ultimately turning exceptional individual performance into replicable team capabilities.
Feishu Projects is evolving from a "data recording" system into an "action execution" system, connecting requirements, tasks, defects, iterations, releases, and documents, group chats, and meeting minutes within Feishu—enabling AI not just to answer questions, but to take the next step, freeing people to focus on judgment, decision-making, and creativity.
Event scene
In this era where "individuals can be very fast," Ding Peng emphasized: models bring speed; work systems turn speed into organizational capability. Thus, the next phase of competition hinges not only on who has stronger models, but who can integrate AI into processes, context, and collaboration modes fastest. Only then can one person’s 10x speed become a team’s sustained, stable, and replicable enterprise delivery capability.
Event scene
Tools solve efficiency problems; organizations and processes solve value problems. R&D management in the AI era ultimately won’t compete on whose model is stronger, but on who can integrate AI into their core processes faster. More and more enterprises are using Feishu Projects to transform AI from a personal coding tool into part of their organizational delivery capability.
At NIO, Agents are no longer isolated islands, but collaborative relays along core workflows. At Insta360, thousands of components behind a single camera are being reorganized into a unified hardware-software collaboration system. Meanwhile, Feishu and Feishu Projects are dedicated to converting individual speed into organizational efficiency—through Harness, AI Workflow, and shared context—ultimately forming a delivery-oriented workflow chain.
Bringing AI into core workflows is a shared challenge for enterprises today. What Feishu Projects offers is precisely the ability to "embed AI into nodes and collaboration relationships," transforming individual efficiency into organizational capability.
The true value of AI to enterprises lies not in boosting individual employees’ typing speed, but in changing the way entire organizations move forward.
If you're interested in use cases of Feishu Projects, click the link in the upper right corner to get a "customized solution."