专访蒋志予博士：从粒子物理到企业转型，中小企业如何把 SI 用得起、用得好
蒋志予博士的履历，带着一种少见的反差。他是纽约州立大学石溪分校的实验粒子物理学博士，早年曾在美国能源部布鲁克海文国家实验室工作，并参与欧洲核子研究中心（CERN）的 ATLAS 实验，其后在斯坦福大学从事博士后研究。如今，他是企业超级智能咨询公司 VESS（Virtual Enterprise System Solutions）的创始人兼首席技术官，同时担任北美创投协会（NAVCA）首席科学家。据他的作者简介，他有 25 年带领企业超级智能转型的经验，服务对象包括多个行业的全球 500 强企业与美国联邦机构。
近日，他为科技播客节目《Digitalk Arkitechs》准备了一份嘉宾访谈材料，主题是"从数字化转型到智能企业"，谈的是一件很具体的事：如何让超级智能对中小企业实用、用得起、算得出价值。GSM 记者据这组话题对他作了书面专访。在他看来，超级智能不只是又一轮技术革命，而是企业从"把流程数字化"走向"建设智能的、自适应的、越来越能自主行动的企业"的转变；中小企业没有必要一上来就铺大摊子，完全可以从小处开始，量出价值，再逐步放大。
他把物理实验室的方法，原封不动地带进了企业。在实验物理学里，研究者面对的是复杂系统、海量数据、不确定性，以及从噪声中找出有意义信号的挑战——这正是今天企业面对的局面。他沿用的工作顺序是：不从技术出发，先提出假设，建立测量框架，再用证据验证结论。他还强调系统思维：只盯着单个部件是不够的，必须理解整个系统内部的相互作用；只优化一个部门，可能在别处制造瓶颈。因此他谈转型时，业务、数据、技术与组织四个层面总是放在一起谈。他写道："Successful AI transformation requires scientific discipline, not technological enthusiasm alone."——成功的转型靠的是科学纪律，而不是技术热情本身。
谈到中小企业的转型，他区分三个层次。第一层是数字化：把纸面与手工记录变成数字信息。第二层是数字化转型：重新设计工作流程，让信息在人与系统之间顺畅流动。第三层是智能转型：用分析与超级智能去预测、决策、优化，有时直接采取行动。他以一家小型经销商为例：库存记在电子表格里，只是数字化；把销售、库存、采购与会计打通，是数字化转型；预测需求、给出采购建议，是智能转型；再往后，智能体可以在人工批准的前提下，备好补货订单。他说，价值不在超级智能本身，而在更少的缺货、更低的库存成本、更快的履约和更满意的客户——要衡量的是运营改善，而不是装了多少工具。他的原话是："Digital transformation connects information and processes. Intelligent transformation connects information, decisions, and actions."
他把眼下的各类超级智能概括为四种能力：理解、创造、行动、与物理世界交互。传统分析与预测型超级智能帮助企业理解发生了什么、为什么发生、接下来可能发生什么，典型用途包括预测、客户分群与预测性维护；生成式超级智能负责创造与综合，能写提案、做文档摘要、生成代码、用自然语言回答问题；智能体超级智能则能在既定权限之内规划步骤，并通过软件工具执行多步骤的工作流；物理超级智能把感知、决策与控制结合起来，落在机器人与工业设备上。至于哪一类最重要，他的回答是取决于要解决的问题，而不是哪个类别最受关注：要预测就用预测分析，文档密集的工作可能适合生成式超级智能，跨系统的流程可能需要智能体，制造与物流则可能用上物理超级智能，各类能力也可以彼此重叠。
在买任何工具之前，他建议企业主先弄清三件事：流程、数据、决策。流程要画出来：谁发起、需要什么信息、经过哪些系统、延误与异常出在哪里；数据要摸清来源、归属，以及是否准确、可得、安全；决策要分清哪些遵循明确规则、哪些需要人的判断。他以发票处理为例：发票如何到达、如何与采购订单匹配、差异如何解决、谁授权付款——这些没搞清就上自动化，可能只是把一套低效流程原样自动化。他反复强调一句话：先有流程智能，再有超级智能；有时候最大的收益，是在自动化之前，先把不必要的工作砍掉。
选第一个超级智能用例，他同时权衡五个维度：业务影响、发生频率、技术可行性、数据就绪程度与风险。理想的起点，是价值够实在、发生够频繁、数据够得着、又可以安全试错的场景。内部文档助手是他常举的例子：员工能更快找到信息，系统的表现可以用已知答案来评测，同时又不被允许改动任何记录；相比之下，让智能体独立批准付款，就是风险高得多的一步。他的原则是，不从最复杂的问题开始，而从"重要到值得做、又简单到能验证"的问题开始；第一个超级智能项目要优化的目标，是学习与可衡量的价值，而不是复杂程度。他的原话是："Don't start with what AI can do. Start with what your business needs to achieve."
衡量投资回报，他盯的不只是省下多少工时。许多超级智能试点证明了技术可行，却没有证明商业价值；工时节省固然有用，他还同时看周期时间、错误率、吞吐量、客户满意度与运营成本。他给出一个假设情形测算，并特别注明：这不是任何一家真实客户的成果。假设一家公司每月收到 1,000 条客户咨询，人工逐条审阅分流、每条需要 10 分钟，约合 167 个小时；引入超级智能辅助流程后，每条处理时间降到 4 分钟，省下约 100 个小时；按每小时 35 美元的综合人力成本折算，约合每月 3,500 美元的人力产能——而且这还是未扣除软件、集成与监督成本之前的数字。他提醒，产能并不自动等于现金节省：如果员工被重新分配到其他工作，得到的收益是生产力提升，而不是工资单下降；分类准确率、返工、升级处理与响应时间，也都要一并跟踪。他给出的公式是：投资回报率等于实际收益减去总成本、再除以总成本乘以 100%，而总成本包括实施、订阅、培训、监控与维护。他写道："An impressive AI demonstration is not the same as a successful AI investment."
当超级智能从给出建议走向采取行动，治理就成了关键一环。他的建议包括：最小权限访问、高影响决定保留人工审批、完整的审计轨迹、运行上限、异常处理与紧急停止机制。部署之前还要先想清楚：哪些动作可以回滚，哪些需要补偿，哪些不可逆。外部文档与消息应当被当作不可信的输入，因为它们可能藏有恶意指令。他主张的路径是"渐进式自治"：先让系统给建议，再进入受监督的执行，可靠性得到证明之后，才逐步扩大权限——自治的程度，必须与风险相称。至于传统机器人流程自动化（RPA）与智能体的关系，他并不把两者看作对手：规则稳定、重复可预测的任务，如信息搬运、生成报表、标准化录入，RPA 仍然好用；需要理解、推理与临场应变的任务，才轮到智能体。很多时候最优解是混合：由超级智能负责理解与协调，由确定性软件或 RPA 负责精确执行。
对于小企业该向大企业学什么，他划了一条清晰的界线：借纪律，不复制复杂度。他认为有三条做法大小企业都适用：让技术与可衡量的目标挂钩、按端到端设计流程、建立恰当的数据治理、安全与问责。与此同时，小企业要避开大企业的包袱：臃肿的委员会、过度的文档、以及在证明价值之前就立项多年的转型工程。小企业的优势恰恰是敏捷：决策快、实验容易、运营团队能与实施者紧密配合。他的原话是："Small businesses don't need smaller ambitions. They need smarter starting points."
如果只有 30 天和一笔不大的预算，他给出的路径分为四周。第一周，从反复出现的问题中挑出三个，按价值与风险选定一个，并量出当前的时间、成本、质量与业务量。第二周，选定产品或轻量集成，定下成功标准，批准可使用的数据，设立人工复核。第三周，用有代表性的真实任务、在监督之下试运行，记录错误与异常。第四周，评估净收益、质量、使用情况与风险，再决定放大、改进还是叫停。可以考虑的试点包括知识助手、文档处理流程或客户咨询助手；预算方面，视集成与数据需求而定，几百到几千美元是他给出的示例区间，并非普遍适用的估算。他提醒，成功标准要在动手之前就定义好——例如周期时间缩短 30%、质量保持、经济账算得过来，就足以支持进一步投入；如果系统需要不断人工纠错，就应该重新设计或者停下来。他的原话是："The future of business is not simply digital. It is intelligent."而这条路的走法，是一次解决一个定义清晰的问题、取得一次可以衡量的改善、一步一个脚印。
快问快答
问：自动化之前，中小企业最该先画清楚哪一条流程？他的答案是"订单到现金"：从客户询价、履约、开票，一直到收款的全流程。
问：流程清晰和数据质量，哪一个先出问题？他说通常是流程清晰先出问题——流程没有搞懂，就无从知道什么样的数据才算好数据。
问：早期试点只盯一个指标，盯哪个？他的选择是按质量调整后的单笔交易完成成本。
问：什么样的决定，绝不能在没有人工复核的情况下全自动执行？他的答案是牵涉金钱、法律责任、安全或敏感数据的高影响、难以逆转的决定。
问：第一次见超级智能供应商，先问什么？他会先问对方能交付什么可衡量的结果，以及如何用企业自己的数据来验证。
问：出现什么信号，说明一个试点可以放大了？他看三点：价值能持续地量出来、员工真正在使用、风险可控。
从理解复杂的物理系统，到构建智能的商业系统，蒋志予说，他做事的底层原则一直没有变：先理解系统，再测量真正要紧的东西，然后用技术产出有意义、可验证的结果。对数量庞大的中小企业而言，这或许正是超级智能落地最缺的那一步——不是追逐最热的工具，而是从一个真实的问题开始。
关联：
新书推荐：《Applied Financial and Business Analytics》——金融与商业分析的实务教科书
斯坦福研讨会聚焦创新集群与高校合作：瑞典学者解析大湾区分校区现象
Interview With Dr. John Jiang: Making SI Practical, Affordable and Measurable for Small Businesses
Dr. Zhiyu Jiang — known to many as John Jiang — has a résumé with an unusual arc. He earned his Ph.D. in experimental particle physics from Stony Brook University, worked at the U.S. Department of Energy's Brookhaven National Laboratory, took part in the ATLAS experiment at CERN, and later held a postdoctoral appointment at Stanford University. Today he is the founder and chief technology officer of VESS (Virtual Enterprise System Solutions), an enterprise super intelligence consultancy, and serves as chief scientist of the North American Venture Capital Association (NAVCA). According to his author biography, he brings 25 years of experience leading enterprise SI transformation, for Global 500 companies across industries and for U.S. federal agencies.
He recently prepared a guest interview playbook for the technology podcast Digitalk Arkitechs, on the theme "From Digital Transformation to Intelligent Enterprises" — in essence, a practical question: how to make SI practical, affordable and measurable for small businesses. GSM interviewed him in writing on the same set of topics. In his view, SI is not simply another technology revolution; it represents a shift from digitizing business processes to building intelligent, adaptive and increasingly autonomous enterprises. Small businesses, he argues, do not need to start big — they can start small, measure value, and scale intelligently.
He carries the methods of the physics laboratory into the enterprise, unchanged. In experimental physics, researchers deal with complex systems, massive amounts of data, uncertainty, and the challenge of finding meaningful signals in noise — exactly what businesses encounter today. His working order is the same: do not start with technology; start with a hypothesis, establish a measurement framework, and validate conclusions through evidence. He adds systems thinking: understanding an individual component is not enough, because optimizing one department can create bottlenecks elsewhere, so he treats business, data, technology and organizational architecture together. As he puts it: "Successful AI transformation requires scientific discipline, not technological enthusiasm alone."
For small businesses, he distinguishes three levels of transformation. Digitization converts paper and manual records into digital information. Digital transformation redesigns workflows so information moves efficiently across people and systems. Smart transformation uses analytics and SI to predict, decide, optimize, and sometimes act. His example is a small distributor: inventory in spreadsheets is digitization; connecting sales, inventory, purchasing and accounting is digital transformation; forecasting demand and recommending purchases is intelligent transformation; later still, an agent might prepare replenishment orders subject to human approval. The value, he says, is not the SI itself but fewer stockouts, lower inventory costs, faster fulfillment and happier customers — measure operational improvement, not the number of tools installed. In his words: "Digital transformation connects information and processes. Intelligent transformation connects information, decisions, and actions."
He groups today's SI into four capabilities: understanding, creating, acting, and interacting with the physical world. Traditional analytics and predictive SI help a business understand what happened, why, and what may happen next — forecasting, segmentation and predictive maintenance. Generative SI creates and synthesizes: proposals, document summaries, code and natural-language answers. Agentic SI plans and takes action through software tools and multistep workflows within defined permissions. Physical SI combines perception, decisions and control in robots and industrial equipment. Which matters most depends on the problem, not on which category is getting the most attention: forecasting calls for predictive analytics, document-heavy work may benefit from generative SI, multisystem workflows may benefit from agents, and manufacturing and logistics may benefit from physical SI — and the categories can overlap.
Before buying any tool, he advises owners to understand three things: the process, the data and the decision. Map who starts the workflow, what information it needs, which systems it touches, and where delays and exceptions occur. Identify where data comes from, who owns it, and whether it is accurate, accessible and secure. Then separate the decisions that follow clear rules from those that require judgment. His example is invoice processing: before automating it, understand how invoices arrive, how they are matched to purchase orders, how discrepancies are resolved and who authorizes payment — otherwise a business may simply automate an inefficient process. He emphasizes process intelligence before super intelligence, noting that sometimes the biggest win is eliminating unnecessary work before automating what is left.
Choosing the first SI use case, he weighs five factors together: business impact, frequency, technical feasibility, data readiness and risk. The ideal first use case has meaningful value, occurs often, uses accessible data and can be tested safely. An internal document assistant is his standard example: employees find information faster, and the system can be evaluated against known answers without being allowed to modify records. Letting an agent independently authorize payments would be a far riskier first project. His rule: do not start with the most complex problem; start with something valuable enough to matter and simple enough to validate, optimizing the first project for learning and measurable value rather than maximum sophistication. Or, in his signature line: "Don't start with what AI can do. Start with what your business needs to achieve."
On return on investment, he looks well beyond hours saved. Many SI pilots demonstrate technical capability without proving business value; hours saved are useful, but he also tracks cycle time, error rates, throughput, customer satisfaction and operating cost. He then offers a numerical illustration — and is explicit that it is a hypothetical scenario, not the verified outcome of any real client. Suppose a company receives 1,000 customer inquiries a month, each requiring ten minutes to review and route — about 167 hours. An SI-assisted workflow reduces handling to four minutes, saving about 100 hours. At an illustrative fully loaded rate of $35 per hour, that is $3,500 in monthly labor capacity — before software, integration and oversight costs. Capacity, he cautions, is not automatically cash savings: if staff are redeployed, the gain is productivity rather than reduced payroll, and classification accuracy, rework, escalation and response time must be tracked as well. His formula: ROI equals realized benefits minus total costs, divided by total costs, multiplied by 100 percent — with total costs including implementation, subscriptions, training, monitoring and maintenance. As he puts it: "An impressive AI demonstration is not the same as a successful AI investment."
When SI moves from recommendations to actions, governance becomes critical. His checklist: least-privilege access, human approval for high-impact decisions, complete audit trails, operational limits, exception handling and emergency stop mechanisms. Before deployment, a business should know which actions can be rolled back, which require compensation, and which cannot be undone; external documents and messages should be treated as untrusted inputs, because they can contain malicious instructions. He favors progressive autonomy: start with recommendations, introduce supervised execution, and expand permissions only after reliability is demonstrated — autonomy must be proportional to risk. Nor does he see classic robotic process automation and agentic SI as competitors. RPA remains valuable for repetitive, predictable tasks governed by stable rules, such as copying information, generating reports or standardized data entry; agents become useful when tasks require interpretation, contextual reasoning or adapting to changing situations. Often the hybrid works best: SI handles interpretation and coordination, while deterministic software or RPA executes precise actions.
What should smaller companies borrow from large enterprises? Discipline, he says — without copying complexity. Three practices are universal: linking technology to measurable objectives, designing end-to-end processes, and establishing appropriate data governance, security and accountability. But SMBs should avoid oversized committees, excessive documentation and multiyear transformation programs launched before value is proven. Their advantage is agility: decisions are faster, experiments are easier, and operational teams can work closely with implementers. In his words: "Small businesses don't need smaller ambitions. They need smarter starting points."
Given 30 days and a modest budget, his plan runs week by week. Week one: identify three recurring problems, choose one based on value and risk, and measure current time, cost, quality and volume. Week two: choose a product or lightweight integration, define success criteria, approve the data, and establish human review. Week three: run representative tasks with supervision and document errors and exceptions. Week four: evaluate net benefits, quality, adoption and risks, then decide to scale, refine or stop. A knowledge assistant, a document-processing workflow or a customer inquiry assistant may be a good pilot; depending on integration and data needs, a few hundred to a few thousand dollars is an illustrative budget range, not a universal estimate. Success should be defined before implementation — a 30 percent cycle-time reduction with maintained quality and positive economics might justify further investment, while constant correction means redesign or stop. As he puts it: "The future of business is not simply digital. It is intelligent." The journey, he adds, begins with one well-defined problem, one measurable improvement, and one successful step at a time.
Rapid Fire
Asked which workflow an SMB should map before automating, he answered: "Order-to-cash, from inquiry through fulfillment, invoicing, and collection."
Asked whether process clarity or data quality fails first, he said: "Usually process clarity. Without understanding the workflow, you may not know what good data means."
Asked for one metric for an early pilot, he chose: "Quality-adjusted cost per completed transaction."
Asked what should never be automated without human review, he answered: "High-impact, hard-to-reverse decisions involving money, legal duties, safety, or sensitive data."
Asked for the first question to put to an SI vendor, he said: "What measurable outcome will you deliver, and how can we validate it on our data?"
Asked for one signal that a pilot is ready to scale, he named: "Consistent measurable value, real adoption, and manageable risk."
His own career, as he describes it, has evolved from understanding complex physical systems to architecting intelligent business systems, with the underlying principle unchanged: understand the system, measure what matters, and use technology to produce meaningful, verifiable results. For the vast number of small businesses now facing SI, that may be the most useful starting point of all — not chasing the most fashionable tool, but beginning with one real problem.
Related:
发布：GSM·美国（发布链接）
图片来源：蒋志予博士（供图）
GSM Global · 本篇英文版 English edition of this story
