{"slug": "where-the-enterprise-ai-advantage-comes-from", "title": "Where the enterprise AI advantage comes from", "summary": "Alibaba.com's CommerceAgentBench, an open-source benchmark of 107 end-to-end e-commerce tasks, found that the strongest AI agent tested completed 61.7% of real-world commerce tasks, with different models leading on different workflows such as requests for quotation, claims settlement, and product publishing. In one estimate covering 44 tasks, Alibaba's AI business agent Accio cost $1.72 in total tokens versus $3.79 for Codex and $3.91 for Claude Code, which Alibaba cites as evidence that enterprises should match model capability to each workflow's requirements. Alibaba says the results make model selection a workflow-level decision rather than a single-model choice.", "body_md": "The strongest [AI](https://www.fastcompany.com/section/artificial-intelligence) agent we recently tested completed 61.7% of 107 real-world commerce tasks.\n\nThat number captures where agentic AI stands today. A system that can complete six out of every 10 complicated business tasks can already take meaningful work off someone’s plate. But for businesses, model choice is only part of the decision. The larger task is matching each workflow with a system that can complete it reliably at an operating cost that makes sense.\n\nFor much of the AI boom, progress has been measured through model intelligence. Bigger and more capable models generally meant better results. As AI moves into everyday operations, however, companies also have to account for efficiency. Different tasks demand different levels of reasoning and computing power, and using maximum capability can add cost without improving the outcome.\n\nThis is one dimension of precision delegation: matching the capability assigned to the workflow with what the work requires.\n\nToday, business outcomes depend on the entire system around a model: how work is structured, what tools the model can access, and the context available when it makes a decision.\n\nThis is especially true in specialized domains. At Alibaba.com, 27 years of commerce experience and real-world business data give our AI systems context on how commercial workflows operate. Clearly defined e-commerce tasks may be well-suited to smaller, lightweight models, while more complex work requires greater reasoning capacity.\n\nThat changes the enterprise AI equation. Each workflow needs enough capability to meet its required standard at a sustainable cost.\n\nWe saw this while building CommerceAgentBench, an open-source benchmark based on e-commerce operations. We distilled real business activity into 107 end-to-end commercial tasks and graded the final outcome: Was the listing published correctly? Was the shipment booked according to specifications? Was the purchase order created and dispatched?\n\nDifferent tasks favored different models. The model that led on requests for quotation and market research trailed the one that performed best on claims settlement and listing compliance. A third led on publishing products and handling returns.\n\nThose results make model selection a workflow-level decision.\n\nA complicated business request may combine market research, spreadsheet analysis, supplier comparison, and purchase order creation. Each step places different demands on reasoning, speed, or data handling.\n\nIn Accio, Alibaba’s AI-powered business agent, we continuously evaluate models across different kinds of commercial work. A complex request can be broken into individual steps, with each routed based on its difficulty and requirements, including quality, speed, price, and data processing. The result is closer to managing a team of specialists than asking one person to do everything.\n\nThat also changes the economics. In one estimate covering a set of 44 tasks, Accio’s total token cost was $1.72, compared with $3.79 for Codex and $3.91 for Claude Code.\n\nAs the number of high-quality models grows, the ability to evaluate their real-world performance continuously and assign work accordingly will become an increasingly valuable enterprise capability.\n\nChoosing the right model can lower the cost of each step. The next opportunity is reducing how many steps are needed in the first place.\n\nAI agents repeatedly process context, call tools, and pass information through a workflow. Poorly designed systems may misread information they already have or repeat computation that adds nothing to the final result.\n\nA more efficient system can reuse previous computation and compress context when the full history is unnecessary. Specialized agents can also coordinate work so that each component processes only the information it needs.\n\nLower model prices will make AI cheaper to operate. Better orchestration can reduce the amount of computing required to complete the initial work.\n\nThe objective is a reliable business outcome using only the resources the task requires.\n\nThe effects are especially visible in small businesses, where a founder may personally handle research, sourcing, merchandising, and operations.\n\n[Harrison Nott](https://www.linkedin.com/posts/alibaba-com_a-british-genius-owns-his-destiny-at-15-activity-7396944124769329152-YRY6/) started with an idea for a cooling towel after playing sports. He used Alibaba.com to develop the product with manufacturers, then Accio to identify demand for pet cooling products and source them. At 16, he has turned that initial idea into a broader cooling-products business that has generated $1 million.\n\nAI gives entrepreneurs access to capabilities that previously required more people, more time, or more specialized expertise. But that leverage becomes much more valuable when those capabilities are reliable and affordable enough to use repeatedly.\n\nCost, performance, and efficiency are only part of the precision delegation equation. Companies also need to decide how much authority an agent should receive.\n\nA weak draft of a social post carries one kind of consequence. An incorrect supplier payment carries another.\n\nRoutine work with consistently strong performance may run with minimal intervention, while higher-risk decisions may still require approval. This is where precision delegation becomes important: assigning work and authority workflow by workflow based on demonstrated performance, cost, and the consequences of getting the decision wrong.\n\nThe stakes will rise as agents connect directly to payments, inventory, logistics, and other operational systems. At that point, choosing which model handles the work is only the beginning. Companies also need evidence about how reliably the system performs and clear rules governing what it can do on their behalf.\n\nThe next phase of enterprise AI will favor companies that make those decisions at the workflow level. The strongest systems will allocate computing power according to what the job requires and expand an agent’s authority as its performance earns greater trust.\n\n*Kuo Zhang is president of Alibaba.com.*", "url": "https://wpnews.pro/news/where-the-enterprise-ai-advantage-comes-from", "canonical_source": "https://www.fastcompany.com/91609789/where-the-enterprise-ai-advantage-comes-from", "published_at": "2026-09-22 12:00:00+00:00", "updated_at": "2026-09-22 12:23:40.601326+00:00", "lang": "en", "topics": ["ai-agents", "ai-products", "ai-tools", "ai-research"], "entities": ["Alibaba.com", "CommerceAgentBench", "Accio", "Codex", "Claude Code"], "alternates": {"html": "https://wpnews.pro/news/where-the-enterprise-ai-advantage-comes-from", "markdown": "https://wpnews.pro/news/where-the-enterprise-ai-advantage-comes-from.md", "text": "https://wpnews.pro/news/where-the-enterprise-ai-advantage-comes-from.txt", "jsonld": "https://wpnews.pro/news/where-the-enterprise-ai-advantage-comes-from.jsonld"}}