AI + 数据预测 2026

of enterprise and a new kind of worker The Year of Agents and Ecosystems It seems like every new year is labeled a “breakthrough year for enterprise AI.”And pretty much every year is. Just looking at the recent past, 2024 was theyear businesses got creative, experimenting with explosively hot generativeAI technologies to varying levels of initial success. By 2025, the focus wason ROI and proving that all the hype and FOMO wasn’t getting in the way ofsolid business value. (Countless surveys, including one published by Snowflake,found that many AI-forward organizations were measuring solid return on theirgen AI investments.)So here comes 2026, and another evolution in enterprise-class artificial This year’s report considers the ever-shifting state of AI technology, the outlookand opportunities for enterprise adoption, and the challenges, particularlyaround security and data governance. It draws from in-depth conversationswith more than a dozen leaders and experts at Snowflake. Click through thesummaries for an expanded view of these topics: Choke points and paths of progress Feedback loops will improve agents intelligence. Two things, broadly, are on the horizon: extending ROI from thelevel of individual projects to comprehensive, strategic AI ecosystems, andagentic AI. The former reflects a maturing of the enterprise itself, as the CDOand other leaders integrate individual successes into a data and AI strategy thathelps every team and contributor perform better. The latter reflects a maturingof the technology, as large language models evolve into large reasoning models,and as these LRMs become increasingly capable of reliably taking action withless human supervision. EUregulationsmayadvance,nothinder,innovativeAIinitiativesAdominantAIprotocolwillfacilitateagenticdevelopmentandpreventvendor lock-inOpen-sourcefoundationmodelswillbreaktheholdofthehandfulofgiantsAgentswillstartsmall,stitchingtogether‘micro-agents’intoeffectivetoolsContextwindowsandmemorywillbethekeystobetterAIagentsPostgreswillbeafoundationaltechnologyforagenticAI AI drives cybersecurity (for better and worse)Agentsascyberweaponswillhelphumanattackersrefineandscaletheir Enterprise uptake attackswithinayearCybercriminalswilluseincreasinglysophisticated‘darkAIs’Inthreeyears,AIagentsandtoolswillfinallyclosethesecuritytalentgap DatastrategywilldetermineAIreadiness—andAIoutcomesWhat’snotinyourdatawilllimitagenticdecision-making The AI-augmented workforceWorkerswill(haveto)masterhuman-AIcollaborationandcommunication AIcodingassistantsshouldmakedevelopers33%moreproductiveby2027Everyonewill(haveto)becomeastrategicthinkerWhatjunior engineers?Workerswill(needto)developcross-functionalandorchestrationskillsForworkersandtheenterprisetothrive,leadersmustprovidecontinuous,contextuallearningcapabilitiesAI-createdapplicationswillbethenewgeneral-purposespreadsheetHumanswillremainintheloopasinterpretersandqualitycontrollers GenerativeAIwillaccelerateandimprovehumancreativity—butitwill takesmart,skillfulapproachestoavoiditbeingacrutch Retail:GenAIisflippingcustomer360onitshead Financialservices:Shiftingbacktoadata-firstmindsetManufacturing:AIadoptiontakescenterstage The AI Landscape In 2025, trying to understand the state of AI was an exercise in industrial-scalecognitive dissonance. Last summer, Anthropic CEO Dario Amodei said thatgenerative AI technologies were developing so quickly that massive layoffsacross the global workforce were nigh. About two weeks later, Apple reportedthat its research with large reasoning models showed that these ostensiblybrighter cousins of LLMs couldn’t reason their way through a puzzle that mightbamboozle a high school sophomore — which might cast doubt on the near-termpotential of agentic AI.So which is it? Is the AI industry flying down the freeway at 90 miles an hour,or is it (and are we) stuck in stop-and-go traffic while obnoxious radio adsimplore us toACT NOW! DON’T MISS OUT! on this once-in-a-lifetime opportunity?The general consensus at Snowflake is, “a little bit of both.”“It’s like my commute home from the office,” says Anupam Datta, AI ResearchLead at Snowflake. “It goes pretty fast, but there are choke points. That’s AI inthe coming years: In general, it works well, but there are choke points we haveto invest in to get AI agents to scale.” Choke points and paths of progressDatta,whoco-leadsSnowflake’sAIresearchteam,breaksthe“Enterprisesarealreadydemandingthatreliabilitybequantified, chokepointsslowingtheadoptionofLLMandLRMsystemsintoathree-partchallenge.“Tomakeagentsmorereliableandtrustworthy,threefactorsaregoingtobeatplay:evaluation,prompt-basedandpost-trainingoptimization,andAIsafetyandsecurity.”Evaluation,orvalidationofamodel’ssuccess,isperhapsthebecausethat’swhatittakestosucceedintheenterprise,”notesCEOSridharRamaswamy.“Imean,thereisexactlyoneanswerto‘HowmuchmoneydidSnowflakemakeyesterday?’That’snotamatterofdoubtoropinion.” key2026challengeonthemindofSnowflakeDirectorofAIInfraMonaAttariyan.“Itw