潜力新秀候选名单

ⓘ 关于这套评估框架

每一家公司都由侦察员依据 AI 应用业务中会复利累积的五件事读过——客户关系、专有数据、评测体系、工作流以及领域专用模型——每一项都分为完全具备 / 部分具备 / 不具备 / 未知，各自附上引用与日期。

除这五项外，每张卡片还标注AI 落地深度——分为深 / 中 / 浅 / 未知，引用方式相同。它是并列的另一条轴，而不是第六项：五项问的是若这套论点成立，价值是否会归属于这家公司；而这一条问的是它究竟有没有在运转。它从不计入“五项中具备几项”，两者也从不相加——一家公司可以既不拥有数据也不拥有模型，却仍在运行全世界最深入的部署之一。凡是尚未做出判断的，卡片会直说，而不是猜一个出来。

按垂直领域分组，组内按名称字母排序。这些计数是一种格局，既不是排名也不是分数：五项俱备与其中三项出色，是性质不同的两件事，没有任何单一数字能表达。未知的意思是没人找到证据，而不是答案为否。无权重——本页任何内容都不会进入信号，也都不是对回报的预测。

198 家公司中已覆盖 13 家。侦察是人工进行、由条件驱动的，因此另外 185 家没有档案——那是研究的缺席，绝不是“五项全无”的判定。

Coding

依据以下条件侦察：a company that simultaneously owns the customer relationship, the data, the evals, the workflow, and a domain-specific model

GitLab (GTLB)

编辑估算截至 2026-08-05

2 项完全具备 · 2 项部分具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：中

GitLab's Duo Agent Platform, generally available since January 2026, finished its first full quarter of consumption at a paid consumption run rate of nearly $20 million, a figure that includes minimum usage commitments made before that launch plus on-demand credits. It attached to four of the quarter's top ten deals and added more net new ARR than Duo Pro and Duo Enterprise combined had in any prior quarter. One disclosed account shows the intensity: a top-ten US bank whose developers averaged 1.5 hours saved per task is in active rollout across hundreds of developers, and the customer's own plan points to an active user base nearly 20 times larger at full deployment. Management assumes no material revenue contribution in fiscal 2027 and frames the year as converting pilots to production. 来源：s204.q4cdn.com

GitLab owns the delivery workflow and the customer relationship, rents the model, and is barred by its own policy from compounding the one data asset that would otherwise be its moat. It is the system of record for the software lifecycle, agents run inside it, and GitLab Credits meter that agent activity, while the models are entirely third-party and customer tenant data is walled off from training. Evaluation is genuinely first-party but leans on an external tracking platform and an industry benchmark for scoring.

客户关系	完全具备	GitLab sells "through a direct sales organization, a self-service web purchasing experience, and a global partner ecosystem", with partners "designed to complement direct sales efforts." Free tiers and trials feed the funnel, and the installed base is named: 1,456 customers at $100k+ ARR, up from 1,229, net retention of 118%, and "no individual customers whose revenue represented more than 10% of total revenue" in FY2026, FY2025 or FY2024. Customer counts track each end customer separately even when sold through a channel. The counterweight is intermediation: at January 31, 2026 two channel partners held 17% and 13% of accounts receivable, in a book where over 70% of ARR is public sector and enterprise, and the filing carries a channel-partner risk factor. 来源：sec.gov
专有数据	部分具备	GitLab has privileged inference-time access to an unusually complete software-lifecycle corpus, but it is customer tenant data and GitLab has foreclosed compounding it: "GitLab does not train generative AI models" and "For GitLab Duo requests, GitLab has a zero data retention policy with Fireworks AI." Some exhaust is kept: Duo Chat and Duo Agent Platform "retain chat and workflow history", "GitLab retains trace data" under expanded logging, telemetry covers prompt lengths, model, editor and accept/reject counts, and since 18.9.1 an opt-in setting logs "Full prompt and response text" and "Tool calls and their results" — "not for training AI models." Prompt caching and "limited vendor-side data retention" apply to certain Anthropic and OpenAI models. 来源：docs.gitlab.com
评测体系	部分具备	GitLab runs a public, first-party evaluation harness. The Centralized Evaluation Framework "is a command-line interface (CLI) framework designed to provide a centralized solution for evaluating AI features at GitLab", maintained by its AI Model Validation and Research group and covering "GitLab AI features (such as Duo Chat, Code Suggestions, etc.)". Its outputs are performance metrics and comparison reports letting developers "test and validate prompt engineering changes"; that these set per-feature defaults is an inference, as the docs say only "GitLab might update default models to optimize performance." It stays below owns because "CEF relies on LangSmith for evaluation tracking and experiment management" and agentic scoring uses an external benchmark via "one-click SWE-bench evaluations." 来源：gitlab.com
工作流	完全具备	GitLab "is a technology company and its primary offering is 'GitLab', an intelligent orchestration platform for DevSecOps delivered as a single application", and the issue, merge request, pipeline and vulnerability record live in its system of record. The Duo Agent Platform, "Generally available in January 2026", puts agents inside that workflow: "teams orchestrate AI agents that execute tasks autonomously across planning, development, security, and deployment. This eliminates the need for manual handoffs between software lifecycle stages." Monetization meters the same surface: "Usage-based pricing (GitLab Credits) applies specifically to GitLab Duo Agent Platform capabilities." Authoring is the limit — code is written in "third-party code editors and integrated development environments, or IDEs." 来源：sec.gov
领域专用模型	不具备	Every model behind GitLab Duo is third-party: Anthropic (Claude Sonnet 4.5, Sonnet 4.6, Sonnet 3, Haiku 3, Haiku 4.5, several via Google Vertex), Google (Gemini 2.5 Flash Vertex), Mistral Codestral 25.01 and 25.08, and OpenAI — the guidance recommends starting with "Claude Haiku 4.5 or GPT-5.4 Mini." Defaults are Claude Sonnet 4.6 Vertex for code generation and chat, Claude Sonnet 4.5 Vertex for Code Review, and Codestral 25.08 Fireworks for code completion. No GitLab-authored model appears and there is no fine-tuning claim: "GitLab does not train generative AI models." The 10-K books the dependency as a risk — "We rely on third-party vendors for the provision of the AI models which power many of our AI features." GitLab owns routing, per-feature model switching (GA in 18.5) and prompt engineering. 来源：docs.gitlab.com

Defense

依据以下条件侦察：a company that simultaneously owns the customer relationship, the data, the evals, the workflow, and a domain-specific model

Palantir Technologies (PLTR)

编辑估算截至 2026-08-05

2 项完全具备 · 1 项部分具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Palantir's quarter ended June 30, 2026 shows 1,049 trailing-twelve-month customers, up from 849 a year earlier, on quarterly revenue of $1.94 billion, a 93% increase, with US commercial revenue of $764 million up 149%. The depth sits in expansion rather than logo count: average trailing-twelve-month revenue across the top twenty customers reached $124 million, up 67%, and $407 million of the $495 million commercial revenue increase came from customers already on the books at December 31, 2025. AIP, deployed since 2023, adds an agent and automation toolchain and an evaluations framework for governing AI workflows in production atop the Foundry Ontology. The installed base stays narrow at roughly a thousand customers, and remaining performance obligations stood at $4.9 billion. 来源：sec.gov

Palantir owns the customer relationship and the workflow, holds a partial position on evaluations, and deliberately owns neither the data nor the models. It sells direct to 954 customers and positions its platforms as the central operating system for their daily work, while acting as a data processor over corpora walled off per engagement and routing to third-party or open-source LLMs. The edge is the Ontology, permissioning and deployment layer between those customer-owned inputs and the operational decisions.

客户关系	完全具备	Palantir sells through a "Direct Sales Force": the FY2025 10-K says it has "invested and may continue to invest in an account-based sales force to identify and capture new customers" and that "We embed directly with customers across numerous industries", with AIP bootcamps that "deliver real workflows on actual customer data in days". It reported 954 customers as of December 31, 2025, a customer being "an organization from which we have recognized revenue during the trailing twelve-month period". Two caveats sit against full first-party control: Japan and Korea joint ventures that "limit our ability to independently sell our platforms", and "channel sales relationships and strategic alliances with various global system integrators" whose partners may take on "some portion of sales, marketing, implementation services". 来源：sec.gov
专有数据	不具备	Palantir disclaims ownership of customer data: "With regards to customer data, we act as a data processor, not a data controller", "We do not collect, store, or sell personal data", and "We don't use personal data to train proprietary AI or machine learning models to share or resell to other customers." That post dates from 11 November 2020, but the current AIP security documentation says the same today: Palantir "secures strict technical and contractual guarantees from third-party-hosted model service providers to ensure that no customer data submitted in prompts or contained in completions is used for model training". The FY2025 10-K is consistent — the platforms "integrate our customers' data" and deploy "in customer networks, on-premises at customer sites, on edge devices". Any retraining stays in the customer's deployment. 来源：blog.palantir.com
评测体系	部分具备	Palantir ships a real first-party evaluation harness, named in the 10-K as "a broad evaluations framework for governing AI workflows in production", but the criteria are authored per customer. Users "Set up tests manually to create an empty evaluation suite" in a test editor "where you can configure inputs for each test case", and the shipped evaluator is generic: teams "use the built-in Exact string match evaluator", while "depending on the nature of your function, you may need to use other evaluators or write custom ones." AIP can bootstrap a starting point — "select Generate evals, if your Logic function is eligible, to have AIP bootstrap useful tests and evaluators for you" — but that is scoped to one customer's function rather than a published domain standard. 来源：palantir.com
工作流	完全具备	The Q2 FY2026 10-Q, for the quarterly period ended June 30, 2026, states that "The Company builds and deploys software platforms that serve as the central operating systems for its customers", and the FY2025 10-K says "our platforms have become central operating systems for entire industries and sectors". The product description backs this at the level of daily work: "data engineers can integrate new data sources, analysts can clean and transform data, data scientists can write models, business users can conduct daily workflows, and senior leaders can make critical decisions". Gotham "facilitates the hand-off between analysts and operational users, helping operators plan and execute real-world mission plans." "Central operating system" is the filing's own characterisation, so the band rests on that task list rather than the phrase. 来源：sec.gov
领域专用模型	不具备	AIP routes to models Palantir did not train: it "supports a wide range of LLMs (large language models) and text embedding models from leading providers, including xAI, OpenAI, Anthropic, Meta, and Google", covering Grok, GPT, Claude and Gemini alongside an "Open source (Palantir-hosted)" tier including Llama 3.1/3.3, Mixtral 8x7B, Whisper Large V3 and Document Information Extraction. The 10-K describes AIP as providing "secure connectivity to third-party-provided large language models (“LLMs”), a development toolchain for building AI-powered agents and automations", with no reference to fine-tuning or proprietary training data. Customers can also connect "their own LLMs or accounts". The value sits in the Ontology, permissioning and deployment layer rather than in weights. 来源：palantir.com

Health

依据以下条件侦察：a company that simultaneously owns the customer relationship, the data, the evals, the workflow, and a domain-specific model

Abridge

未上市编辑估算截至 2026-08-05

2 项完全具备 · 2 项部分具备 · 1 项未知 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Abridge is live at more than 300 health systems and supports over 100 million clinical conversations a year, at partners collectively serving roughly 250 million patients, with note generation in 28-plus languages. Named enterprise deployments include Northwestern Medicine, Johns Hopkins Health System, Emory Healthcare, UCHealth and Sutter Health, where 78% of clinicians report improved work satisfaction — daily encounter workflow, not an occasional assist. At Reid Health, nursing use accompanied a vacancy rate falling from 18% to 8.6% and a 70% cut in incidental overtime. The June 2026 platform launch extends the same ambient capture beyond ambulatory documentation into nursing, inpatient and revenue-cycle workflows, surfaces it had only begun to address. 来源：abridge.com

Owns the enterprise relationship and the whole encounter-to-billing chain at 300+ health systems, but the corpus belongs to those customers and the model is an adaptation of a third-party base. Evals are undocumented — notable for a product whose output enters the medical record.

客户关系	完全具备	Contracts with the health system itself — Northwell Health, the largest not-for-profit system in the Northeast, across 28 hospitals and 1,000+ outpatient facilities — rather than selling seats to individual clinicians; 300+ health systems in total. 来源：abridge.com
专有数据	部分具备	Trains on "de-identified data" from roughly 100 million conversations a year, but no proprietary corpus or exclusive right to it is described — the conversations originate in customer health systems, so the volume is real while the ownership is not established. 来源：abridge.com
评测体系	未知	Looked for a published validation methodology, an accuracy benchmark or a model card and found none; the cited evidence of quality is a KLAS market-survey ranking, which measures customer sentiment rather than clinical accuracy. Absence of evidence, not evidence of absence. 来源：abridge.com
工作流	完全具备	The chain from encounter to billable record runs inside the product — pre-visit preparation, in-visit documentation and decision support, post-visit note generation, billing-code assignment, nurse documentation and clinical-trial screening, across 55+ specialties and 28+ languages. 来源：abridge.com
领域专用模型	部分具备	Building a foundation model for clinical conversations on NVIDIA Nemotron, trained on NVIDIA Blackwell infrastructure — the Nemotron base layer is a third party's open model and Abridge's adaptation sits on top, so the weights are customised rather than owned end to end. 来源：abridge.com

Hippocratic AI

未上市编辑估算截至 2026-08-05

2 项完全具备 · 3 项未知 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Hippocratic AI's April 2026 Polaris 5.0 release says the model was built on more than 180 million real-world patient interactions and validated by over 7,500 U.S.-licensed clinicians. The AI is the product rather than a feature: voice agents run post-discharge, medication-adherence and screening outreach calls as standing workflow at Universal Health Services, WellSpan Health, Cincinnati Children's, University Hospitals and Sheba Medical Center, with customer case studies citing a 30% readmission reduction and an 89% connection rate. The agent catalogue exceeds 1,000 prebuilt agents. Headroom remains visible: University Hospitals describes initial pharmacy-related deployments ahead of wider rollout, and the life-sciences arm dates only to a January 2026 acquisition. 来源：hippocraticai.com

Owns the customer and performs the patient interaction itself at real volume, but three of five dimensions are undocumented in the cited source. Included at that grain deliberately: the rubric records what is evidenced, and a company can carry serious traction while most of the stack stays unestablished.

客户关系	完全具备	Contracts directly with 50+ large health systems, payors and pharma clients across six countries, including named providers such as Cleveland Clinic, Northwestern Medicine, Ochsner Health and Moffitt Cancer Center. 来源：hippocraticai.com
专有数据	未知	Looked for a proprietary corpus or an exclusive data right and found none described; 115 million clinical patient interactions are reported as a volume of work performed, which is not the same as a claimed data asset. Absence of evidence, not evidence of absence. 来源：hippocraticai.com
评测体系	未知	The announcement names a "Polaris Safety Constellation Architecture" it will keep investing in, but does not define it, describe its methodology, or establish it as an evaluation system Hippocratic owns — so the band cannot be raised on the strength of the name alone. 来源：hippocraticai.com
工作流	完全具备	The patient interaction itself happens inside the product: 115 million clinical patient interactions completed in the 15 months since commercialisation, rather than the product handing a task back to staff to perform. 来源：hippocraticai.com
领域专用模型	未知	The announcement does not state whether the company trains its own models or builds on a third-party base. Looked for a model description in the cited source and found none. 来源：hippocraticai.com

OpenEvidence

未上市编辑估算截至 2026-08-05

2 项完全具备 · 2 项部分具备 · 1 项未知 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

OpenEvidence is used by more than 40% of physicians in the United States, and founder Daniel Nadler describes it as the most widely used AI platform among American doctors — a point-of-care tool for high-stakes clinical decisions, not an occasional assist. The company says it has gathered hundreds of millions of real-world clinical consultations from verified physicians, and passed $100 million in annualized revenue in 2025 on advertising rather than subscriptions, with 95% of new users hearing about it from another physician. Depth is concentrated in one clinician type in one country: roughly six in ten U.S. physicians sit outside that base, and Nadler points to small practices without IT departments or software budgets as where most American care happens. 来源：cnbc.com

Owns the clinician relationship outright and trains its own vertical model, but rents the corpus that model is trained on and does not own the surface where care is recorded. The interesting asymmetry in the set: enormous reach on top of a licensed foundation.

客户关系	完全具备	Reaches the clinician directly and gates access itself — verification by NPI scanning or hospital email — with 40%+ of U.S. physicians using it daily and 65,000+ new verified registrations a month. No health-system contract stands between the product and its user. 来源：sacra.com
专有数据	部分具备	The corpus is LICENSED, not owned: NEJM and JAMA journals, Wiley biomedical literature, Cochrane systematic reviews and specialty-society content, with no exclusivity terms specified. A renewal it does not control sits under the whole product. 来源：sacra.com
评测体系	未知	Looked for a published benchmark or evaluation methodology and found none; deterministic citation linking is a retrieval mechanism that makes answers traceable, not a standard for judging whether they are clinically right. 来源：sacra.com
工作流	部分具备	Consulted during care rather than being where care is recorded — the report names Epic and Oracle-Cerner as the EHR incumbents it sits against, so the clinical system of record remains someone else's surface. 来源：sacra.com
领域专用模型	完全具备	A vertical LLM trained exclusively on licensed medical texts, deliberately avoiding public-internet contamination, rather than a wrapper over a general frontier model. 来源：sacra.com

Tempus AI (TEM)

编辑估算截至 2026-08-05

4 项完全具备 · 1 项部分具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Tempus runs its AI in routine clinical practice, not pilots. Its FY2025 10-K reports connections to more than 55% of all oncologists practicing in the United States across over 700 data connections, about 367,500 clinical oncology samples sequenced in 2025 against roughly 31,000 in 2018, and revenue of $1,271.8 million. The Tempus Next care-gap engine covers 15 disease areas, is live in around 150 hospitals and screens more than 60,000 patients a month; its tumor-origin algorithm has been in clinical use since 2021; and the models draw on 8.6 million de-identified patient records and over 450 petabytes of data. Tempus notes that algorithmic diagnostics still represent only a small proportion of the diagnostics market, leaving Algos the least penetrated part of the platform. 来源：sec.gov

The rare case where the data rights are a contractual consequence of the customer relationship rather than a separate purchase: covering 55% of U.S. oncologists is what produces the longitudinal corpus its own foundation model is trained on. Evals are internal validation, not an owned standard.

客户关系	完全具备	The FY2025 10-K states its provider network "covers more than 55% of U.S. oncologists" — a direct contracting relationship with the ordering physician, not access resold through a distributor. 来源：sec.gov
专有数据	完全具备	The same 10-K sentence states the network "provides us with broad data rights, including the rights to longitudinally updated data" — the rights travel with the clinical relationship rather than being licensed in, and the library is described as 500+ petabytes across 45M+ de-identified patient journeys. 来源：sec.gov
评测体系	部分具备	Reports its own validation — a C-index of 0.802 for overall survival (p<0.001) in EGFR-mutant NSCLC — but this is internal measurement of its own model, not a published benchmark the field is scored against. Looked for a domain evaluation standard it owns and defines; found performance results instead. 来源：biospace.com
工作流	完全具备	The 10-K describes running next-generation sequencing on behalf of the patient and providing clinical trial matching and data structuring services — the ordering, sequencing, reporting and matching steps run through Tempus rather than being handed back at the assay. 来源：sec.gov
领域专用模型	完全具备	Constructed and trained its own multimodal transformer foundation model on 2.5M longitudinal records — 250M+ pages of clinical notes, 450,000 digitized images and 500,000 genomic and transcriptomic sequences — rather than adapting a third-party base. 来源：biospace.com

Legal

依据以下条件侦察：a company that simultaneously owns the customer relationship, the data, the evals, the workflow, and a domain-specific model

EvenUp

未上市编辑估算截至 2026-08-05

4 项完全具备 · 1 项未知 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

EvenUp operates as production infrastructure inside the plaintiff personal injury bar rather than as an occasional assist. About 30% of the top 100 US personal injury firms use the platform, which handles more than 10,000 cases a week, over 520,000 a year, spanning roughly $14 billion in damages across active matters. Because it generates the artifacts these firms produce repeatedly, demand packages, medical chronologies and case valuations, the AI sits in the daily drafting path rather than beside it. Its Pre-Litigation-as-a-Service tier, which pairs the models with US-based case staff across the full pre-litigation lifecycle, booked over $10 million in subscription sales in early testing with demands delivered 47 days faster; the other 70% of that top-100 tier remains unsigned. 来源：lawnext.com

The clearest data-compounding loop in the legal set: the settlements its customers resolve on the platform become the corpus its own model is priced on. Evals are undocumented, which is the one gap in an otherwise closed loop.

客户关系	完全具备	Contracts directly with plaintiff firms — over 2,000 use the platform, including 20% of the top 100 U.S. personal-injury firms — rather than reaching them through a case-management vendor. 来源：evenuplaw.com
专有数据	完全具备	A settlement repository built from its own customers' resolved matters — over 200,000 cases resolved on the platform, and models trained on hundreds of thousands of injury cases and millions of medical records. The asset accrues from the work rather than being licensed in. 来源：evenuplaw.com
评测体系	未知	Looked for a published evaluation methodology, an accuracy benchmark or a model card and found none in the cited source. Absence of evidence, not evidence of absence. 来源：evenuplaw.com
工作流	完全具备	Spans the personal-injury lifecycle in-platform — intake, treatment documentation, demand generation, negotiation, discovery and trial preparation — rather than producing one document and handing the matter back. 来源：evenuplaw.com
领域专用模型	完全具备	Piai, described as its own model trained on the largest dataset in personal injury, rather than a domain prompt over a third-party frontier base. 来源：evenuplaw.com

Harvey

未上市编辑估算截至 2026-08-05

3 项完全具备 · 2 项部分具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Harvey reached $300M in annual recurring revenue in May 2026, up from $195M at the end of 2025, with 1,500+ customers in 60+ countries covering half of the Am Law 100 and 50 asset managers. Usage is daily production work rather than trial: 142,000+ lawyers on the platform, 700,000+ tasks executed daily, 50M+ terms extracted weekly, 9.75M active files, monthly queries up 5.5x and weekly active users up fourfold year over year. Median seat count inside an account doubles within twelve months. Harvey-commissioned adoption research reported by Legal IT Insider has 55% of law firms describing it as foundational technology rather than a tool, while 32% of organizations still have fewer than 30% power users. 来源：sacra.com

Owns the customer relationship, the workflow and — unusually — the evaluation methodology the whole vertical is measured against, but adapts third-party frontier models rather than owning one, and the proprietary data is largely the client's. Four-fifths of the thesis, missing the model layer.

客户关系	完全具备	Sells directly into 1,300+ organisations across 60+ countries including the majority of the AmLaw 100, 500+ in-house teams and named customers NBCUniversal, HSBC and DLA Piper International, with legal engineering teams embedded at the customer. 来源：harvey.ai
专有数据	部分具备	Models are augmented with proprietary data for specific legal tasks, but the worked example is a custom tax model trained on a professional-services firm's own internal tax data by that firm's in-house experts — the corpus belongs to the client, not the platform. 来源：microsoft.com
评测体系	完全具备	Built BigLaw Bench, an evaluation framework designed by its own legal research team of practising attorneys, scoring tasks derived from lawyer time entries against custom rubrics on two metrics — Answer Score (share of a lawyer-quality work product completed) and Source Score (share of claims correctly cited). 来源：harvey.ai
工作流	完全具备	25,000+ custom agents execute M&A, due diligence, contract drafting, document review and fund formation inside the platform, including long-horizon agents running multi-step workflows over extended periods. 来源：harvey.ai
领域专用模型	部分具备	Engineers primarily use Azure OpenAI Service o1-preview, o1-mini and GPT-series models, post-trained and fine-tuned with OpenAI and Microsoft; training its own models is described as something the company is looking at, not something it has done. 来源：microsoft.com

Legora

未上市编辑估算截至 2026-08-05

2 项完全具备 · 2 项未知 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Legora's March 2026 Series D announcement states the platform "supports tens of thousands of lawyers each day across 800 customers in more than 50 markets," naming White & Case, Cleary Gottlieb, Goodwin, Linklaters, Dentons and Deloitte among them - daily workflow use, not occasional assist. CEO Max Junestrand frames US buyers as having moved "from experimentation to embedding AI across their organisations." Where firms disclose intensity it runs high: on Legora's customer page, BAHR's managing partner reports "as much as 80% of our people are active users, and as high as 30% use Legora more than ten times a day." Headroom is still being built against, with more than 300 US employees and new Houston and Chicago offices planned by end-2026. 来源：legora.com

Owns the customer relationship and the workflow at real scale, and is the clearest counter-example in this vertical: it declines the model layer as a matter of strategy rather than capability. Data and evals are undocumented in the sources fetched.

客户关系	完全具备	Serves 1,000+ law firms and in-house legal teams across 50+ markets with tens of thousands of legal professionals using it daily; named clients include White & Case, Cleary Gottlieb, Goodwin, Linklaters, Dentons, Deloitte, Blackstone and Barclays. 来源：bvp.com
专有数据	未知	Looked for a proprietary corpus, an exclusive data licence or a data asset the company claims as its own and found none in the fetched sources; the described strategy is scaffolding and UX on top of foundation models, which is silent on data ownership. Absence of evidence, not evidence of absence. 来源：bvp.com
评测体系	未知	Looked for a published legal evaluation methodology, a named benchmark or a model card and found none in the fetched sources. The company is reported to evaluate and route across roughly ten models, but that is internal model selection rather than a domain benchmark it owns and publishes. 来源：bvp.com
工作流	完全具备	Contract review, due diligence, legal research, multi-document Tabular Review and AI Agents run in the platform, surfaced into Microsoft Word and SharePoint for annotation and redlines — the analysis is Legora's, the surface is Microsoft's. 来源：microsoft.com
领域专用模型	不具备	Explicitly does not build proprietary models: the stated bet is that betting against the model companies is a losing strategy, so the company builds the scaffolding, software and user experience on top of third-party frontier models, using GPT-4o and o1 via Azure OpenAI Service. 来源：bvp.com

Luminance

未上市编辑估算截至 2026-08-05

5 项完全具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Luminance reports its contract AI in production at over 1,000 enterprises across 70+ countries, having analysed more than 18 million contracts in the twelve months to its January 2026 platform relaunch — volume that sits inside daily negotiation and review workflow rather than alongside it. Its models were built on over 220 million documents, and the company claims a 70-80% cut in contract negotiation time, rising to 90% with the relaunch, and over 30% of legal teams' time returned. Deloitte, Quantinuum, Ingram Micro and Baringa are named as relaunch design partners. Account penetration looks uneven: Luminance describes 2025 as the year of its first eight-figure enterprise deal, implying most of the installed base sits well below that. Figures are company-reported. 来源：luminance.com

The only candidate in this vertical banded owns on all five: a proprietary contract model trained on its own 220M-document corpus, an in-house benchmark defining accuracy, the full contract lifecycle in-platform, and a direct enterprise customer base. The evals and model claims rest on a single trade-press account of the company's own announcement.

客户关系	完全具备	The company's own site claims 1,000+ customers and displays named enterprises including AMD, Danone, Ralph Lauren, Tesco, NTT DATA and E.ON as the organisations running the platform, i.e. the enterprise is the contracting customer rather than a channel. 来源：luminance.com
专有数据	完全具备	Luna Crescent was fine-tuned on a curated subset of what the company describes as its own proprietary legal dataset, comprising more than 220 million verified legal documents. 来源：lawyersweekly.com.au
评测体系	完全具备	The company built ContractIQ Bench, an internal benchmarking framework of more than 180,000 manually reviewed data points, and uses it to score interpretation of liability clauses, termination rights and confidentiality obligations. The benchmark is internal and self-reported, not independently administered. 来源：lawyersweekly.com.au
工作流	完全具备	The product spans the contract lifecycle in-platform — Draft (automated contract generation), Negotiate, Analyze, Comply, Investigate and Collaborate — rather than handing work off at review. 来源：luminance.com
领域专用模型	完全具备	Luna Crescent is a proprietary vertical model series purpose-built for contract work and built on the company's own AI technology stack, not a wrapper over a third-party frontier base. 来源：lawyersweekly.com.au

Other

依据以下条件侦察：a company that simultaneously owns the customer relationship, the data, the evals, the workflow, and a domain-specific model

Duolingo (DUOL)

编辑估算截至 2026-08-05

3 项完全具备 · 2 项部分具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Duolingo's fiscal 2025 annual report describes machine learning that evaluates every learner's answer to every exercise every day and adaptively builds each lesson, putting AI in the core daily loop for the 133.1 million monthly and 52.7 million daily active users reported for the fourth quarter. Roleplay, Explain My Answer, DuoRadio and Video Call are live generative features, and the company credits AI with generating course content materially faster across 250-plus courses. The cost is visible in the accounts: gross margin fell to 72.2% from 72.8%, attributed partly to increased AI costs in features like Video Call. Duolingo Max, the AI-heavy tier, is still offered to only a portion of the user base, and about 9% of monthly active users were paid subscribers at year-end 2025. 来源：sec.gov

Duolingo owns the three consumer-side dimensions outright — the learner relationship, the first-party exercise corpus, and the daily learning session all sit inside its own product. It is partial where it depends on outside parties: the frontier generative layer is licensed from OpenAI, and its assessment is scored against the externally owned CEFR and IELTS yardsticks.

客户关系	完全具备	Duolingo is pure direct-to-consumer: learners come to the brand directly and hold Duolingo accounts. The FY2025 10-K reports 133.1M MAUs (three months ended 2025-12-31 average) and 12.2M paid subscribers at period end, up from 9.5M, and says word-of-mouth growth lets it "focus our capital investments on product innovation and data analytics, rather than relying primarily on brand or performance marketing." Pricing is its own: "We determine the prices at which these subscriptions and features are sold." The app stores are a real dependency — 61.6%, 23.4% and 10.3% of FY2025 revenue processed by Apple, Google and Stripe, a "generally 15-30%" cut, and agreements "terminable... without cause with 30 days prior written notice" — but no partner sits between Duolingo and the learner. 来源：sec.gov
专有数据	完全具备	The training corpus is first-party learner telemetry generated inside Duolingo's own product by its own consumer users, with no enterprise tenant's data and no licensed-in corpus doing the work. The FY2025 10-K states: "Large data moat. With nearly 2 billion exercises completed every day on our platform, we believe we have built the world's largest collection of language-learning data", and separately that learners create "what we believe to be the world's largest learning dataset"; 52.7M DAUs for Q4 2025 confirm the volume. The filing adds that Duolingo does "not sell personal data to outside parties." One qualifier: its AI risk factors concede it may be "deemed to not have sufficient rights to the data we use to train our AI technologies", so not every training input is unambiguously its own, even though the load-bearing asset is. 来源：sec.gov
评测体系	部分具备	Duolingo owns a real assessment instrument but not the standard it is measured against. The CIOL-validated paper calls CEFR "the leading international standard adopted by governments, educational institutions, and assessment providers" and "the standard that Duolingo uses to reference all of our learning and assessment products, including the Duolingo English Test." Legitimacy is externally conferred: the alignment was "validated externally by ACS Ventures, LLC, an independent psychometric company", and the methodology "was reviewed and validated by the Chartered Institute of Linguists (CIOL)." Scores are anchored to IELTS via "3,586 test takers." The test has its own adaptive scale and "over 6,100 education programs" accept it, but it is small and shrinking: $42,006K of 2025 revenue vs $45,640K in 2024, 4.05% of the $1,037,589K total. 来源：ciol.org.uk
工作流	完全具备	The day's work — the learning session — happens entirely inside Duolingo's product, with no external system of record to hand off to. The FY2025 10-K reports roughly 52.7M DAUs for the three months ended 2025-12-31 versus 40.5M, "representing an increase of 30%", a DAU/MAU ratio that "increased to 39.6% from 34.7% a year ago", and "about 43 million daily active users with a 7-day streak or longer, and about 15 million daily active users with a 365-day streak or longer." Adjacent subjects and the assessment run on one stack: "Products across our platform, like our flagship Duolingo App (including our Math, Music, and Chess courses), and the Duolingo English Test, share a singular technology infrastructure." A 39.6% ratio means roughly 60% of monthly learners are not daily, and the learner's ultimate goal is consummated outside the app. 来源：sec.gov
领域专用模型	部分具备	Duolingo owns a genuine domain-specific model but rents the frontier generative layer. The FY2025 10-K describes "our machine learning tools that evaluate every learner's answer to every exercise every day and learn to predict the probability that any learner will answer any given exercise correctly" — the model Duolingo publicly calls Birdbrain, used "to figure out which exercises in a particular lesson are going to be the best match for a learner's level of knowledge." The generative features are third-party: "Duolingo Max Uses OpenAI's GPT-4 For New Learning Features", with Video Call and Roleplay "powered by the newest, most powerful generative AI technology." The 10-K warns that if model providers "terminate their relationship with us... our business will be harmed", and gross margin fell to 72.2% from 72.8%. 来源：blog.duolingo.com

Customer support

依据以下条件侦察：a company that simultaneously owns the customer relationship, the data, the evals, the workflow, and a domain-specific model

Salesforce (CRM)

编辑估算截至 2026-08-05

2 项完全具备 · 3 项部分具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

Salesforce reported $1.2 billion of Agentforce annual recurring revenue for the quarter ended April 30, 2026, up 205% year over year, inside nearly $3.4 billion of combined Agentforce and Data 360 ARR. Usage carries the evidence: 3.8 billion Agentic Work Units delivered to date across Agentforce and Slack, up 111% quarter over quarter, and 28.6 trillion tokens processed to date, up 152%. Chair and CEO Marc Benioff described Agentforce as powering every Customer 360 application across tens of thousands of businesses. Headroom sits in the same release: more than half of Agentforce and Data 360 bookings came from existing customers, and that $1.2 billion runs against full-year FY27 revenue guidance of $45.9 billion to $46.2 billion. 来源：sec.gov

Salesforce owns the enterprise relationship and the workflow surface outright, and holds partial positions on data, evaluation and models. The pattern is a company that owns where the work happens and who it is sold to, while the reasoning layer it runs on is largely bought.

客户关系	完全具备	Salesforce contracts directly with the enterprise: the FY26 10-K states it sells "to businesses worldwide, primarily on a subscription basis, through our direct sales efforts and also indirectly through partners," with subscription periods "typically 12 to 36 months." The direct sales motion, the contracting counterparty and the multi-year renewal cycle put Salesforce rather than an intermediary at the centre of the relationship, and the partner channel is disclosed as secondary. The product also reaches the enterprise's own end customers: "Our Agentforce Service offering enables companies in every industry to bring all of their customer, employee, IT and field service needs onto one integrated, AI-powered platform to deliver trusted, highly personalized customer support at scale." 来源：sec.gov
专有数据	部分具备	The corpus Agentforce reasons over is the customer's own tenant data, retrieved at run time rather than owned. The Einstein Trust Layer "takes any direct merge fields inside of the prompt and grounds it with record data from Salesforce," via Flow data providers and Data Cloud retrievers, substitutes "a designated placeholder" for PII, and commits that "This 'zero retention' architecture ensures that customer data is never stored outside of Salesforce" — contracting away cross-tenant learning. The FY26 10-K discloses no training or fine-tuning on customer data. What Salesforce holds is the CRM schema and metadata model plus the unification and residency layer: structure and custody, not an owned corpus. The 112 trillion records Data 360 ingested in FY26, up 114% Y/Y, measure that custody volume. 来源：developer.salesforce.com
评测体系	部分具备	Salesforce AI Research publishes its own benchmark, CRMArena-Pro ("Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions," 24 May 2025), covering "nineteen expert-validated tasks across sales, service, and 'configure, price, and quote' processes" with "multi-turn interactions guided by diverse personas and robust confidentiality awareness assessments." It reports that "leading LLM agents achieve only around 58% single-turn success on CRMArena-Pro, with performance dropping significantly to approximately 35% in multi-turn settings." The dataset is released under cc-by-nc-4.0 for "research purposes only in support of an academic paper," no in-product pass/fail bar or grader set is publicly reachable, and outside labs do not measure themselves against it. 来源：arxiv.org
工作流	完全具备	The work happens inside Salesforce's own objects. The FY26 10-K states that "Agentforce is infused across every Salesforce application, so intelligence is embedded directly into the workflows where work actually happens. Instead of switching between tools, people can collaborate with agents inside and across the systems they already use every day." It adds that agents "can access live business data through Data 360, follow company policies defined in Salesforce metadata, and take action through Salesforce applications and MuleSoft Application Programming Interfaces ('APIs')," with "deterministic business logic, workflows, and policies" instantly accessible without rebuilding existing processes. Slack sits inside the same perimeter, at $8,882M FY26 subscription revenue in its segment. 来源：sec.gov
领域专用模型	部分具备	Salesforce AI Research trains real domain models — the xLAM action and function-calling family and xGen — and the company says of xLAM that "A significantly more advanced, proprietary version is powering Agentforce," describing xGen-Sales as "a proprietary model trained and designed to power autonomous sales tasks with Agentforce." The reasoning and generation layer is bought: the default is "a managed mix of trusted models (currently including GPT-4o) optimized by Salesforce," and every selectable model is external — GPT-4o through GPT-5.x and o3/o4-mini, Anthropic Claude via Bedrock, Gemini via Vertex, Amazon Nova, NVIDIA Nemotron — plus bring-your-own-LLM. The proprietary claim rests on the filing's own characterisation from September 2024; the FY26 10-K names no such model. 来源：salesforce.com

ServiceNow (NOW)

编辑估算截至 2026-08-05

3 项完全具备 · 2 项部分具备 ，共 5 项 — 这是一种格局，不是分数

AI 落地深度：深

ServiceNow's AI business crossed $1 billion in annual contract value in the second quarter of 2026, and the company reported that agentic deployments of ServiceNow AI rose ninefold over the preceding nine months. The footprint is broad rather than confined to lighthouse accounts: nearly all 50 U.S. states run the ServiceNow AI Platform, with state agencies citing outcomes such as a 66% service desk cost reduction, and the quarter closed with 658 customers above $5 million in annual contract value. Set against quarterly subscription revenues of $3,877 million, that $1 billion AI base remains a modest slice of the installed estate; management's Financial Analyst Day target of 30% of ACV from AI by 2030 measures how much of its own customer base it treats as unconverted. 来源：sec.gov

ServiceNow owns the customer relationship, the workflow execution layer and the evaluation standard for agents on its platform, and the economics reflect it: $13.28B FY2025 revenue at 97% subscription, $28.2B RPO up 27%, and 603 customers above $5M ACV against a roughly flat ~8,700 logo base. The softer dimensions are data, where a multi-instance architecture keeps the operating corpus per-tenant by design, and models, where in-house LLMs stay available but lose the default inference path to third-party providers from July 2026.

客户关系	完全具备	ServiceNow sells directly: "We sell our products and services to enterprises across a wide variety of industries through subscription agreements facilitated by our global direct sales organization," and "As of December 31, 2025, we had approximately 8,700 customers across a wide variety of industries." The relationship is quantified, not asserted — "Revenues from our direct sales organization represented 78% of our total revenues for each of the years ended December 31, 2025 and 2024." The remaining ~22% is indirect: "We also sell through managed service providers and resale partners," plus integrator ties with "Accenture, Cognizant, Deloitte, EY, Infosys and KPMG," and a risk factor noting "we and our customers rely on our partners to provide professional services, including custom implementations." 来源：sec.gov
专有数据	部分具备	The operating corpus is per-tenant customer data that ServiceNow deliberately does not centralize: "We operate a multi-instance architecture that provides each customer with a dedicated application layer and database," and Workflow Data Fabric works by "connecting to real-time external data without copying or moving the data from its source." What ServiceNow does own is structure rather than corpus — the CMDB, workflow schema and process taxonomy — plus first-party telemetry from the "Now on Now" program, which the filing describes as using "our own products to collect real-time feedback on new and existing features," a product-feedback loop rather than a training corpus. The filing discloses no cross-customer aggregated training set. On that evidence this sits at the generous end of partial. 来源：sec.gov
评测体系	完全具备	WorkArena is a peer-reviewed artifact, not a product name: "a remote-hosted benchmark of 33 tasks based on the widely-used ServiceNow platform" measuring agents on "the typical daily work of knowledge workers utilizing enterprise software systems," by 12 ServiceNow Research staff, shipped with BrowserGym as a "unified testing ground for MiniWob, WebArena, WorkArena" (arXiv 2403.07718; ICML 2024, PMLR v235). ServiceNow also gates production models: "all integrated models are regularly evaluated on Platform-representative data," "models vetted for reliability and performance," and "ServiceNow product teams test and validate OOTB skills against the assigned third-party provider before any default changes." The scope limit is real: this defines good on and for the ServiceNow platform, not an industry-wide support standard. 来源：servicenow.github.io
工作流	完全具备	The 10-K frames the platform as the execution layer AI models lack: they "cannot independently execute transactions, route approvals, update systems of record, or support compliance with business rules and regulatory requirements. Our platform addresses this fundamental gap by providing the underlying infrastructure that connects AI-generated insights to the operational systems and processes where work actually gets done," and "AI agents developed using our solutions follow a human-in-the-loop governance model." Subscription revenues were $12,883M of $13,278M total — 97% of revenue, growing 21%; the auditor cites "subscription revenues of $12.9 billion for the year ended December 31, 2025." The moat rests on "expertise in the operational processes specific to different industries, functional areas and user roles." 来源：sec.gov
领域专用模型	部分具备	ServiceNow offers its own models but is moving them off the default inference path: "Beginning July 9, 2026, ServiceNow will begin rolling out third-party model providers as the default for out-of-the-box (OOTB) Now Assist skills and AI agents," and "If you've set the Now LLM Service as your instance level model provider, we recommend starting the process of getting at least one third-party model provider approved." Any skill "still running OOTB default of Now LLM Service will switch to the new third-party defaults when you update," against a table of "deprecation targets, and retirement targets." In-house models remain selectable — "organizations can choose to leverage ServiceNow's language models or integrate third-party or proprietary models" — but third-party providers are becoming the defaults. 来源：servicenow.com