Overview
Humanloop changed its product while helping experts teach software what good work looks like. It began with data labeling in 2020, moved into large language model development in 2022, and built an enterprise evaluation platform. Its team joined Anthropic in August 2025, and the standalone platform closed the following month. [1][2]
The team recognized that better models threatened its original annotation business. Humanloop moved toward the work those models created: testing outputs, preserving prompt history, and bringing domain experts into development. That adaptation explains the product's trajectory. Public evidence does not explain the financial terms or full motivation for the exit. [3]
Image 1 / 1
Founding Story
Raza Habib and Peter Hayes met through machine learning research at University College London. Jordan Burgess joined from Amazon Alexa. Habib had worked at Google and Monolith AI; Hayes previously co-founded fintech company Certua. The company joined YC's Summer 2020 batch. UCL also identifies professors David Barber and Emine Yilmaz in the founding team. [1][4][5]
Their July 2020 launch described a concrete bottleneck. Lawyers and other specialists labeled text because outsourced annotators lacked the required knowledge. Humanloop trained a model as the expert worked, then selected useful examples for the next annotation. Early work included two large law firms. [6]
Programmatic, a later product, let experts express labeling rules. In July 2022, investor Index Ventures described a $2.6 million seed round. Habib explained the problem: “getting the data into a form that the algorithm can use is the biggest challenge.” [7]
Better general-purpose models challenged that premise. Habib told YC that the team explored a pivot before ChatGPT's release. It planned two weeks to find ten paying customers for an LLM development tool. Habib described the result: “In the end, it took us two days.” This was a founder-reported commercial test, with no published cohort or retention data. [3]
Timeline
- July 29, 2020: The founders launched their annotation, training, and deployment platform on Hacker News. [6]
- July 28, 2022: Index published the Programmatic launch and $2.6 million seed announcement. [7]
- October 5, 2022: Humanloop opened its LLM development product. [8]
- Late 2023: The team established a San Francisco office while retaining engineering in London. [5]
- November 21, 2024: The enterprise evaluation platform reached general availability. [9]
- August 13, 2025: Humanloop announced the Anthropic move and a September 8 platform sunset. Billing had stopped on July 30. [2]
What They Built
The earlier annotation product also changed through customer use. Habib wrote in July 2021 that teams often revised their label categories after exploring real examples. Humanloop added editable categories and discussion around individual data points. Experts needed to agree on the task before the model could learn it. [23]
The October 2022 product connected user feedback to prompt and model experiments. Developers replaced a model call with Humanloop's SDK, which recorded inputs, outputs, and model settings. They could compare prompts using user outcomes and select examples for fine-tuning. The launch described early work with OpenAI and GPT-3 adopters. [8]
By 2024, the central workflow was evaluation: define what an application should do, run it against examples, and compare changes. Humanloop supported code checks, model judgments, and human reviews. Versioned prompts, tools, evaluators, and multi-step flows connected results to the configuration that produced them. Its interface let engineers and subject experts work on the same examples. [9]
Tracing recorded the steps inside an application. A weak answer might originate in retrieval, a tool call, or a prompt. Humanloop's Flows view exposed those steps. OpenTelemetry integration helped teams instrument existing applications, while useful production logs could become evaluation datasets. [10]
In a November 2023 interview, Habib explained the product's practical advantage. A domain expert could find a bad production example, reopen it in the editor, change the prompt, and test for regressions. Keeping those steps together reduced the handoffs between monitoring tools, source code, and deployment. The claimed speed improvement was his account, rather than an independently measured benchmark. [11]
Market Position
Target Customers
Humanloop targeted teams already building AI applications. Habib described the strongest sales signal as buyers who had tried building evaluation tools themselves. Very small teams had fewer collaboration problems, while very large enterprises brought long procurement cycles. [11]
Customer materials named Gusto, Filevine, Dixa, FMG, Athena, and Fundrise. Their use cases spanned support, legal software, financial marketing, and internal assistants. Humanloop published customer performance claims, but those stories cannot isolate its contribution to broader customer results. [12]
In Fundrise’s March 2026 podcast transcript, CEO Ben Miller described Humanloop as a product Fundrise had used. The episode introduced Habib as a member of Anthropic’s technical staff. [24]
Market Size
No reliable market-size estimate or Humanloop revenue series was established. Distribution evidence is more concrete. An Index case study reports that the US provided 75% of revenue at the time of its account. Habib and Burgess worked in San Francisco, with Hayes and most engineering in London. The dual presence brought customer access alongside salary, visa, and time-zone costs. [5]
The High Agency podcast was another distribution channel. Its interviews with builders at Ironclad, Sourcegraph, Zapier, and Hex addressed the implementation problems Humanloop sold tools to solve. Its commercial contribution was not disclosed. [3][13]
Competition
As of October 2026, evaluation and tracing remain available from other vendors. Langfuse offers an open-source platform with self-hosting and cloud service. Its ClickHouse acquisition announcement preserved both paths. Its advertised Core cloud plan starts at $29 per month. [14][15][16]
Braintrust supports experiments, production scoring, human annotation, and turning traces into datasets. Its advertised Starter plan is free and Pro starts at $249 per month. A new entrant therefore needs a specific workflow advantage beyond collecting traces or running model judges. [17][18]
Business Model
Humanloop sold a software workflow alongside separately billed model usage. The historical pricing page offered a free tier with two members, 50 evaluation runs, and 10,000 monthly logs. Enterprise sales included access controls, private-deployment options, and support. Customers supplied their own provider keys and paid model providers separately. Those surviving pages describe historical terms; they are not an available service. [19][20]
In its November 21, 2024 general-availability announcement, Humanloop reported $8 million in total funding. It named YC Continuity, Index Ventures, Localglobe, and founders of Datadog and Segment among its backers. Funding does not establish contract values, retention, profitability, or acquisition consideration, which remain undisclosed in the reviewed sources. [9]
Traction
Peter Hayes's December 2024 retrospective reported more than 300 production deployments culminating in 50 product releases that year. It also reported millions of daily logs across thousands of AI products. These measure engineering activity and usage, with the company as their source. They do not count paying organizations or establish revenue. [10]
Post-Mortem
Human expertise survived the product change
Humanloop's original annotation task became less necessary as pretrained models improved. The team responded by moving the expert's judgment later in the workflow, toward evaluating applications built with those models. The customer problem remained recognizable across both products. This continuity is the strongest explanation for why its pivot was more than a change of branding. [3][6]
Its later interface coupled observation with intervention. A report about poor outputs has limited value if the person who understands the error cannot test a correction. Humanloop made domain expertise part of the development loop. That mechanism appears in both the product documentation and Habib's founder interview. [9][11]
The team moved; customers had to migrate
Humanloop's announcement called Anthropic the “ideal home to amplify our impact.” The statement supports the founders' expressed preference, without explaining private negotiation terms. [21]
TechCrunch reported a team acquisition; an Anthropic spokesperson said Humanloop's assets and intellectual property were excluded. [22]
The operational notice required customers to export their data before September 8, 2025. It promised prorated refunds for annual subscriptions extending beyond that date. The result separated the team’s next employer from the continuation of its customers' evaluation workflow. [2]
A competing explanation is that the company chose a talent deal because standalone economics were weak. The public record reviewed here cannot confirm or reject that explanation. Named customers and usage show adoption; they do not settle whether the business could sustain itself. No reviewed source links the exit to losses or retention.
Key Lessons
- Humanloop tested a potentially necessary pivot with paying customers before committing. Its reported two-day result gave a clearer signal than interest in mockups alone. [3]
- Humanloop kept experts close to both the evidence and the correction. Combining tracing, editing, and regression checks reduced handoffs that separate dashboards would leave to customers. [11]
- Humanloop's US presence followed commercial demand, while engineering remained near its research base. The arrangement also imposed costs and a narrow overlap in working hours. [5]
- Humanloop's customer migration shows why buyers need portable datasets, prompt versions, and evaluation records. A useful tool's team can move even when customers still need the workflow. [2]
Sources
- YC company profile
- Platform sunset notice
- YC founder interview
- UCL spinout profile
- Index US-expansion case study
- Founder launch discussion
- Index investment announcement
- LLM product launch
- General-availability announcement
- Product retrospective
- Founder interview with Matt Turck
- Customer stories
- High Agency podcast
- Langfuse product
- Langfuse acquisition
- Langfuse pricing
- Braintrust evaluation documentation
- Braintrust pricing
- Historical pricing
- Historical quickstart
- Team announcement
- Deal reporting
- Founder lessons from 2021
- Fundrise customer and founder sequel