Guardrails AI: Revenue, sale and how it was built
Guardrails AI (Open-source checks for language-model output, a managed enterprise version and a simulation service for testing AI agents, United States): Acquired by Harvey with its team, September 2026. Evidence grade E: no revenue figure was checked. Team: 2 co-founders named by the buyer (Shreya Rajpal, CEO, and Zayd Simjee); headcount at the sale not published. Launched August 2023; sold 37 months after launch.Added 7 November 2026
- Outcome
- Acquired by Harvey with its team, September 2026
- Evidence
- E no revenue figure was checked
- Launch
- August 2023
- Status
- Sold, 37 months after launch
- Team
- 2 co-founders named by the buyer (Shreya Rajpal, CEO, and Zayd Simjee); headcount at the sale not published
- Outside funding
- US$7.5M seed led by Zetta Venture Partners (February 2024, the company said); investors at the sale also include Bloomberg Beta, Pear VC, Factory and Microsoft
- Main channel
- Open source
What is not verified
No sale price, revenue figure or count of paying customers was found. We read Harvey's announcement, the company's own post, Dealroom's item, the 2024 press release, the blog, the home page and archived copies of the pricing and Pro pages; none gives an amount. The figure of more than 250,000 monthly downloads is Harvey's statement and was not checked. The headcount at the sale was not published. The home page footer shows SOC 2 Type II and HIPAA badges with no link to an audit report or a trust page, and we found no source that confirms them. The Changi Airport case was part of a government testing pilot, and the page does not say whether it was paid. A reader could look for a post by the lead investor, Zetta Venture Partners, or a later interview with the founders. This story is grade E: the business is real and its product, price and customers are described from public pages, but no revenue figure was checked. Aiolum shows it as an idea worth studying, not as proof that it makes money.
Guardrails AI made software that checks the output of a large language model (LLM), a program that writes text by predicting the next word, before the answer reaches a user. Each check is a validator, a small rule or a small trained model that tests one thing, for example whether an answer leaks personal data or invents facts; validators combined around one model call form a guard. The framework was free and open source. The company sold Guardrails Pro, a managed version for large companies, and Snowglobe, a simulation testing service in which AI-generated users hold many conversations with a customer's chatbot to find where it fails.
Shreya Rajpal, who had built AI systems for self-driving cars at Drive.ai and for autonomous systems at Apple, said the open-source Guardrails project launched in 2023; the buyer names her and Zayd Simjee as the co-founders. The company formally launched on 15 February 2024 with Guardrails Hub, and its press release said it had closed a US$7.5 million seed round led by Zetta Venture Partners. On 9 September 2026 Harvey announced that it had acquired Guardrails AI, that the co-founders and their team would join its product and engineering organization, and that the framework was downloaded more than 250,000 times a month.
No sale price, revenue figure or count of paying customers was published, and neither was the team size at the sale, so this story has grade E: a real business with checked facts but no money figure. The strongest source is Harvey's own announcement.

How much did Guardrails AI make, and what was it sold for?
Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.
| Fact | Value | Grade | How we know | Sources |
|---|---|---|---|---|
| Outcome, 9 September 2026 | acquired by Harvey; the co-founders and their team joined Harvey; price not disclosed | A | Harvey's own announcement: "Harvey today announced its acquisition of San Francisco-based Guardrails AI". It calls this "Harvey's fourth acquisition of 2026". The same text is on the Guardrails AI blog. Dealroom reported it on 10 September 2026. None of them states a price. | [1, 2, 3] |
| Founders | Shreya Rajpal (co-founder and CEO) and Zayd Simjee (co-founder) | A | Named by Harvey as "Guardrails' co-founders". The 2024 press release calls Rajpal co-founder and CEO and says she built AI systems at Drive.ai and Apple. | [1, 4] |
| Start | open-source project launched in 2023; first company blog post 31 August 2023; based in San Francisco | C | Rajpal in the press release of 15 February 2024: "With the launch of Guardrails in 2023". The blog post of 31 August 2023 begins "We're Guardrails AI, creator and maintainer of the open-source project Guardrails." Harvey wrote that the team "has spent three years" on the problem. | [4, 6, 1] |
| Formal launch, 15 February 2024 | company launch and Guardrails Hub, a free catalogue of validators | C | Press release: launched "during the opening keynote at the AI in Production conference". DBTA reported that the Hub "already has 50 pre-built validators". The Guardrails Pro post of November 2024 says 65 or more. | [4, 5, 7, 8] |
| Funding | US$7.5 million seed, February 2024; Microsoft among the investors at the sale | C | The company's press release: a seed round led by Zetta Venture Partners, with Bloomberg Beta, Pear VC, Factory and GitHub Fund, to "expand the company's engineering and product teams". Harvey's post lists Zetta, Bloomberg Beta, Pear VC, Factory and Microsoft. | [4, 1] |
| First paid product, 8 November 2024 | Guardrails Pro, a managed service for enterprise AI platform teams | C | Launch post: "our managed enterprise offering", built on the open-source framework, deployed in the customer's VPC, with a private hub and an observability dashboard. No public price; a demo call was the way to buy. | [8, 13] |
| Snowglobe price, October 2025 | US$0.25 per generated message, first 250 messages free; Enterprise on request | C | Archived pricing page of 7 October 2025: "Pay for what you simulate". The Enterprise plan adds a forward deployed engineer and "Guaranteed KPIs on agent performance". | [9] |
| Named customers | MasterClass and Changi Airport Group (Snowglobe) | C | Case studies of 13 and 14 August 2025. MasterClass's head of AI is quoted: "We have completely switched to using SnowGlobe for this data." Changi Airport tested its AskMax chatbot "As part of the AI Verify Pilot". Neither page says what was paid. | [12, 11, 14] |
| Framework downloads, September 2026 | more than 250,000 a month | A | Stated by Harvey in its announcement, with no further detail. Not checked against the package index, whose page answered our download with a bot challenge. | [1, 3] |
| Team size at the sale | not published | C | Harvey writes of the co-founders "and their team". No source gives a headcount. | [1, 3] |
| Revenue and sale price | not published | C | Not found in Harvey's post, the company's post, Dealroom's item, the press release or the archived pricing pages. | [1, 2, 3, 4, 9] |
| Time from the first dated company post to the sale | 37 months | calc | Calculated: August 2023 (the blog post of 31 August 2023) to September 2026. | [6, 1] |
How to read this table: a grade A value is stated by the buyer, Harvey, in its own announcement. A grade C value is a statement by Guardrails AI on its own pages, in its press release or on an archived copy of its pricing page. A "calc" value was computed on this page. No money amount for the sale or for revenue exists in the sources.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Snowglobe Self-service (October 2025) | 0.25 per generated message | first 250 messages free; up to 3 AI applications; 250 scenarios an hour | |
| Snowglobe Enterprise (October 2025) | custom pricing | forward deployed engineer, unlimited simulations, VPC or on-premise deployment | |
| Guardrails Pro (from November 2024) | on request | managed service, fully hosted or in the customer's VPC; private hub; monitoring dashboard | |
| Open-source framework and Guardrails Hub | 0 | open source; Hub validators installed with an API key from the Hub website |
Snowglobe was priced by usage, per generated message, so the yearly column is empty. Enterprise plans for Snowglobe and Guardrails Pro had no public price and were sold after a demo call. The open-source framework and the public Guardrails Hub were free; the Hub needs an API key from its website.
When did Guardrails AI launch, and how did it grow?
From an open-source project in 2023 to a sale to a legal AI company, 37 months after the company's first dated blog post. Each entry carries its grade.
- 2023CThe open-source Guardrails framework launches, as Rajpal later said; on 31 August 2023 the company's first blog post introduces Guardrails AI as its maintainer. [4, 6]
- 15 February 2024CFormal company launch and Guardrails Hub. The press release says the company closed a US$7.5 million seed round led by Zetta Venture Partners. [4, 5, 7]
- 8 November 2024CGuardrails Pro, a managed enterprise service built on the open-source framework, goes on sale. [8]
- 13 and 14 August 2025CSnowglobe, a simulation service for testing chatbots, is introduced with case studies from MasterClass and Changi Airport Group. [10, 11, 12]
- 7 October 2025CThe archived Snowglobe pricing page lists self-service at US$0.25 per generated message and an Enterprise plan. [9]
- 9 September 2026AHarvey announces that it has acquired Guardrails AI, its fourth acquisition of 2026; the team joins Harvey's Applied AI work. [1, 3]
Why did Guardrails AI work?
Four causes, in the order they mattered. Each one is a reusable lesson.
Open source first
The framework and the Hub were free, so developers could add checks to a model call without a sales contract. Guardrails Pro was, in the company's words, the answer to platform teams that wanted the open-source framework without the work of running it. Harvey still described the company by its open-source framework and its download count.
Small models for each check
The company wrote that many of its guardrails are small fine-tuned machine-learning models chosen for accuracy and speed, run on GPUs in the customer's own cloud. A narrow model for one check is cheaper and faster than asking a large model to judge every answer.
From runtime checks to testing before release
Snowglobe moved the work earlier: instead of blocking a bad answer in production, it generated many realistic conversations before release. The Snowglobe post compares this with the simulation testing of self-driving cars. Harvey bought the company for this work on agent reliability.
How it was built
These points are the company's own statements on its blog and archived product pages, grade C. [7, 8, 13, 10, 9, 11, 15]
- A Python framework: validators are combined into guards around a model call, and validators from the Hub are installed with a command-line tool and an API key.
- Guardrails Pro deploys a central guardrail service in the customer's virtual private cloud (VPC), or fully hosted, with machine-learning models on GPU infrastructure, an OpenAI-compatible endpoint and a monitoring dashboard.
- Snowglobe builds user personas from the customer's own context, runs multi-turn conversations with the customer's agent, and scores them with automated judges based on language models; datasets can be exported to Hugging Face.
- The Changi Airport test ran about 100 multi-turn conversations per topic, and the case study says that aligning the automated judges with human reviewers was the hard part.
- The website is built with Framer; demos were booked through Cal.com and support ran through a Discord community.
Tools named in the sources
Each one opens its page in the Tools tab, with every story that uses it.
Who tried this before Guardrails AI, and why did they fail?
Three companies in our database of failed startups sold developers tools that test software or catch its errors, and failed. The values below are from that database and are not independently verified.
Testmunk
The promiseAutomated testing of mobile apps on real devices.
Why it diedBetter-funded competitors and frequent operating-system updates, which forced constant rework.
Proxino
The promiseReal-time error detection for JavaScript in the browser.
Why it diedMany competitors and no money left to keep up.
Verelo
The promiseWebsite monitoring for uptime, malware and performance from several countries.
Why it diedIt used up its budget in six months without convincing businesses to pay for specialized monitoring.
The failed startups come from our database of 1,749 dead startups, retold here in our own words. Its entries are AI-assisted summaries of public sources, so these cards carry no evidence grade.
What changed
Testmunk, Proxino and Verelo sold testing and monitoring as a paid tool from the start, and they lost to larger competitors or ran out of money. Guardrails AI gave its core tool away as open source, so adoption did not depend on sales, and charged only for the managed and testing products that large companies need.
The other difference is timing. Language models produce wrong or unsafe answers that ordinary software tests do not catch, and Harvey calls Guardrails' framework the first open-source AI guardrails. The end was a sale, not a shutdown: a buyer with its own agents took the team and the know-how.
Can one person copy Guardrails AI now?
The open-source route, yes. The scale, hardly.
A solo developer can still publish a free library that solves one painful problem for AI developers and sell the hosted or enterprise version. The company in this story raised US$7.5 million, as it said in its press release, and its CEO had worked in AI for a decade, including self-driving cars, so the research and the enterprise sales are hard to copy alone.
Where it is still open: Testing and checking for AI agents in fields with high stakes, such as law, finance and healthcare administration, where buyers ask how an agent will behave before they use it.
- 1Pick one failure of AI products that developers meet every day, such as leaked personal data or invented facts.
- 2Publish the fix as a free library with a catalogue that others can add to.
- 3Sell the version that companies cannot easily run themselves: hosted, inside their cloud, with monitoring.
- 4Publish case studies with named customers and their own words.
What could go wrong for Guardrails AI?
The risks are the mirror image of the causes above.
What is not verified
No sale price, revenue figure or count of paying customers was found. We read Harvey's announcement, the company's own post, Dealroom's item, the 2024 press release, the blog, the home page and archived copies of the pricing and Pro pages; none gives an amount. The figure of more than 250,000 monthly downloads is Harvey's statement and was not checked. The headcount at the sale was not published. The home page footer shows SOC 2 Type II and HIPAA badges with no link to an audit report or a trust page, and we found no source that confirms them. The Changi Airport case was part of a government testing pilot, and the page does not say whether it was paid. A reader could look for a post by the lead investor, Zetta Venture Partners, or a later interview with the founders.
Competition
On this site, Agnost AI (monitoring and custom models for AI agents) and Kyvvu (a security layer that checks each action of an AI agent) sell to the same AI teams.
Dependence on model providers
The home page says the framework works "across any LLM", and the Pro page says guardrails run "independently of your AI provider" through an OpenAI-compatible endpoint. Many checks are the company's own small models. Snowglobe generates personas and judges conversations with language models whose providers the company does not name.
The products after the sale
On 7 November 2026 the site still showed Snowglobe with a "Try Now" button and the open-source framework, under a notice of the Harvey deal, but the pricing pages were gone. Harvey did not say whether it will keep selling Guardrails Pro or Snowglobe; we did not test a signup.
Sources
The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.
- A Harvey blog, 9 September 2026: Guardrails AI Joins HarveyRead in full. The buyer's own announcement. Source for the acquisition, the co-founders' names, the team joining, the fourth acquisition of 2026, the download figure and the investors.
- C Guardrails AI blog: Guardrails AI Joins Harvey (read 7 November 2026)Read in full. The same text as Harvey's post, on the company's own site.
- B Dealroom News, 10 September 2026: Harvey acquires Guardrails AI in its fourth deal of 2026Read in full from the copy the pipeline downloaded on 7 November 2026. Reports the buyer's statements and quotes both chief executives.
- C The AI Insider, 16 February 2024: Guardrails AI is Solving the LLM Reliability Problem for AI Developers With $7.5M in Seed Funding (GlobeNewswire press release of 15 February 2024)Read in full. The company's press release. Source for the formal launch, the 2023 start, the seed round and Rajpal's background.
- C Database Trends and Applications: Guardrails AI Promises to Stop 'Hallucinations' Plaguing LLMs with $7.5M in Seed FundingRead in full. A news item based on the press release. Source for the 50 validators at launch.
- C Guardrails AI blog, 31 August 2023: Hello World!Read in full. The company's first dated post.
- C Guardrails AI blog, 15 February 2024: The Future of AI Reliability Is Open and Collaborative: Introducing Guardrails HubRead in full. Source for the Hub, validators, guards and the command-line installation.
- C Guardrails AI blog, 8 November 2024: Meet Guardrails Pro: Responsible AI for the EnterpriseRead in full. Source for the managed enterprise product, the private hub, the 65 or more guardrails and the small fine-tuned models.
- C Snowglobe pricing page, Internet Archive copy of 7 October 2025Read the price table and the plan lists. The live address now redirects to the Guardrails AI site, which has no pricing page.
- C Guardrails AI blog, 14 August 2025: Introducing SnowglobeRead in full. Source for how Snowglobe works and the comparison with self-driving simulation.
- C Guardrails AI blog, 14 August 2025: Testing Changi Airport's ChatbotRead in full. Case study with a quote from Changi Airport Group; source for the test design and the automated judges.
- C Guardrails AI blog, 13 August 2025: MasterClass' need for synthetic dataRead in full. Case study with a quote from MasterClass's head of AI.
- C Guardrails Pro page, Internet Archive copy of 16 April 2025Read in full. Source for the deployment options, the OpenAI-compatible endpoint and the provider independence.
- C Guardrails AI home page, read 7 November 2026Read in full. Source for the Harvey notice, the two case-study customers and the compliance badges in the footer.
- C Snowglobe page, read 7 November 2026Read in full. Source for the live product page, the Try Now link, Cal.com and Discord.
- C Database of failed startups, local exportLocal export. AI-written summaries, not independently verified. Source for the failed predecessors.
Run this company? Send corrections or a verified figure to aiolum@livyden.com