# Instinct (AI assistant)

> Source: https://aiwiki.ai/wiki/instinct_ai
> Updated: 2026-08-27
> Fact-checked: 2026-08-27
> Categories: AI Agents, AI Companies, AI Tools & Products
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Instinct (AI assistant)." aiwiki.ai, 27 Aug 2026. https://aiwiki.ai/wiki/instinct_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Instinct** is a private-access [AI agent](https://aiwiki.ai/wiki/ai_agents) designed to act as a personal assistant across a user's applications and devices. Spear Street Technology, Inc. operates the service under the Instinct name. Founder Noah Shinn publicly introduced the product on August 26, 2026, describing an invite-only beta that users can contact by text or phone.[1][2][3]

Instinct is intended to do more than return written answers. Its current terms authorize the service to access, copy, collect, and index data from connected services, exchange data with those services, and take actions on a user's behalf. Those actions can include purchases and agreements with third parties.[3] Its broad access to personal context means that data controls and action reliability affect the consequences of mistakes.

| Item | Publicly documented detail |
|---|---|
| Operator | Spear Street Technology, Inc. d/b/a Instinct[3] |
| Founder | Noah Shinn[2][5] |
| Public introduction | August 26, 2026[2] |
| Availability at introduction | Private access through a waitlist or an invitation from an existing member[1][2] |
| Main interface | Text messages and phone calls[1][2] |
| Connected context | Email, messaging, screen, audio, location, and other applications or device data, subject to the permissions a user grants[1][4] |
| Reported financing | $250 million Series B, $350 million total funding, and a $2.5 billion valuation, according to reporting published on August 26, 2026[7] |

## Product design and access

The company's product description frames Instinct as a form of [computer use](https://aiwiki.ai/wiki/computer_use) that can work through a phone and computer rather than requiring the user to manage every task in a separate interface. It can receive instructions by text or call, use connected information to determine context, and communicate or transact with outside services. Instinct's site gives follow-ups, airport rides, and home-service bookings as examples of intended tasks.[1][3] These are company descriptions of the product, not results from an independent evaluation.

Shinn reported that early users had used the beta for road-trip planning, groceries, concert tickets, subscription cancellations, and wedding planning.[2] Those examples are founder-reported anecdotes. They do not establish a task-completion rate, the level of human intervention involved, or how the service performs across a broader user population.

The product site says Instinct's model is trained to use a phone and computer and to respond to personal context.[1] The available public sources do not support a more specific account of its model architecture. The terms also say the service may incorporate third-party software, including third-party models, but they do not identify which models perform a given task.[3]

Instinct remained limited to a private-access group when it was introduced. The company said it was adding compute capacity and offered a waitlist as well as member invitations.[1][2] The terms refer to Mac OS and mobile applications, but the public product page does not describe a general release schedule.[3]

## Development and financing

Before founding Instinct, Shinn worked as a research scientist at Sierra. Sierra's profile says his earlier work included machine-learning and programming-language research at Northeastern and MIT.[8] He was the first author of [Reflexion](https://aiwiki.ai/wiki/reflexion), a 2023 framework in which language agents retain verbal feedback in an episodic memory, and a co-author of [tau-bench](https://aiwiki.ai/wiki/tau-bench), a 2024 benchmark for interactions among tool-using agents, users, and domain policies.[9][10] Those papers establish his prior work on agents, but neither paper describes Instinct's internal design.

Shinn's August 26 introduction said the team had been building Instinct for several months.[2] Independent reporting identified the operator as a San Francisco company and described Shinn as a former Sierra research scientist.[5] Forbes reported that Shinn founded the startup in April.[6]

The startup raised several rounds while its product was still privately accessible. Forbes reported that Sarah Guo and Neil Mehta of Greenoaks led an early round at a valuation above $100 million. It also reported a $75 million Series A led by Kleiner Perkins at a valuation above $500 million.[6] Later on August 26, TechCrunch reported that Instinct had told The Wall Street Journal about a $250 million Series B co-led by Index Ventures and Benchmark. That report put total funding at $350 million and the company's valuation at $2.5 billion.[7]

## Connected data and model training

Instinct's Terms of Service and Privacy Policy were both revised on August 26, 2026. The documents distinguish connected-service data, user materials, information placed in a feature called Vault, and information received through Google Workspace APIs.[3][4]

| Area | Current disclosed treatment |
|---|---|
| Connected services | Instinct may access, copy, collect, and index data, exchange data with the connected service, and act through it. Disconnecting a service does not by itself delete already indexed input. The user must make a separate deletion request through Instinct's Workspace controls.[3][4] |
| Materials and model training | The company says it may use user inputs and outputs to provide, maintain, develop, and improve its products, including training models. Users can opt out of model training, but the opt-out applies going forward, does not cover other types of product improvement, and has an exception for material flagged for safety review. Models trained or improved before the opt-out may continue to be used.[3][4] |
| Vault | Materials placed in Vault are to be used only to provide the service to that user and are not to be used for model training.[3] |
| Google Workspace API data | The Privacy Policy says information received from Google Workspace APIs is not used to evaluate, fine-tune, train, or improve models and is not used for advertising. It says third-party processors may handle that information only to provide the service and for limited purposes such as security or legal compliance.[4] |
| Usage and device information | Depending on permissions and use, the service may process messages, email, documents, audio, account credentials, payment information, or health-related information. The policy also lists clicks, keystrokes, cursor positions, shared location, and device or log data among information that may be collected.[4] |

The Privacy Policy says Instinct does not sell user information or disclose it to registered data brokers or third parties that resell information. It separately lists disclosures to service providers, third-party model providers, business partners under stated conditions, parties with which the assistant interacts on the user's behalf, and recipients required for legal or security purposes.[4]

The deletion controls are separate from connection controls. The current terms state that Instinct may continue using indexed input from a disconnected service until the user follows the deletion process in Workspace.[3] The Privacy Policy similarly says that disconnecting a third-party integration does not automatically delete collected data. For Google Workspace, it also describes revoking access and deleting previously collected information through Instinct's account controls.[4]

## Actions and user responsibility

The terms give Instinct authority to take actions that it determines are responsive to a user's input. For a purchase, the service may share relevant information and the user's payment method with a connected service or payment provider. The terms state that the user, rather than Instinct, is the buyer and remains subject to the outside provider's pricing, cancellation, refund, and other rules.[3]

The same document appoints the service as the user's agent for agreements, commitments, transactions, and acceptance of third-party terms associated with an action. It says these commitments bind the user as if the user had entered them directly.[3] This is broader than permission to draft a message or recommend a purchase because an incorrect action can create an external consequence.

Instinct's terms warn that outputs may be incomplete or wrong and that actions may contain errors or may not be reversible. They place responsibility for financial, contractual, legal, and reputational consequences on the user. They also say users must apply meaningful human review where relying on an output or action without such review would be unlawful, unsafe, or otherwise inappropriate. The company may add safeguards or confirmation requirements, but the terms do not warrant that those controls will prevent unintended actions.[3]

## Changes to the terms and public scrutiny

TechCrunch's August 24 report quoted an earlier version of Instinct's terms that granted a broad `perpetual and irrevocable` license over user materials.[5] That phrase does not appear in the current terms dated August 26. The current document instead says the company does not claim ownership of user materials, while permitting their use for service operation, product improvement, and model training under the opt-out and Vault provisions described above.[3] The older licensing characterization should therefore not be presented as the current terms.

The same TechCrunch report documented concerns raised by individual private-beta testers. Claire Vo said Instinct continued to summarize email after she disconnected Google access. Peter Yang said he initially could not delete indexed Gmail records and later reported that the team added an external-data deletion control. Alex Cohen said he tested whether instructions placed in an email could influence the agent and deleted his account after the test. Katie Jacobs Stanton said the assistant sent an email on her behalf without first asking for confirmation.[5] These accounts describe individual experiences with an early product and are not controlled measurements of the service.

Instinct's revised Privacy Policy acknowledges related risk categories. It says an autonomous assistant may make unintended payments or communications and may encounter misleading instructions from third parties intended to influence its behavior, a form of [prompt injection](https://aiwiki.ai/wiki/prompt_injection).[4] The policy advises users to review actions and remain attentive, while the terms assign users responsibility for configuring permissions and monitoring the service.[3][4]

Agent reliability was also a subject of Shinn's research before Instinct. In the 2024 tau-bench paper, the tested function-calling agents completed fewer than half of tasks under the paper's setup, and the retail-domain `pass^8` measure was below 25 percent.[9] Those results concern the benchmark's agents and test conditions. They are not performance results for Instinct.

## Availability and evidence limits

Instinct was not generally available when Shinn announced it. Access came through invitations and a waitlist, and the company described the release as a beta while it expanded compute capacity.[1][2] Public descriptions therefore rely heavily on the company's legal documents, its product page, the founder's account, and reports from a limited group of testers.

The product's example tasks should be read as intended uses or reported experiences, not as independently established capabilities. The public sources used for this article do not provide a controlled product evaluation, reliability rate, or a technical specification for Instinct's underlying models. The terms and privacy policy provide the clearest current account of what the service is authorized to access and do, but they are contractual disclosures rather than measurements of product behavior.

## References

1. Instinct. "Instinct." Accessed August 27, 2026. https://instinct.co/
2. Noah Shinn. "I'm Noah, the founder of Instinct." X, August 26, 2026. https://x.com/noahrshinn/status/2092691344456351744
3. Instinct. "Terms of Service." Last revised August 26, 2026. https://instinct.co/terms
4. Instinct. "Privacy Policy." Last revised August 26, 2026. https://instinct.co/privacy-policy
5. Sarah Perez. "Instinct's powerful AI assistant is raising privacy and security concerns." TechCrunch, August 24, 2026. https://techcrunch.com/2026/08/24/instincts-powerful-ai-assistant-is-raising-privacy-and-security-concerns/
6. Iain Martin. "VCs Are So Obsessed With This AI Assistant That Its Valuation Jumped Fivefold In Weeks." Forbes, August 26, 2026. https://www.forbes.com/sites/iainmartin/2026/08/26/vcs-are-so-obsessed-with-this-ai-assistant-that-its-valuation-jumped-fivefold-in-weeks/
7. Lucas Ropek. "Viral AI startup Instinct has raised $350 million at a $2.5 billion valuation." TechCrunch, August 26, 2026. https://techcrunch.com/2026/08/26/viral-ai-startup-instinct-has-raised-350-million-at-a-2-5-billion-valuation/
8. Sierra. "Noah Shinn." Accessed August 27, 2026. https://sierra.ai/uk/author/noah-shinn
9. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. "tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains." arXiv:2406.12045, submitted June 17, 2024. https://arxiv.org/abs/2406.12045
10. Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. "Reflexion: Language Agents with Verbal Reinforcement Learning." arXiv:2303.11366, revised October 10, 2023. https://arxiv.org/abs/2303.11366