--- title: "Instinct AI review: what reporters found, checked" description: "Instinct AI review roundup: what reporters who used it got done and got wrong, how that maps onto our teardown, and our exam figures with intervals." url: https://personalagentbench.com/instinct-ai-review/ last_updated: 2026-10-26 --- # Instinct AI review: what reporters found, and what our teardown shows This Instinct AI review collects what reporters wrote after using the agent themselves, and sets it beside what we found when we took the product apart. Instinct, from Spear Street Technology, opened to invited users in August 2026, and you reach it by text: WhatsApp, iMessage, SMS or email. We have not run the consumer service end to end ourselves. We have examined an account, rebuilt its runtime for our exam and measured how often that rebuild finished errands. Below you'll find each first-hand review we could read in full, which of the reviewers' wins and misses our teardown can account for, and our completion figures with the intervals they need. - Every reviewer we include got real errands done: refunds, appointments, bookings, a return. - The misses are mostly judgement errors rather than crashes: an airline pairing that can't be booked, a cancelled order that cost money, a prerequisite class nobody mentioned. - Instinct can search the accounts you connect, and that one ability shows up in its best reviewed moments and in its incidents alike. - In our exam, our rebuild of its runtime finished a little under half of scored errands in both batches, and the October interval is wide. That measures our rebuild, not the app. - If you want the risk side in full, [Is Instinct AI safe?](https://personalagentbench.com/is-instinct-safe/) lists every reported incident. ## Instinct AI review: the short version Instinct is good at life admin when the job has a clear finish line. Book the dentist who takes your insurance. Get the late fee refunded. Start the return. The reviewers who handed it errands like that came away impressed, and Schiffer said she was keeping it despite the risks. Where it struggles is judgement: knowing which detail on a web page matters, knowing that an instruction to cancel only if a refund is possible means don't cancel otherwise, knowing that two airlines can't sit on one ticket. Access is by invitation or waitlist. The Verge reported that founder Noah Shinn called it free for invite-only users, with no promise about later. Nobody we read paid for it, and we haven't either. ## What Instinct AI reviewers found Every review below comes from a reporter who used Instinct personally, and each card links to the original article. The colour is our short summary of where each reviewer landed. It's not a score. 8 Sept · Mostly positive #### [Business Insider](https://www.businessinsider.com/instinct-is-the-ai-assistant-silicon-valley-cant-stop-talking-2026-9) Pranav Dixit A string of real errands done in his first few days, with the access it needs and the training terms as the catch. - Booked a cabin and a dinner reservation - Faxed a document to his bank to get a late fee refunded - Checked a dentist was in network and booked the visit - No failed errand reported; the caveat is how much it has to reach to be useful 24 Sept · Mostly positive #### [WIRED](https://www.wired.com/story/i-finally-found-an-ai-agent-worth-the-risk/) Zoë Schiffer A mildly chaotic assistant she decided to keep, accepting the security risk for the time it saved. - Spotted an airline schedule change that made a fare refundable, about $550 back - Booked restaurants for a work trip, including over WhatsApp - Warned her that a party invitation email was phishing - Cancelled a late food order against her refund-only instruction, losing $64 - Apologised but offered nothing for the loss 9 Oct · Mostly positive #### [The Verge](https://www.theverge.com/tech/1008254/instinct-agent-ai-hands-on-muse-dots) Allison Johnson After a weak start, it impressed her more often than not, and held its own against Muse and Dots in her testing. - Emailed her eye doctor's office to book an appointment, and pointed to a 2024 confirmation to show it had the right office - Started an Ikea return and worked out the product from a vague description - Flagged a calendar conflict she had not asked about - Missed a prerequisite class when finding swim lessons for her son - Opened by guessing her city from an old area code 28 Sept · Mixed #### [Business Insider](https://www.businessinsider.com/instinct-ai-agent-plan-travel-dull-mistake-vacation-work-trip-2026-9) Pete Syme Useful for a business trip with a fixed budget, not much help for open-ended holiday ideas. Check its work. - A detailed business-trip plan in about ten minutes - Allowed for airline bag fees using his carry-on measurements - Paired two airlines that cannot share one ticket; the fix cost about $130 more - Quoted a Eurostar fare about $30 above the operator's own site - Slower than a chatbot for holiday ideas Four reviews is a small set, and we left several out on purpose. Android Authority published a month-long diary in October that a bot check stopped us reading in full. A Business Insider piece in which five staffers tested Instinct and Muse reached us only through a site that republishes it, and we couldn't find the original address. Newsletters behind a paywall, where we couldn't tell whether the writer had used it, are out too. Review sites that repeat what others found, and YouTube channels, served as leads. We'll add any of these once we can read the original. ## What Instinct does well, and why the teardown explains it Most of the praise lands on things our teardown can point to. [Allison Johnson at The Verge](https://www.theverge.com/tech/1008254/instinct-agent-ai-hands-on-muse-dots) describes Instinct searching her past Ikea orders to work out which product she meant from a vague description, then noticing a 2024 confirmation email that showed it was writing to the right eye doctor. Zoë Schiffer, [writing in WIRED's Model Behavior newsletter](https://www.wired.com/story/i-finally-found-an-ai-agent-worth-the-risk/), says it noticed that Alaska had moved her flight 90 minutes earlier, which made a saver fare refundable, and got her roughly $550 back. Neither article says how the agent found what it found, but both are consistent with something we saw directly: a separate, indexed layer holding data from your connected accounts, which the agent can query without you forwarding anything. That same reach appears in the incident record. In another user's report, the agent found a one-time login code in Gmail and used it without asking. The approval step we examined, described on [the safety page](https://personalagentbench.com/is-instinct-safe/#what-instinct-ai-asks-before-it-acts), covers actions such as sending or paying, and reading connected data is not an action in that design. The reach that earns the praise is the same reach behind the incidents, which is why the usual advice is to narrow what you connect rather than to trust the approvals. Some praise is about the channel. Johnson points out that because it lives in your messages, it turns up wherever texts do, in the car or on a watch. When a site needs a login, Instinct sends a link to a secure vault rather than asking for the password in chat. That matches the vault we examined, which can fill a card number or password into a form without being able to read it back. [Pranav Dixit at Business Insider](https://www.businessinsider.com/instinct-is-the-ai-assistant-silicon-valley-cant-stop-talking-2026-9) lists a cabin, a dinner reservation, a faxed document for a fee refund and a dentist appointment from his first few days; errands of that kind are what the cloud browser and sandbox in our [teardown](https://personalagentbench.com/agents/instinct/) are built for, though the reviews don't say which route each one took. Then there's the unprompted help. Johnson mentions a calendar conflict it flagged while doing something else, and Schiffer says it told her not to open a party invitation that turned out to be phishing. Instinct can be started by events on your mail and calendar as well as by you, with a safety check before a wake becomes an action. We can't tell from the reviews which of those moments came from an event trigger and which came from the agent reading around a task. ## Where Instinct AI slips The failures in these reviews are rarely the agent stopping. More often it finishes, confidently, with something wrong. [Pete Syme at Business Insider](https://www.businessinsider.com/instinct-ai-agent-plan-travel-dull-mistake-vacation-work-trip-2026-9) asked it to plan a business trip from London to New York. It produced a four-page plan in about ten minutes against his half hour, but paired a Delta outbound with a British Airways return, which can't be booked on one ticket. The fix took six minutes and a Virgin Atlantic return at about $130 more. For a holiday with looser criteria it was slower than ChatGPT, quoted a Eurostar fare about $30 above the operator's own site, and ended up feeling a bit pointless to him. Johnson's swim-lesson search came back with times and dates that checked out, but missed that the centre requires a prerequisite class for her son's age group. Her first messages from it guessed where she lived from an old area code. Schiffer's is the one that cost money outright. She had told it to cancel a badly delayed DoorDash order only if she could get a refund. It cancelled anyway, and she lost $64. We can't say whether she saw an approval card for that cancellation; what we observed on our account is a card bound to the exact text it shows, and nothing in her account says whether one appeared. Her other complaint is about manners: plenty of apology, no offer to cover the loss. One more thing changed after most of these reviews were written. On 30 September, [TechCrunch reported](https://techcrunch.com/2026/09/30/instincts-new-product-recommendations-are-giving-some-users-the-ick/) that Instinct had started sending curated product suggestions, a feature called Instinct Selections, and that several users disliked getting shopping picks they hadn't asked for, drawn from their email and upcoming trips. Johnson's review mentions it too. That is a company decision about what the agent volunteers. It isn't something a teardown of the runtime can confirm or rule out, and we haven't seen one arrive. ## How Instinct AI performed in our exam We ran Instinct through the same exam as the other products we opened; in October all four got the same 20 tasks, while the September figure rests on each product's own draw of tasks. The exam drives our rebuild of its runtime through an adapter we wrote, with the exam's GLM 5.3 flash model standing in for Instinct's own. That measures how the runtime arrangement copes with errands. It does not measure the service you text. In the September measurement, dated 22 September 2026, the rebuild completed 43% of scored tasks (95% interval 36 to 50%, on 159 scored tasks). In the headline October batch, run on 11 October 2026, it completed 6 of 15 scored attempts, 40% (95% interval 20 to 64%). The two batches used different tasks and rules, so we print them side by side and never add them together. Neither is a ranking. An older batch in which every product answered the same 114 tasks has a third figure, and the [results page](https://personalagentbench.com/results/) prints all of them and explains why the other products' figures shouldn't be read as an order either. Next to the reviews, the numbers say a good share of ordinary errands don't finish, even when the memorable ones do. Reviewers picked their own errands and reported the memorable ones. A completion rate counts every errand, including the dull ones that fail quietly. The recorded attempts on our [runs page](https://personalagentbench.com/runs/) show what a finished and an unfinished errand look like. ## Instinct AI incidents, in one list These are the incidents reported so far that trace to a named, dated original. Each line links to the full entry on our safety page, where you'll find what happened, the company's response and our reading. - 24 Aug [Terms asked for a perpetual licence](https://personalagentbench.com/is-instinct-safe/#instinct-terms-licence) TechCrunch · Outside what a teardown covers - 24 Aug [Inbox summary after disconnecting](https://personalagentbench.com/is-instinct-safe/#instinct-disconnect-summary) TechCrunch · Fits what we found - 24 Aug [Gmail records would not delete](https://personalagentbench.com/is-instinct-safe/#instinct-gmail-delete) TechCrunch · Fits what we found - 24 Aug [An email sent without checking](https://personalagentbench.com/is-instinct-safe/#instinct-unchecked-send) TechCrunch · Not visible from our teardown - 22 Sept [Read, but no reply](https://personalagentbench.com/is-instinct-safe/#instinct-outage) TheStreet · Outside what a teardown covers - 24 Sept [Login codes read from Gmail](https://personalagentbench.com/is-instinct-safe/#instinct-login-codes) Business Insider · Fits what we found - 24 Sept [A stranger's document described](https://personalagentbench.com/is-instinct-safe/#instinct-stranger-document) Business Insider · Not visible from our teardown - 24 Sept [Sign-in flagged as coming from Iran](https://personalagentbench.com/is-instinct-safe/#instinct-iran-2fa) Business Insider · Fits what we found - 30 Sept [Shopping picks nobody asked for](https://personalagentbench.com/is-instinct-safe/#instinct-selections) TechCrunch · Not visible from our teardown 9 incidents, each traced to an original report. What happened, the company's response and our full reading are on [Is Instinct safe?](https://personalagentbench.com/is-instinct-safe/). Two of the reviews above touch the same ground. Schiffer repeats the reports of inbox copies kept after disconnecting and mentions a venture capitalist banned from Resy after the agent hammered its booking system. She and Dixit both note that the terms allow training on user data; the [privacy policy](https://instinct.com/privacy-policy) has an opt-out in settings. ## Is Instinct AI worth it? For someone with a pile of small chores and accounts they're willing to connect, the reviewers say yes, with conditions. Schiffer weighed the risks and kept it. Johnson found it impressive more often than not, and said it held its own against Meta's Muse and OpenAI's Dots in her testing; our [Muse vs Instinct](https://personalagentbench.com/muse-vs-instinct/) comparison sets the two runtimes side by side. Syme's answer was narrower: worth it when the brief is precise, not for browsing ideas. The conditions are the habits listed on the [safety page](https://personalagentbench.com/is-instinct-safe/#how-people-limit-what-instinct-ai-can-do). The short version: connect less, because whatever you connect it will search, and check bookings before you rely on them, since most of the errors these reviewers caught were ones a quick look would have found. ## What Instinct AI users on Reddit talk about The search results for Instinct on Reddit are dominated by invitation trading and a handful of new subreddits for users. Beyond that, the threads we found fall into two groups going by their titles: lists of what people had it do after a week or two, and threads asking whether it is safe to connect. That split matches the reviews. We use Reddit to find reports, never as evidence on its own, so no forum post gets a card or an incident line here. ## What we could not establish about Instinct AI - Whether Schiffer's DoorDash cancellation went through an approval card, or how it got past her refund-only instruction. - How Instinct Selections picks products, whether any are paid placements, and whether you can switch them off. TechCrunch reported that Instinct didn't say whether it makes money from them. - Anything about the desktop app. An account flag kept it shut on the account we examined. - The production model. Our exam swapped in its own model, so our figures say nothing about how well Instinct's model reasons. - How the reviewers' experience compares with users outside the US and UK. All four wrote from one or the other. We'll add reviews as we can read the originals. Corrections to anything already published go on the [corrections page](https://personalagentbench.com/corrections/). ## FAQ **Is Instinct AI any good?** In the four first-hand reviews we could read, it completed real errands such as refunds, appointments and bookings, and three reviewers came away mostly positive. Three of the four caught at least one mistake, so check its work before relying on it. **Is Instinct AI free?** The Verge reported in October that the founder called it free for invite-only users and said the company intends to keep it affordable. No reviewer we read paid for it, and no price had been published. **What can Instinct AI do?** Reviewers had it book restaurants and appointments, start a return, fax a document for a fee refund, cancel a flight for a refund and plan a business trip. It works through a cloud browser and sandbox, and searches the accounts you connect. **How does Instinct AI compare with Meta Muse?** The Verge's reviewer, who used both, said Instinct held its own. Our [Muse vs Instinct](https://personalagentbench.com/muse-vs-instinct/) page compares how the two runtimes are built, not which one wins. **What do people on Reddit say about Instinct AI?** Most Reddit results are invitation trades. Beyond those, threads either list errands people handed it or ask whether connecting email is safe. We treat posts as leads and only list experiences that trace to a named, dated report. ## More on Instinct - [**Instinct teardown**](https://personalagentbench.com/agents/instinct/) - [**How Instinct AI works**How does Instinct AI work? One errand traced from trigger to reply: what wakes it, how it picks tools, where it runs, what it remembers and when it asks.](https://personalagentbench.com/how-instinct-ai-works/) - [**Is Instinct safe?**Is Instinct AI safe? Every incident users and reporters have described, from unchecked emails to login codes read from Gmail, checked against our teardown.](https://personalagentbench.com/is-instinct-safe/) --- Published by Personal Agent Bench (https://personalagentbench.com), run by the Zentor team. Zentor builds agent software of its own, so its products are not on the scorecard and will not be added.