Meta’s Muse landed on iOS, Android and the web on 8 September, built to book, message and shop on a person’s behalf from inside a dedicated cloud computer. xAI’s Grok Bot has been signing into subscribers’ inboxes and CRMs since 11 August, running as a small team of persistent “bots” rather than a single assistant. Apple spent the summer rebuilding Siri around the same premise, and a fast-growing startup called Instinct has drawn both funding interest and privacy scrutiny for going further still, reading a user’s messages, screen and location to act before being asked.

The rush matters to B2B SaaS marketers less because of who wins the race to become a person’s default agent, and more because of a parallel, better-evidenced shift already under way in enterprise software buying. New survey data from G2 and TrustRadius shows business buyers bringing AI into research earlier than ever, while stopping well short of letting it authorise a purchase. The opportunity for marketers is not building for agents that transact — that capability barely exists yet, even in consumer products — but for the accurate, verifiable research they are already doing.

What Meta, xAI, Apple and Instinct have actually shipped

Meta’s Muse is the most fully specified of the four. It runs on what Meta calls Muse Secure VM, a dedicated virtual machine that houses the agent and a person’s connected data, policed by a separate “Sentinel” process that must approve anything Muse sends to the internet. Meta says Muse asks for permission before sensitive actions such as sending an email or making a purchase, can check out using a one-time-use card through Stripe’s Link, and does not feed conversations into its ad systems. It is free for most tasks, with paid tiers for heavier use, and Meta says it is coming to its AI glasses.

xAI’s Grok Bot, launched in beta on 11 August, is available to SuperGrok, SuperGrok Plus and SuperGrok Heavy subscribers and to Cursor’s Pro, Pro+, Ultra and Teams customers on desktop and iOS, with an enterprise waitlist for wider access. Rather than one assistant, it gives each user a small roster of named “bots” that share a cloud computer, sign into work tools directly, and can message each other to divide up a task. Apple’s rebuilt Siri, announced 8 June and still in developer testing, adds personal-context understanding, onscreen awareness and broader web knowledge, with a public beta due later this year.

Instinct sits apart from the other three: no company press release describing its full feature set was available to verify, and reporting is the primary source. TechCrunch reported that Instinct added its own email addresses on 9 September and phone-calling on 17 September, on top of existing access to a user’s messages, screen, audio and location. Axios and Forbes report the startup is in talks to raise at a valuation near $10bn, up sharply from earlier this year. TechCrunch, the Atlantic and Business Insider have separately reported privacy and security concerns, including terms that reportedly let Instinct retain indexed data after a user disconnects an account and grant it authority to bind users to agreements — claims the company has not addressed in any primary source found for this piece.

Where OpenAI’s personal-agent plan actually stands

The Axios story that surfaced this comparison, published 20 September, frames OpenAI as the company still to enter the race, expected to follow after hiring Peter Steinberger, the creator of the open-source agent OpenClaw. That hire is not new: Sam Altman announced it on 15 February, saying Steinberger would “drive the next generation of personal agents” and that OpenClaw would move into an independent, OpenAI-supported foundation rather than become an OpenAI product. Seven months on, OpenAI has not announced a personal-agent product, a name or a launch date.

The gap is worth holding onto. Meta and xAI moved from announcement to a product real users can install inside months; Apple went from a WWDC preview to developer testing on the same day it was unveiled. OpenAI’s timeline has run the other way; a headline hire followed by silence on shipping. For marketers weighing where to prioritise integration or content work, that makes OpenAI’s entry a plan to watch rather than a platform to build for yet.

How far B2B buyers will actually let agents go

G2’s 2026 Buyer Behavior Report, based on a June survey of 1,038 B2B software decision-makers, found that 61% of buyers currently use or plan to use AI agents somewhere in the buying process, with another 19% open to it for select use cases. But the same report drew a hard line at the point of payment: only 9% of buyers said they were comfortable letting an AI agent execute a purchase within pre-approved guardrails, and just 2% would allow one to buy without prior approval. Agents are being welcomed in as researchers, not as decision-makers.

TrustRadius’s 2026 B2B Buying Disconnect Report, drawn from a January survey of 1,862 buyers and 444 vendors, found a similar pattern from a different angle: 63% of buyers used AI somewhere in their most recent purchase journey, and of those, 94% said they fact-check what the AI tells them at least some of the time. Adoption and scepticism are rising together, not trading off against each other.

Read alongside the personal-agent launches, the two data sets point the same way. Consumer agents are still mostly asking permission before they act; enterprise buyers are still mostly checking the answers before they trust them. The near-term task for a B2B SaaS marketing team is not preparing for an agent to buy the software unsupervised — none of the four products above can reliably do that today, and buyers say they would not let them regardless. It is making sure the same product facts, pricing and integration details that a sceptical human would fact-check are the ones an AI agent finds first, and finds consistent, when a buyer sends it out to look.

Whether Muse, Grok Bot, Siri or an eventual OpenAI product becomes the assistant people trust with their calendars, G2 and TrustRadius have already measured what enterprise buyers do next: they let the agent gather the shortlist, then check its homework themselves.