A far cry from the 600 million Alexa devices in the market, but Amazon is making progress on Alexa+. The upgraded assistant first debuted in February, when Amazon said it would roll out to users in batches over the coming months.
The new digital assistant is designed for more natural conversation and will eventually gain agentic capabilities to use third-party applications on the user’s behalf. Alexa+ should generate original responses on the fly — similar to voice modes in OpenAI’s ChatGPT and Google’s Gemini — rather than relying on the pre-programmed responses of the old Alexa and Siri systems.
However, as The Washington Post reported at launch, the shipping version of Alexa+ is missing several key features demoed in February: third-party app integrations like GrubHub, bedtime stories for kids, and gift-idea brainstorming are not yet available. When they will arrive is unclear. Jassy on the call: “We plan to add more features over the coming months.”
The Agent Accuracy Problem, Stated Candidly
In his remarks, Jassy claimed Alexa+ is among the first consumer-facing action-oriented AI agents. But he also admitted the technology remains “pretty raw” and “inaccurate.” Most multi-step AI agents today, he said, have low success rates — between 30 and 60 percent. Jassy set a target for Nova Act, the web-browsing agent powering Alexa+, to reach 90 percent accuracy in that domain. That gap between current reality and target is the honest centerpiece of Amazon’s AI story — and the reason the 100,000-user figure should be read as a beta population, not a launched product.
Amazon’s rollout pace makes Apple’s LLM-powered Siri look slow by comparison. On Apple’s earnings call the same Thursday, CEO Tim Cook, asked about the Siri delay, said the company needs “more time to finish the work.” Both giants have reportedly hit obstacles and delays augmenting traditional assistants with generative AI — the biggest hurdle being getting LLMs to use tools and integrate with other systems. That is what lets Alexa and Siri perform practical tasks like setting timers and reading texts, and it has proven harder than anticipated.
The agent-accuracy admission carries implications well beyond Amazon. When a CEO of Jassy’s stature says publicly that multi-step AI agents operate at 30-60 percent accuracy, it punctures the assumption — promoted aggressively across the industry — that agentic AI is ready for mainstream deployment. The statement functions as both an expectation reset for Amazon’s own users and an implicit challenge to competitors who have been less forthcoming about their agents’ real-world success rates. If the best-resourced companies are landing in that range, the gap between AI-agent demos and AI-agent products is wider than most marketing suggests, and anyone building on agent frameworks should be planning for the same reliability ceiling rather than assuming their use case is the exception.
Nova Act’s 90 percent target is also more specific than most accuracy commitments in the industry, and worth parsing. A 10 percent failure rate on multi-step web tasks sounds high until you consider what those tasks involve: navigating websites designed for humans, handling pop-ups and cookie banners, dealing with layout changes, and recovering from unexpected states. A 90 percent success rate on that kind of unstructured browsing would actually represent a genuine breakthrough — current web agents struggle to exceed 50 percent on standard benchmarks, and the gap between benchmark performance and production reliability is consistently wider than teams expect. If Nova Act reaches 90 percent on Amazon’s own defined task set, it will be because Amazon controlled the test conditions; whether that transfers to the open web is the harder question that Jassy’s framing quietly avoids.
The competitive timing against Apple deserves a final note. Both companies are attempting the same product transformation — replacing a scripted voice assistant with a generative AI agent — and both hit the same wall: tool use and system integration are much harder with LLMs than their providers anticipated. Amazon’s decision to ship at 30-60 percent accuracy to a gated beta of 100,000 users is a fundamentally different strategy from Apple’s decision to delay indefinitely rather than ship anything imperfect. The honest comparison is not which company is faster but which approach generates more useful training data: Amazon’s 100,000 users are generating real-world failure cases that improve Nova Act, while Apple’s delay means Siri’s LLM upgrade ships with less battlefield experience when it finally arrives. Whether that trade — shipping raw to learn faster versus shipping polished to protect the brand — proves correct will become clear over the next twelve months.
Why the 100,000 Number Is the Interesting One
The gap between 600 million Alexa devices and 100,000 Alexa+ users is not a failure metric — it is a definition of the beta. Amazon is deliberately gating access while the agent’s accuracy climbs from the 30-60 percent range toward Jassy’s 90 percent target, because shipping an agent that completes only half its multi-step tasks to a base of 600 million would destroy the brand overnight. The staged rollout is a data-collection strategy: every early user’s failure teaches the Nova Act system something a lab test cannot. Jassy’s candor about the accuracy problem is also strategically shrewd — by naming the 30-60 percent range publicly, Amazon manages expectations downward while implying the upside when agents cross the reliability threshold. If Nova Act hits 90 percent, Alexa+ stops being a beta and starts being a reason to buy an Echo. The race with Apple is less about who ships first and more about who crosses the reliability threshold without burning user trust — and Amazon is currently the only one of the two that has put a generative agent in front of real consumers at all.
