\n
![](\"https:\/\/images.ctfassets.net\/jdtwqhzvc2n1\/4P5b8GF6swZD4TbE4YQYeC\/d313c89fdbd0abca57787aca76175c61\/cfr0z3n_realistic_hyper_detailed_minimalist_sci-fi_splash_page__733c277b-dfb5-4961-b0eb-ce28078a96e6.png\")
<\/p>\n
\n<\/p>\n
For more than a decade, conversational AI has promised human-like assistants that can do more than chat. Yet even as large language models (LLMs) like ChatGPT, Gemini, and Claude learn to reason, explain, and code, one critical category of interaction remains largely unsolved \u2014 reliably completing tasks for people outside of chat<\/i>. <\/p>\n
Even the best AI models score only in the <\/b>30th percentile on Terminal-Bench Hard,<\/b> a third-party benchmark designed to evaluate the performance of AI agents on completing a variety of browser-based tasks, far below the reliability demanded by most enterprises and users. And task-specific benchmarks like TAU-Bench airline, which measures the reliability of AI agents on finding and booking flights<\/b> on behalf of a user, also don't have much higher pass rates, with only 56% for the top performing agents and models<\/b> (Claude 3.7 Sonnet) \u2014 meaning the agent fails nearly half the time. <\/p>\n
New York City-based Augmented Intelligence (AUI) Inc.<\/b>, co-founded by Ohad Elhelo<\/b> and Ori Cohen<\/b>, believes it has finally come with a solution to boost AI agent reliability to a level where most enterprises can trust they will do as instructed, reliably. <\/p>\n
The company\u2019s new foundation model, called Apollo-1<\/b> <\/b>\u2014 which remains in preview with early testers now but is close to an impending general release \u2014 is built on a principle it calls stateful neuro-symbolic reasoning.<\/i><\/p>\n
It's a hybrid architecture championed by even LLM skeptics like Gary Marcus, designed to guarantee consistent, policy-compliant outcomes in every customer interaction.<\/p>\n
\u201cConversational AI is essentially two halves,\u201d said Elhelo in a recent interview with VentureBeat. \u201cThe first half \u2014 open-ended dialogue \u2014 is handled beautifully by LLMs. They\u2019re designed for creative or exploratory use cases. The other half is task-oriented dialogue, where there\u2019s always a specific goal behind the conversation. That half has remained unsolved because it requires certainty.\u201d<\/p>\n
AUI defines certainty<\/i> as the difference between an agent that \u201cprobably\u201d performs a task and one that almost \u201calways\u201d does. <\/p>\n
For example, on TAU-Bench Airline, it performs at a staggering 92.5% pass rate<\/b>, leaving all the other current competitors far behind in the dust \u2014 according to benchmarks shared with VentureBeat and posted on AUI's website.<\/p>\n
Elhelo offered simple examples: a bank that must enforce ID verification for refunds over $200, or an airline that must always offer a business-class upgrade before economy. <\/p>\n
\u201cThose aren\u2019t preferences,\u201d he said. \u201cThey\u2019re requirements. And no purely generative approach can deliver that kind of behavioral certainty.\u201d<\/p>\n
AUI and its work on improving reliability was previously covered by subscription news outlet The Information<\/i>, but has not received widespread coverage in publicly accessible media \u2014 until now. <\/p>\nFrom Pattern Matching to Predictable Action<\/b><\/h3>\n
The team argues that transformer models, by design, can\u2019t meet that bar. Large language models generate plausible text, not guaranteed behavior. \u201cWhen you tell an LLM to always offer insurance before payment, it might \u2014 usually,\u201d Elhelo said. \u201cConfigure Apollo-1 with that rule, and it will \u2014 every time.\u201d<\/p>\n
That distinction, he said, stems from the architecture itself. Transformers predict the next token in a sequence. Apollo-1, by contrast, predicts the next action<\/i> in a conversation, operating on what AUI calls a typed symbolic state<\/i>.<\/p>\n
Cohen explained the idea in more technical terms. \u201cNeuro-symbolic means we\u2019re merging the two dominant paradigms,\u201d he said. \u201cThe symbolic layer gives you structure \u2014 it knows what an intent, an entity, and a parameter are \u2014 while the neural layer gives you language fluency. The neuro-symbolic reasoner sits between them. It\u2019s a different kind of brain for dialogue.\u201d<\/p>\n
Where transformers treat every output as text generation, Apollo-1 runs a closed reasoning loop: an encoder translates natural language into a symbolic state, a state machine maintains that state, a decision engine determines the next action, a planner executes it, and a decoder turns the result back into language. \u201cThe process is iterative,\u201d Cohen said. \u201cIt loops until the task is done. That\u2019s how you get determinism instead of probability.\u201d<\/p>\n
Unlike traditional chatbots or bespoke automation systems, Apollo-1 is meant to serve as a foundation model<\/i> for task-oriented dialogue \u2014 a single, domain-agnostic system that can be configured for banking, travel, retail, or insurance through what AUI calls a System Prompt<\/b>.<\/p>\n
\u201cThe System Prompt isn\u2019t a configuration file,\u201d Elhelo said. \u201cIt\u2019s a behavioral contract. You define exactly how your agent must behave in situations of interest, and Apollo-1 guarantees those behaviors will execute.\u201d<\/p>\n
Organizations can use the prompt to encode symbolic slots \u2014 intents, parameters, and policies \u2014 as well as tool boundaries and state-dependent rules. <\/p>\n
A food delivery app, for example, might enforce \u201cif allergy mentioned, always inform the restaurant,\u201d while a telecom provider might define \u201cafter three failed payment attempts, suspend service.\u201d In both cases, the behavior executes deterministically, not statistically.<\/p>\n
AUI\u2019s path to Apollo-1 began in 2017, when the team started encoding millions of real task-oriented conversations handled by a 60,000-person human agent workforce. <\/p>\n
That work led to a symbolic language capable of separating procedural knowledge<\/i> \u2014 steps, constraints, and flows \u2014 from descriptive knowledge<\/i> like entities and attributes.<\/p>\n
\u201cThe insight was that task-oriented dialogue has universal procedural patterns,\u201d said Elhelo. \u201cFood delivery, claims processing, and order management all share similar structures. Once you model that explicitly, you can compute over it deterministically.\u201d<\/p>\n
From there, the company built the neuro-symbolic reasoner \u2014 a system that uses the symbolic state to decide what happens next rather than guessing through token prediction.<\/p>\n
Benchmarks suggest the architecture makes a measurable difference. <\/p>\n
In AUI\u2019s own evaluations, Apollo-1 achieved over 90 percent<\/b> task completion on the \u03c4-Bench-Airline benchmark, compared with 60 percent<\/b> for Claude-4. <\/p>\n
It completed 83 percent<\/b> of live booking chats on Google Flights versus 22 percent<\/b> for Gemini 2.5-Flash, and 91 percent<\/b> of retail scenarios on Amazon versus 17 percent<\/b> for Rufus.<\/p>\n
\u201cThese aren\u2019t incremental improvements,\u201d said Cohen. \u201cThey\u2019re order-of-magnitude reliability differences.\u201d<\/p>\n
AUI isn\u2019t pitching Apollo-1 as a replacement for large language models, but as their necessary counterpart. In Elhelo\u2019s words: \u201cTransformers optimize for creative probability. Apollo-1 optimizes for behavioral certainty. Together, they form the complete spectrum of conversational AI.\u201d<\/p>\n
The model is already running in limited pilots with undisclosed Fortune 500 companies across sectors including finance, travel, and retail. <\/p>\n
AUI has also confirmed a strategic partnership with Google<\/b> and plans for general availability in November 2025<\/b>, when it will open APIs, release full documentation, and add voice and image capabilities. Interested potential customers and partners can sign up to receive more information when it becomes available on AUI's website form.<\/p>\n
Until then, the company is keeping details under wraps. When asked about what comes next, Elhelo smiled. \u201cLet\u2019s just say we\u2019re preparing an announcement,\u201d he said. \u201cSoon.\u201d<\/p>\n
For all its technical sophistication, Apollo-1\u2019s pitch is simple: make AI that businesses can trust to act \u2014 not just talk. \u201cWe\u2019re on a mission to democratize access to AI that works,\u201d Cohen said near the end of the interview.<\/p>\n
Whether Apollo-1 becomes the new standard for task-oriented dialogue remains to be seen. But if AUI\u2019s architecture performs as promised, the long-standing divide between chatbots that sound human and agents that reliably do human work may finally start to close.<\/p>\n