With the rapid popularization of generative AI, the challenges faced by enterprises when adopting AI have shifted from the early question of "can it be used?" to "can we trust it?" In the past, many AI models, even without accurate information, would still give seemingly confident "illusionary" responses, which often led to serious operational crises in real business scenarios.
In response, Appier shared the latest forward-looking findings of its international AI research team today (April 15), emphasizing that the key to the future development of Agentic AI lies in giving the system "self-awareness." Through four key technologies, Appier enables AI to accurately ask questions, assess risks, and accurately grasp its own capability boundaries at critical moments, driving commercial AI to evolve from a simple generation tool into a reliable decision-making brain for enterprises.
Solving the pain point of AI "hard-to-answer" problems for enterprises
From customer service misinterpretations to fabricated event content, the situation where AI, unaware of its own capabilities, insists on providing answers has already cost many companies dearly.
Appier CEO and co-founder Chih-Han Yu pointed out that future AI agents will accelerate the connection between people, tools, and software, forming a more complex "Agent society." The key to whether companies can seize the initiative in Agentic AI lies in whether AI has trustworthy decision-making power.
He further emphasized that Taiwan not only has hardware advantages in the global AI industry chain, but also has the software capabilities to solve world-class problems. Appier is bringing trustworthy agentic AI into real business scenarios through its massive data barriers and domain knowledge.
Four key technologies: bridging the final mile to improve return on investment.
Appier has long been committed to AI industry-academia collaboration, having published over 400 papers in authoritative international associations such as NeurIPS, ACL, and EMNLP. Addressing the current obstacles to enterprises adopting Agentic AI, Appier proposes four core solutions:
Preventing "catastrophic amnesia" in continuous learning: Many models "forget" their original logical reasoning abilities after being fine-tuned for specific tasks. Appier proposes a stable fine-tuning method that identifies and avoids highly perplexing tokens from the source. Data shows that this method not only requires only 8 minutes of preprocessing time but also reduces the degradation rate of non-target tasks to near 0%, enabling AI to learn stably and efficiently in enterprise environments.
"Precise Questioning" that Rejects Blind Guessing: When faced with vague instructions, AI often engages in subjective blind guessing or excessive questioning, causing confusion for users. Appier incorporates verifiable external feedback and performs cross-validation with other large language models (LLMs) before answering, making AI's questions more precise and necessary, improving the balance between task accuracy and user experience by more than 30%.
Context-oriented "risk assessment": Appier employs a "skill decomposition" architecture, separating problem-solving, confidence assessment, and expected value decision-making. This enables AI to possess risk perception capabilities, determining when to answer, refuse to answer, or report to a human based on the context, successfully reducing high-risk expected losses by 60% to 70%.
Low-cost inference "ability calibration": Unlike the traditional approach that only considers the correctness of a single answer, Appier has created a brand-new ability calibration mechanism that allows AI to estimate the "probability of answering correctly" before actually answering, thus more accurately defining its own ability boundaries. This inference cost is extremely low, even less than 1 token.
From B2C to B2B: The "mistake-proofing" performance in real-world scenarios
These cutting-edge technologies have already been integrated into the Appier AI Agent's operational processes and are reflected in the interaction scenarios between businesses and consumers.
In consumer-facing Sales and Service Agents, when users ask beauty brand agents irrelevant questions such as "restaurant recommendations for Mother's Day," a discerning agent will not give random answers or fabricate product shades to cater to the user. Instead, the agent will clarify the question and redirect the user to relevant brand products at the appropriate time, reducing the risk of inappropriate interaction.
In the Audience Agent application used within an enterprise, if a marketer requests to analyze audience data from the past five years, but the system can only access data from the past year, the Agent will not force an answer. Instead, it will truthfully inform the user of the data limitations, proactively clarify the conditions, and provide alternative solutions that include explanations of their advantages and disadvantages.
According to Appier's actual verification, the current AI Agent can already successfully block 80% of risk responses for enterprise users, and will continue to be optimized with data iteration. From "all-purpose tool" to "AI colleague", Appier emphasizes that what enterprises need is no longer a system that dares to answer everything, but a smart agent that knows "what it knows and what it doesn't know".
