Over 50% of AI Implementations Fail, Inbenta Reports

Inbenta has expanded its focus beyond strictly conversational AI to address a broader range of customer and employee experience issues. The company now empowers subject matter experts to build and scale AI solutions across various enterprise functions.

Merlin Bise, CTO of Inbenta, stated, "So, you know, obviously Inbenta's in the conversational AI space, but we've evolved to a lot more than that. So, we solve a variety of customer experience issues, employee experience issues."

Inbenta's platform enables enterprises to build automations, connect systems, and create intuitive digital guides while prioritizing safety, governance, and transparency. This approach allows users to explore AI with actionable insights in a secure and controlled environment.

Merlin Bise highlighted, "They can build automations, they can connect systems, they can build digital intuitive guides. Uh and really explore AI in a way they just can't with competitors. So, we've tried to make it really easy, but very safe. It's governed."

Many organizations implemented AI solutions due to market pressure, rather than strategic evaluation, leading to a lack of understanding regarding their purchases. This often resulted in poor data mapping and fragmented knowledge consolidation across departments. Consequently, these siloed implementations frequently create more friction for customers and employees than they resolve, with some companies finding that human labor is still more cost-effective than their poorly integrated AI systems.

Melissa Solis noted, "A lot of people felt pressure because AI became such a hot word. Everybody ran out and just started implementing solutions. First of all, they didn't understand what they were buying... they're finding out that it's actually causing many times more friction than actually picking up the phone and answering it."

Regulators are increasingly asserting their authority, demanding that companies demonstrate the source and reasoning behind AI-generated content. It is no longer acceptable for solutions to create content 'on the fly' without clear audit trails. This heightened scrutiny means companies must be able to prove the origin and validity of every AI-provided answer.

Melissa Solis explained, "Regulators are starting to have a voice and standing up... they're saying, 'Okay, it's no longer okay just to have a solution where content is created on the fly and you can't prove where that answer came from.'"

This high failure rate indicates a widespread 'reckoning' in the AI industry, where vendors often fail to deliver on promised results. The implications extend to management, who face fear of job loss due to failed deployments.

Melissa Solis stated, "Over half of these solutions are either getting repealed and pulled out or they're having to replace them or they're having to do something on top of it to try to make it work. And so when I say the reckoning is coming, it is here."

Many organizations, pressured by the hype around AI, abandoned their standard disciplines for evaluating and procuring products and services. The misconception that AI was a 'magic wand' led them to bypass critical vetting processes, prioritizing speed over robust governance and safety. Inbenta, in contrast, emphasizes built-in governance protection as essential for long-term success, recognizing that shortcuts in deployment ultimately lead to failure.

Melissa Solis commented, "It's amazing how the pressure from AI caused so many people to basically get away of all the disciplines that they normally use, right? To to buy products and services or look at vendors, and they threw it out the window because of that pressure, and we all thought that AI was a magic wand you just waved, and it and we're finding out it's not."

While general-purpose generative AI models like Claude or Hercules offer high quality, they demand significant prompt engineering to achieve specific, accurate results. For instance, building a chatbot for service requests using a top-tier large language model can take three hours, even for an expert prompt tuner. This lengthy process highlights the practical limitations of pure generative AI solutions, which often lack built-in domain context and necessitate extensive training time to perform specialized tasks effectively.

Merlin Bise illustrated, "I'm an expert prompt tuner. It took like 3 hours using Open AI's best model to get it right. I could have done that in 4 minutes, 5 minutes with OnCore."

Inbenta's OnCore system leverages embeddings while maintaining an intent-based architecture, focusing on deterministic answers over large context windows. It employs very small context windows to validate and verify information, ensuring 100% accuracy. This method not only guarantees precision but also significantly reduces costs, making it about 1/15th the price of generic Large Language Model (LLM) approaches.

Merlin Bise emphasized, "We use very small context windows to validate and verify that information. So we're 100% right. And I think that's a unique difference. We're 1/15 of the cost."

Future AI deployments will be scrutinized under two primary responsibilities: financial accountability and accuracy/safety. Organizations will need to demonstrate clear return on investment and prove where money has been allocated. Additionally, governance is anticipated to become a powerful enforcement tool, ensuring that AI solutions are both accurate and safe for widespread use.

Merlin Bise articulated, "I believe the two responsibilities are going to be financial, where did the money go and what is it producing? The second is, is it accurate and safe? Because governance is going to be a big hammer."

The head of AI stressed the importance of selecting the right tools for the right tasks, cautioning against using generative AI as a one-size-fits-all solution. There is often immense pressure to deploy generative AI for every problem, even when it is unsuitable, without fully understanding its specific requirements and limitations.

Melissa Solis described this issue, "Hey, you're not going to you go use a hammer to put a, you know, a tack into a a board, a memory board. And a lot of times because everything as it progresses whether it be a gen AI or whatever, everybody wants to run and throw a thing at a gen AI."

Many AI implementation failures stem from decisions made in silos by IT or operational teams, without the crucial involvement of governance and other relevant departments. CEOs are now actively pushing for cross-departmental representation at the decision-making table to avoid these pitfalls. Successful AI deployment requires the entire team, including compliance officers and those performing the actual work, to be involved from the outset to ensure comprehensive understanding and integration.

Melissa Solis observed, "Why wasn't the whole team involved? Cuz what we're finding out when people sit down with table with us is that, you know, IT made the decision or operational made decision without including IT and governance wasn't involved."

Projections indicate that 80% of call center roles are at risk of being replaced by AI. This trend is particularly concerning given that most of these positions are held by women, raising alarms about the potential for disproportionate gendered impact. Prioritizing the preservation of progress made for minorities and women in technology necessitates a more thoughtful and balanced approach to AI implementation to mitigate adverse effects.

Melissa Solis stated, "80% of call centers are going to be replaced. Well, guess what? Most of those are women."

Artificial intelligence has been in use for over two decades, signifying its long-term evolution rather than a recent emergence. While people were already applying AI appropriately, ChatGPT played a pivotal role in bringing the technology to mainstream awareness. Companies are advised against impulsively adopting trends without a comprehensive understanding of AI's historical development and practical applications.

Melissa Solis explained, "we've been doing AI for 20 years. Like, AI just didn't show up, but people were using it in the right way. And chat GPT really brought that to the forefront."

AI, regardless of its sophistication, lacks the capacity for human judgment and genuine empathy, making human interaction essential in critical situations. For instance, in sensitive cases like filing an insurance claim after losing a family member, individuals often prefer to communicate with a human. Efficiency, while valuable, cannot serve as a substitute for the care and understanding that only a human can provide.

Melissa Solis asserted, "There's times when I don't care how good that AI is, I want to have that caring. You know, if I've lost my son or my daughter or my husband and I'm filing an insurance claim, I don't want to talk to AI."

AI is most effectively utilized for repetitive and mundane tasks, freeing up human talent for more complex and purposeful work. Organizations can achieve better outcomes by retraining employees and reallocating them to roles that align with their intentions and strengths. Investing in training opportunities empowers employees to leverage new technologies and contribute more effectively to the company's strategic goals.

Melissa Solis advised, "retraining and re-looking at opportunities to take those humans and put them in really find their purpose and what their intention was and what they're really good at."

Customers prioritize accurate, real-time answers over complex technical infrastructure, highlighting the importance of content understanding. AI platforms do not necessitate advanced degrees for effective operation, meaning leadership's role is to facilitate the orchestration of existing internal expertise to deliver precise solutions.

Merlin Bise underlined this point, "it's engaging the actual experts within the organization on what you're trying I mean, if you have a data issue, then you're going to have to engage the people that know the content the best."

Employees typically manage a diverse set of tasks daily, often balancing preferred activities with less desirable ones. AI can be strategically deployed to automate those specific tasks individuals would rather avoid, thereby improving overall job satisfaction and efficiency. By delegating such tasks to AI, humans are freed up to dedicate their time to new, challenging, and more engaging work, fostering an environment of continuous learning and growth.

Merlin Bise suggested, "So can AI help you solve the five you don't like the best? Because guess what? You're going to like stuff that's new."