Blog

Large Language Models LLM and AI Artificial Intelligence Concept

<h3>What That Means for Your Enterprise Productivity</h3>Large Language Models (LLMs) have felt like a non-stop rocket ship since ChatGPT first launched, delivering exponential improvements every few months. For the last couple of years, the right strategy was often to wait and see—after all, a groundbreaking update was usually just around the corner.But that rapid pace of progress may be starting to plateau.Here at Nearly Human AI we specialize in high-accuracy, high-ROI enterprise AI solutions. And in the last three to six months, we've started seeing indicators that the frenetic pace of language model improvement is slowing down.The Plateau is Here (For Now)While we’ve seen major new releases, including the most recent versions, their underlying quality hasn't shown the same fundamental leap we saw between earlier generations. The core technology of LLMs appears to be reaching an optimal level of quality—one that still has a noticeable "error gap" we all recognize when using these tools.This is a natural part of machine learning's evolution. What we're observing now is a shift: companies are focusing their efforts on building applications around the existing LLMs rather than overhauling the underlying models themselves.For the enterprise, this marks a critical inflection point in your AI strategy.&nbsp; The New Strategy: Stop Waiting, Start IntegratingFor two years, the substantial, game-changing releases made waiting worthwhile. That benefit is largely gone. Over the next year, until we see another true technological breakthrough, the LLM capabilities we have right now are essentially the ones you'll be working with.This means now is the time to act.While the current technology has a quality gap, it also has immense, untapped utility. The key to extracting value is changing your approach from waiting for better models to mastering implementation and application design around the models you have.Here's how to do it: 1. Shrink the Problem and Narrow the ScopeTo achieve high ROI, you can't just throw a general-purpose LLM at a massive, complicated problem. Focus on:Specific Use Cases: Identify high-volume, repetitive tasks that can be clearly defined. Data Curation and Quality: Invest cycles in ensuring the data fed into the system for those specific use cases is clean, relevant, and high-quality. This is how you drive accurate outputs.2. Bring Human Experts Seamlessly into the LoopThe reality is that LLMs will hallucinate or make errors. The most effective way to address this is by designing systems that integrate human expertise to manage risk:Drive Hallucinations to Zero: We specialize in bringing human experts "into the loop" to review model outputs specifically when there's a risk of error. Massive Workload Reduction: Our policy and procedure bots, for example, can see 80-85% automation on sophisticated problems. This doesn't replace the expert; it frees them up to focus on the crucial 15-20% of issues that truly require advanced knowledge and experience.3. Shift Your Investment from "Improvement" to "Integration"If you're already working on LLM projects, stop spending excessive cycles trying to wring out tiny improvements from the base model. Instead, focus on integrating the current level of quality into your enterprise systems in a way that generates real value.And if you haven't started yet—the time is now. You won't get a "late start advantage" anymore by waiting for the next huge, labor-saving model to drop.&nbsp; The Future is Bolt-OnBy building your core systems now, you're not locking yourself in. When new breakthrough technologies eventually emerge, they will be able to bolt onto your existing, production-ready systems, expanding the set of use cases that have a positive ROI.Don't wait any longer. Move these projects into production, use the current state of the technology, and start extracting massive value today.Need help defining your highest ROI automation projects?&nbsp; Schedule a consultation. We’d be happy to help you with your enterprise AI strategy.

Are LLMs about as good as they are going to get for now?

Businessman using mobile online banking and payments, Online shopping and customer network connection.

Recently, I discussed an MIT study revealing that a staggering 95% of AI projects yield no measurable ROI. This is a stark contrast to what we're seeing: our AI implementations are delivering over 400% ROI. This discrepancy is unsettling, but I believe I've identified the core reason.Here's the bottom line upfront: If you've been working on an AI project for two months or more and haven't yet achieved a measurable 200% ROI, something is fundamentally wrong. The number one reason we're observing this failure in the banking sector is simple: going it alone.<h3>Pitfall #1: The Perils of Going It Alone</h3>Companies struggling to realize ROI from their AI systems are typically attempting in-house solutions. This often involves either existing internal teams or expensive consultants who aren't deeply invested in the long-term success of the organization. The result? Projects drag on endlessly, delivering negative ROI.Think about it: you wouldn't develop your core banking system or CRM in-house, and for good reason. Enterprise-level AI isn't much different. While using tools like Copilot or ChatGPT feels easy, it's akin to having your team manage customer relationships on a shared Excel spreadsheet – it simply doesn't scale.Why does this "go it alone" approach consistently fail?<ul><li>Overlooking Critical Aspects: Initial in-house implementations often miss essential components like proper software tooling, seamless integrations, robust data pipelines, and the long-term maintenance required as data and processes evolve. The initial "zero to one" might seem straightforward, but achieving the necessary quality for ROI quickly becomes challenging.</li><li>Slow Progress: Internal teams often work on AI projects part-time. They're solving complex problems for the first time, and the learning curve is steep. This inevitably leads to sluggish progress.</li><li>Compromised Quality: If you're using the wrong tools or approaches (and a pro tip: if you're currently using Copilot for enterprise-level AI, you likely are), the quality of your solution will suffer.</li><li>Lagging Adoption: A significant hurdle is adoption. If AI tools aren't seamlessly built into your existing workflows, they're often side-chained, difficult to use, and consequently, go unused. Without widespread adoption, you won't see any benefit.</li><li>Missing Empirical Reporting: Crucially, empirical data and reporting are frequently absent. This leaves you in the dark, unable to determine if your AI initiatives are actually working or delivering value.</li></ul><h3>The Rule of 70: Your AI Goalpost</h3>So, how do you know if you're on the right track? I use what I call the "Rule of 70": 70% adoption and 70% automation across the entire process you're targeting, with verifiable data to prove it.We've seen this rule in action with incredible results. For instance, our first customer literally doubled the size of their bank while simultaneously shrinking their compliance and operations departments. During this transformation, the customer and employee experience improved dramatically. This level of success is not only possible but happening right now.<h3>The Hard Truth and Your Next Step</h3>Here's the hard truth: if nothing changes, the 5% of companies successfully leveraging AI are positioning themselves to either acquire or bankrupt the 95% that aren't getting results. I don't want to see that happen to you.Time is money. AI implementations can be quick, delivering value rapidly. Waiting for results is costing you hundreds of thousands, if not millions, of dollars per year, not to mention lost ground to competitors.If you're reading this and thinking, "I don't believe that's true," consider which category you might be falling into. But if you resonate with these challenges, reach out. Visit our website or message me on LinkedIn, and let's book a call. I want to help you. I can either assist you in fixing your current project or provide an out-of-the-box solution that will generate immediate ROI.Don't wait. Be realistic about your AI investments and ensure you're getting those significant returns. <a data-mce-href="https://calendly.com/jamison-rotz/discovery" href="https://calendly.com/jamison-rotz/discovery" title="Book a call">Book a call</a> if you need any help whatsoever. We can help you fix a flagging existing project or offer one of our AI products that can get you going within weeks.

The Rule of 70 - Your AI Goalpost

AI trading bot is a piece of software that analyzes market data and executes trades automatically using artificial intelligence algorithms. Business investment concept. 3d rendering

Earlier this summer, I had the privilege of joining the American Society for AI at the MIT Media Lab. It was an incredible experience to see some of the brightest minds working on cutting-edge projects. This week the MIT Media Lab released a study with a striking finding: 95% of enterprise AI pilots are failing.By “failing,” the study means either never making it into production or never delivering measurable business value.As someone who works daily with organizations to implement AI in the enterprise, I wasn’t surprised. I’ve seen these challenges firsthand. But here’s the good news: while 95% may not be getting it right, the 5% who do are generating massive returns—in as little as a 3 months.<h3>So what separates the 95% from the 5%?</h3>Why Most Enterprise AI Pilots FailThere are a few common reasons AI pilots stall or collapse before generating real value:Bottom-Up and Shadow ITMany AI projects start small—an experimental tool here, a shadow IT project there. It’s like when Excel first showed up in the enterprise: useful and valuable on an individual basis, but not scalable across the entire organization. Seat licenses are expensive, adoption is uneven, and value is hard to measure when everyone uses the tool differently.Internal Projects That Stall OutAI pilots often live as “side projects” for technical teams. They’re exciting, but they don’t always reach the quality bar needed for company-wide adoption. Without seamless integration into workflows, usage remains low, and value never materializes.Poor Adoption StrategiesEven when the technology is solid, adoption can make or break success. If only a handful of staff use the system, the ROI just isn’t there.<h3>What Successful AI Adoption Looks Like</h3>At Nearly Human, we recently rolled out an AI system that reached 70% adoption in the first week—a number I’d never seen before with any enterprise software. Even better, adoption grew over time instead of falling off after launch day.How did that happen? Two critical factors:High-Quality User Experience: We invested upfront in getting the data, implementation, and linguistics right so the AI was genuinely easy to use. The first impression mattered—and it was excellent.AI Embedded Into Workflows: Instead of building a side tool with a separate user experience, we put the AI directly into the heart of the enterprise process. Employees interacted with it because it was part of their natural communication and job flow—not an optional add-on.The result was less friction, faster problem-solving, and more consistency across processes. Staff wanted to return to the system because it made their work easier and better.The ROI of Enterprise AI Done Right When AI systems are implemented correctly—with the right tools, workflows, and adoption strategies—the ROI is staggering. We’re talking 300–600% within a single quarter.But to get there, organizations need to:Work with experienced partners who know which AI tools succeed and which don’t. Choosing the wrong platform can mean building a “bridge to nowhere.”Avoid building everything in-house. Development and maintenance take time and specialized expertise. Updates can improve one area while breaking another if not handled carefully.Factor in maintenance costs. Internal projects often fall into sunk-cost bias, limping along because teams don’t want to admit the investment isn’t paying off. A holistic cost-benefit analysis is essential.How to Move From the 95% to the 5% If your organization is struggling with AI pilots, you’re not alone. But you don’t have to stay in the 95%.<h3>To succeed with enterprise AI adoption:</h3><ol><li>Partner with experts who have deployed successful AI systems.</li><li>Focus on quality and user experience, not just technology.</li><li>Embed AI directly into workflows instead of bolting it on.</li><li>Look beyond the pilot phase to long-term maintenance and ROI.</li></ol>The organizations that follow these steps are seeing transformational outcomes in both customer and employee experience.<h3>Final Thoughts: Making AI Work in the Enterprise</h3>AI in the enterprise isn’t about flashy demos or hype—it’s about embedding high-quality, well-maintained systems into real workflows to deliver measurable business value.If you’re struggling with your AI pilots, let’s talk. You can <a data-mce-href="https://calendly.com/jamison-rotz/discovery" href="https://calendly.com/jamison-rotz/discovery" data-mce-selected="inline-boundary">book a call</a>.&nbsp; For community banks, we already have a product that can be deployed in just one week, offering a low-cost, low-risk way to improve customer and employee experience.New verticals are coming soon, so stay tuned.Until then—keep pushing toward that 5%. The ROI is real, and it’s worth it.

Why 95% of AI Pilots Fail—and How to Achieve Enterprise AI ROI

Small language model, SLM with chipset concept. Abstract stream information, Generative AI.

OpenAI recently unveiled GPT 5, and for everyday use, it appears to be somewhat unremarkable compared to the 4 series models. That said, initial reviews are touting GPT 5 as an absolute game-changer for coding. This comes as a bit of a surprise, as OpenAI has frequently discussed a strategy of leveraging AI to write the code that will further expand AI capabilities, yet a coding-focused model wasn't what I had anticipated from this particular release based on the chatter leading up to it. This shift alters my perspective on the future trajectory of AI. We've been involved in developing "AI" automations for the past 12 years, long before we collectively began referring to machine learning systems as "AI." The statistics-based methods inherent in machine learning have consistently struggled with what I call the "quality gap"—the persistent 10-20% of the time that models tend to get things wrong. This gap is particularly stubborn due to edge cases—rare circumstances that are crucial to handle correctly but are, by their very nature, on the edges of the statistical bell curve. These are inherently difficult to address using traditional machine learning approaches. Attempting to balance data to account for these outliers often poisons the "golden middle" of the bell curve, thereby degrading the model's performance on typical operations. We've typically mitigated this by implementing rules-based backstops around machine learning systems and/or through human-in-the-loop methods. Both approaches are costly and not particularly scalable. I always viewed this as a significant, potentially insurmountable hurdle on the path toward Artificial General Intelligence (AGI). It's a foundational problem for machine learning that seemed unlikely to be resolved simply by adding "more data." With the release of GPT 5, I'm beginning to envision a world where we could establish a fairly scalable ecosystem. In this system, machine learning models could handle the bulk of our tasks, while specialized expert systems could be GENERATED to address the specific edge cases associated with various use cases. While this isn't the fully scalable solution we often imagine when contemplating AGI, it could dramatically accelerate the deployment of highly reliable AI implementations. This aligns perfectly with the Nearly Human and Manada Technology Enterprise AI model, which is incredibly exciting to me. What remains to be seen, is how this development will enable OpenAI and other hyperscalers to advance agent software. This is an area where it feels like there is still a considerable distance between our current capabilities and the envisioned potential of AI agents. It's truly an exciting time to be innovating in the technology space!<a data-mce-href="https://calendly.com/jamison-rotz/discovery" href="https://calendly.com/jamison-rotz/discovery" data-mce-selected="inline-boundary">Book a call with us!</a>

GPT 5 and what it means for the future of AI

Mature couple meeting financial advisor for investment

<h3>Let me be direct: </h3>I believe most AI experts get it wrong when it comes to community banks — because they misunderstand what community banking is really about. Community banking isn’t about technology first. It’s about relationships. It’s about trust built over time. It’s about knowing your customer’s name — and their story. It’s about a real human being on the other end of the line who cares. That human connection is not a cost center. It’s the very core of your brand. And when AI experts push an “AI-first” strategy, they miss the point entirely.<h3>The AI-First Strategy Won’t Work in Community Banking</h3>It might work for tech giants or call center factories — but not here. Why? Because your value isn’t in scale. It’s in relationship capital. And that capital erodes the moment you replace a caring conversation with a cold script or a chatbot. At Nearly Human, we believe something very different: AI should never be the star of the show. If this were a movie, the customer would be the main character. The human banker would be the guide. And AI? It’s the quiet, loyal supporting actor — making everyone else better.<h3>Nearly Human OpsPilot Was Built for a Human-First World</h3>We don’t put bots between your people and your customers. We put AI behind your people — like a co-pilot who never sleeps. That means: <ul><li>Context is ready the moment the call begins</li><li>Reps can focus on listening instead of multitasking</li><li>Emotional intelligence stays front and center</li><li>The human touch gets amplified, not replaced</li></ul>This is the human-first AI model — and it’s the only one that aligns with what community banks actually stand for.<h3>Relationships Are the Differentiator — Don’t Give Them Away</h3>Let’s be honest: Local relationships and high-touch service are the key differentiators for community banks. It’s what sets you apart in a world of impersonal, algorithm-driven interactions. Without that, you’re just a smaller version of today’s highly automated megabanks — and customers will treat you as such. Why would they choose a small bank with fewer features if it feels just as impersonal? The only path forward is doubling down on what makes you unique — the human connection — and using AI to reinforce it, not erase it.<h3>Because You Don’t Just Serve Accounts — You Serve Relationships</h3>Your customers aren’t looking for automation. They’re looking for someone who knows them. That’s why community banks thrive — because you show up when it matters. Our job at Nearly Human is to help your people show up stronger, more focused, and more supported — without compromising who they are or how they serve.<h3>Pro-Relationship. Pro-Performance. Pro-Human.</h3>We believe AI can enhance performance without sacrificing empathy. Community banks using Nearly Human are seeing: <ul><li>25–40% faster resolutions</li><li>Stronger customer satisfaction</li><li>Less burnout on the front lines</li><li>And most importantly — deeper relationships at scale</li></ul>Because when you support your people, they deliver the magic that keeps customers for life.<h3>The Bottom Line</h3>You don’t need an AI-first strategy. You need a human-first strategy — powered by the right kind of AI. Nearly Human was built for that. To protect what makes you special, and scale it smartly — without losing your soul.<a data-mce-href="https://calendly.com/jamison-rotz/discovery" href="https://calendly.com/jamison-rotz/discovery" title="Book a free consultation" target="_blank" data-mce-selected="inline-boundary">Book a free consultation</a> to learn how we help community banks stay personal, powerful, and unforgettable — even in an AI-powered world.

Why I Disagree with 99% of AI Experts When It Comes to AI and Community Banks

Remember those days? The Wild West of spreadsheets. Everyone, from the intern to the CEO, had their own personal Excel masterpiece. Budgets, trackers, project plans – all meticulously crafted, often with a hidden formula that only made sense to the creator (and sometimes not even to them a week later).<h3>The Glorious Chaos of Personal Tools </h3>Excel, in its prime, was the ultimate "bottom-up" tool. Need to track your tasks? Excel. Generate a massive SQL insert script? Excel. Create the executive dashboard? Excel. The beauty was its accessibility. Anyone could jump in, hack something together, and instantly feel like a data wizard.Sales teams loved using 17 different Excel sheets to track leads. Each rep had their own, with their own unique color-coding system and a "secret" column for "gut feeling." It was brilliant for them individually, a nightmare for the sales manager trying to get a unified forecast. "Just consolidate them!" The response was always a blank stare, followed by "But mine works for me!"<h3>Enter ChatGPT: The New Kid on the Block (with a Familiar Vibe) </h3>Fast forward to today, and we're seeing a similar phenomenon with ChatGPT. Everyone's tinkering. Need a quick email follow up? ChatGPT. Brainstorming blog post ideas? ChatGPT. Analyzing reports? ChatGPTIt's empowering, right? Suddenly, you have a personal AI assistant, ready to churn out text, code, or even terrible poetry on demand. It's the ultimate "bottom-up" AI tool.<h3>The Inevitable Scaling Headache </h3>But here's where the comparison gets a bit more… Excel-like. Just as those individual spreadsheets started to buckle under the weight of team collaboration and enterprise demands, so too will our personal ChatGPT hacks.Imagine this: A marketing team, each member using their own ChatGPT prompts and methods to generate campaign copy. Some are getting amazing results, others… not so much. The brand voice is fragmented, the messaging inconsistent. The marketing director throws their hands up and says, "Can't we just have one approved AI for this?"Or consider a support team. One person uses ChatGPT for quick answers, another for empathetic replies, a third for comedic relief. While individually helpful, the lack of a standardized, scalable approach means inconsistent customer experiences and a mountain of unmaintainable, idiosyncratic AI "workflows." Add the demands of corporate policy or worse yet compliance constraints and it all falls apart pretty quickly.<h3>The Future: From Personal Hacks to Enterprise Solutions </h3>Just as we learned to embrace proper databases, CRM systems, and collaborative tools over countless individual spreadsheets, we'll need to do the same with AI. The individual experimentation with ChatGPT is invaluable – it's how we discover what works and what doesn't. But for real business impact, we need to move beyond isolated, unscalable hacks.I’ve been working in the enterprise AI space for 12 years - starting way back when we were calling it “predictive analytics!” My companies have developed platforms to become great at picking up those diamonds in the rough and creating enterprise agents. That means setting up data governance, software maintenance process, responsibility and safety controls and putting experts seamlessly in the loop to close the AI quality gap.&nbsp;We’re seeing a lot of companies who are starting to wonder why they aren’t getting the productivity juice from their AI squeeze, and a few who are ahead of the Enterprise AI Curve seeing 400%+ ROI in just a few months.If you’re in the rarified air, congratulations! You’ll likely dominate within the next 2-3 years. If you need some help getting there, reach out to me.<a data-mce-href="https://calendly.com/jamison-rotz/discovery" href="https://calendly.com/jamison-rotz/discovery" title="Schedule a Demo!" target="_blank" data-mce-selected="inline-boundary">Would you like to see for yourself? Click here to schedule a demo!</a>

Blog

Are LLMs about as good as they are going to get for now?

The Rule of 70 - Your AI Goalpost

Why 95% of AI Pilots Fail—and How to Achieve Enterprise A...

GPT 5 and what it means for the future of AI

Why I Disagree with 99% of AI Experts When It Comes to AI...

ChatGPT is becoming the new Excel