Foxbox Digital is an award-winning engineering-focused product development agency.

Inside the Box Foxbox Digital Blog

Versapay: Digital Payments Mobile Strategy

Discover how Foxbox Digital unpacks the MIT AI Study and the viral “95% of AI Pilots fail” claim. Learn why the stat is misleading, what AI pilots really deliver, and our perspective on building lasting value with AI.

Rethinking AI Pilot Success | Foxbox Digital 

The problem with that viral "95% failure" stat   

 by claiming that 95% of enterprise AI pilot projects fail to deliver any measurable ROI.The headline quickly went viral and fueled talk of an AI hype bubble and a parade of LinkedIn posts about how to “ensure your pilot is part of the 5% success stories!”

But a closer look reveals that 95% stat isn’t telling the full story. The study defined “success” in extremely narrow terms: only pilots that moved into full deployment and showed a profit-and-loss impact within six months counted as successes. 

By that definition, countless AI initiatives that improved efficiency, cut costs, or yielded valuable learnings were labeled failures. In fact, the MIT report’s authors offered almost no data to back the 95% figure, basing it on a small set of interviews and vague criteria.


Even if the headline was accurate, are we missing the bigger picture about where AI is, where it should fit, and what smart investments look like when nobody knows exactly what to expect next?

Even if many AI pilots aren’t immediately profitable, that doesn’t mean those projects are without tremendous value. On the contrary, constant experimentation is both productive and necessary when working with emerging technologies. 

Experimentation is always about learning and generating new insights, not a binary success/fail paradigm.
A stalled pilot can reveal integration challenges, data quality issues, or user adoption hurdles – knowledge that is invaluable as leaders continue iterating toward viable solutions.

In other words, the technology often “works”, it’s the organizational readiness and implementation strategy that lag behind. Innovation requires iteration, and a high early failure rate is normal for any nascent tech. Many executives are learning that “failure” in AI pilots often just means learning what doesn’t work so you can discover what does.

The problem, then, isn’t that pilots sometimes fail – it’s when organizations fail to learn and adapt. Some companies fall into the “experimentation trap,” endlessly running pilots that never connect to real business value or scale beyond the lab.
The recent MIT report itself concluded that the core issue impeding AI ROI is not model quality at all, but a “learning gap” for both the tools and the organizations using them. 

In practice, this means that companies aren’t effectively absorbing the lessons of their AI experiments. Rather than viewing a pilot that didn’t hit a short-term KPI as a waste, leading organizations ask: What did we 

learn about our data, processes, and people

? How can we apply that knowledge to the next AI project?
The real metric of success is how quickly a team can turn today’s pilot missteps into tomorrow’s strategy. 

When failures prompt tweaks and improvements, they cease to be failures at all – they become stepping stones to innovation.

It’s important to remember that many AI success stories emerged only after some trial and error. Far from 95% of companies getting “zero return,” there are numerous examples of AI pilots that paid off after iterative improvement, case studies via 

JPMorgan Chase – Applied AI in risk and operations and reduced financial losses while improving operational efficiency and customer satisfaction. Early prototypes were refined to achieve these gains.

Nestlé – Developed an internal generative AI assistant (“NesGPT”) that employees across departments use daily. The company says this experiment accelerated idea generation from six months to six weeks by empowering teams to iterate quickly with AI.

Novo Nordisk – Created an AI-powered documentation tool (NovoScribe) after piloting the concept. It now automates report writing, cutting resource needs by 80% and shrinking documentation time from weeks to minutes.

These cases show that real ROI is being achieved through AI, often after multiple tweaks and pilot rounds. In fact, analysts have documented 130+ enterprise AI case studies with tangible benefits and direct ROI - evidence that a blanket “failure” label is pessimistic and short-sighted.

 The common thread is that each organization treated initial setbacks as learning opportunities, not reasons to quit.

To maximize value from AI pilots (and join the successful 5% club), companies should take a strategic, learning-oriented approach:


: Define the business problem and target outcome upfront. Successful teams focus on high-impact use cases (e.g. reducing downtime or improving customer retention) rather than chasing shiny AI novelties

 Don’t consider a pilot a mere demo. Right from the start, think about how you’ll integrate the AI into real workflows if it works. This means allocating resources for process redesign, employee training, and change management – not treating AI as a plug-and-play tool

: Many pilots stumble due to poor data foundations. Ensure you have clean, unified data and the necessary infrastructure in place. AI can’t deliver value if it can’t access reliable data at scale.

Don’t judge every pilot solely on immediate revenue uplift. Look at efficiency gains, cost savings, customer experience improvements, risk reduction, etc. as signs of progress. A project that streamlines a process or provides insight can be a win even if it doesn’t boost profits in six months.

 Lack of executive support or unclear policies around AI risk and compliance can doom a pilot that is otherwise promising . Treat AI initiatives as strategic programs with C-level sponsorship, and establish guidelines so teams feel confident moving from pilot to production.

Perhaps most importantly, treat each pilot as a learning cycle. Encourage teams to document what worked and what didn’t. Even a pilot that “fails” is only a failure if nothing is learned or applied afterward.

The drumbeat of “95% of AI pilots fail” makes for attention-grabbing headlines, but it misses the bigger picture. 

Yes, most AI pilot projects won’t hit a home run right away. They’re not supposed to, at least not always in 6 months.
The organizations that embrace failure as feedback and scrutinize why a pilot underperformed are the ones that will ultimately reap the rewards of AI. 

In the end, successful AI adoption isn’t about avoiding failure at all costs; it’s about failing productively, learning relentlessly, and scaling up the ideas that stick. By staying strategic and resilient, businesses can turn those pilot “failures” into the very innovations that make the right headlines for the right reasons.

Beyond the “95% Failure” Myth: Rethinking AI Pilot Success

The past few months have brought real momentum for the Foxbox. We’re finding new ways to build with AI, building new solutions for our clients, and are welcoming three new additions to the team as we continue to build for the future. 

Founder's Update - Summer 2025

AI is no longer a curiosity at the edge of our engineering approach, it’s at the core of how we build. Not for shortcuts, to empower our engineers to move faster, think deeper, and deliver more value. 

to share more about how we're using tools, and why we're bullish on what the future holds.   


Beyond our internal adoption of AI tools, our teams are hard at work building custom AI products for clients. One recent example: an AI-powered voice assistant designed to help seniors live healthier lives. It’s highly sophisticated behind the scenes, but for users it’s an ultra-intuitive, reliable path to the answers they need. We’re thrilled with the work so far and will be eager to share more.

Welcome to the Team: Sai Balamurali, AI Engineer

We’re excited to welcome Sai as our newest AI Engineer. Sai is already hitting the ground running for our clients - analyzing data, architecting AI solutions, and helping find new ways to deliver value. With a Master’s in Computer Science from the University of Illinois Chicago, Sai brings practical machine learning experience to the engineering team and we’re thrilled to have him on board.

After cancelling our trip due to the Maui fires and almost cancelled due to a tsunami, Pat and I were able to successfully unplug 4,000 miles from home on the beautiful Hawaiian Islands of Kauai & Maui. I didn’t even bring a laptop.

I came home feeling energized, excited, and - admittedly - curious to know just what it takes to get a helicopter license.

Discover how Foxbox blends AI-driven code generation with product strategy, user insight, and market foresight to create products with real staying power.

AI can generate code. Partners build products. | Foxbox Digital

A year ago, the question was whether AI could actually 

Now, the answer is obvious: it can. The more important question is: 

AI-generated code is remarkable in its speed and accuracy. With tools like Claude Code, Cursor, and Gemini, we can go from concept to functional code in hours instead of days.

But a product needs much more than functionality to succeed. The real work is in:


Making the business case before a single line is written.


Identifying opportunities (and threats) your stakeholders may not even be considering.


Connecting the dots between user needs, market realities, and technical feasibility.


Building for what’s coming, not just what’s in the brief.

AI can refactor code, generate unit tests, and even interpret design files with surprising precision. What it can’t do is care. It can’t recognize the problem that isn’t on the requirement sheet, the one that’s only revealed in conversation, or hidden between the lines of user feedback. 

It can’t spot the bigger-picture opportunity that changes the entire trajectory of the product. It doesn’t have stakes in the quality of its output or pride in the craft that makes it possible. We do.

When success or failure impacts someone’s business, livelihood, or reputation, it affects the final product. We sweat the details. We design for resilience. We push back when the easy answer is the wrong one. The impact of having stakes in our clients success may be hard to quantify, but it shouldn’t be underestimated.

Of course, great product strategy and design centers on your user’s experience. But when teams start thinking of the 

 as the end goal, the focus can drift toward the wrong things,  like portfolio pieces instead of real business outcomes.

when they fit so seamlessly into a user’s life that reliance becomes second nature.But lasting impact comes when that experience is treated as a 

into a broader strategy that’s also informed by:

That’s the difference between a nice launch and a product with staying power.

The future isn’t AI or humans. It’s AI 

 humans.AI can generate code. But partners build products. 

 that are designed to drive measurable business outcomes.

 whose success is defined by long term viability, not overnight development efficiency.

At Foxbox, our engineers and product strategists aren’t just here to “ship faster.” We’re here to make sure what ships is the 

, built for the right reasons, and positioned to succeed long after the launch window closes.

Speed is easy. Thoughtfulness isn’t. That’s why we build with both. 

Curious what an AI-amplified approach can do for your business?    

AI can generate code. Partners build products. 

AI has moved from novelty to necessity in software engineering. Learn about how Foxbox is using tools like Claude Code and Cursor to accelerate development, maintain code accountability, and deliver greater value to our clients.  

Evolving with intention in the age of AI-amplified engineering 

A year ago, most engineering teams were still treating AI like a novelty. Something to explore on the side, not something you’d bet the business on. Most agencies were still keeping AI tools at an arm’s length, including Foxbox. 

video not  long ago  about “vibe coding”

 and generally concluded that it yielded surprisingly quality results, but was still most applicable to proof of concepts or easily automated tasks.


When updated models of tools like Claude Opus, Gemini and Cursor were released earlier this year, the agency world reached a threshold. I firmly believe we’re now on the other side of it.


So we’re adapting.  This article outlines what we’re doing now, what we’re learning in real time, and how we’re choosing to move forward as a company. It’s not a set of final answers. It’s a snapshot of how we think, how we build, and how we’re staying grounded while everything around us moves quickly.


Dependency upgrades, typically long and arduous processes, can now be completed in a fraction of the time. Which helps eliminate typical bottlenecks to innovation and new feature development.


We’re using Claude Code as a pair programmer working alongside our engineers to build new features. We lead the tool, breaking down the problem for Claude as if we were to write it ourselves. Claude writes the first pass of the code. Then we edit, review, and commit. Look out for security and performance issues. Collaborate with the tool to write unit tests. Claude Code opens pull requests and automates the arduous pull request description writing. 

Here’s the key thing: our engineers are accountable and responsible for every line of code shipped. The tool helps, but we’re accountable.

, we instruct Claude Code on exactly what we want it to look like by providing screenshots of Figma or example code. It does a very good job. Then we use our expertise to fine tune and get the UI just right.


We’re using tools like Claude Code and Cursor to 

. Instead of manually writing failing tests and debugging step by step, we feed the bug into the AI, which helps generate a test case, suggest a fix, and even write the pull request. This automates much of the traditional TDD workflow and helps us resolve issues faster and with more confidence.
These aren’t shortcuts. 

They’re scaffolding that helps our engineers move faster without cutting corners.
Of course, we don’t just hit “accept” on AI output and ship it. Human oversight is built into every step. In many ways, these tools haven’t made software easier to build, they’ve made it more possible to build exceptional software in far less time.

We’ve seen the critiques. That AI-assisted code erodes discipline. That it nudges teams toward shipping faster instead of thinking deeper. That it risks turning engineers into editors. And some of those concerns are valid.
But the solution isn’t to slow down out of fear. It’s to build systems of accountability and lean into the process. 

We’re not handing off thinking to machines. We’re extending our own capacity to think with quicker clarity and greater context. When we automate a test suite or use AI to surface a cleaner refactor, it’s not because we’re lazy. It’s because we want our engineers focused on what only humans can do: make tradeoffs, understand users and build systems that adapt . 


I’m proud that I have a development and engineering team full of people that take tremendous pride in their craft. As an engineer who had to work hard to let go of writing code, I understand the defensive instinct or hesitation lots of engineers are feeling right now. 

The craft of software engineering isn’t about elegant code for its own sake. It’s about solving real problems through thoughtful systems. Great engineering bridges the gap between what a product does and what people actually need. The best solutions are often invisible—seamless, intuitive, and built with care. 

It’s about solving problems. Bridging gaps between product and user. It’s invisible architecture and elegant functionality, using the most advanced tools at our disposal to build things that impact people’s lives. It’s about the insight, creativity and attention to detail that makes a product so seamless and intuitive that it 

That’s the craft. Engineers that learn to be less precious about their code will be the first to remaster it.


We’re choosing to ride the wave of change: document as we go, evolve with intention, and stay focused on maximizing value to our clients. That tracks with what we’re seeing. The developers on our team aren’t just getting more done. 

They’re spending more of their time doing higher-value work.
So what does it mean when all engineers become leads? What blind spots might future engineers have if they never actually learn the basics of programming? 

Nobody knows exactly what’s next. That’s the point.
We’re not trying to predict the endpoint. We’re designing for the middle. For the messy, fast-changing space where smart humans and fast machines solve problems together. 

If we’ve learned anything over the past year, it’s that you can’t wait for certainty to take action.
So that’s what we’re doing. We’re not claiming to have it all figured out. But we’re moving fast, thinking deeply, and having the conversation out loud along the way.

Foxbox in the AI Era

Recap of our 7/10/25 event in Chicago exploring AI-amplified development. From shifting engineering mindsets to emerging workflows with Claude Code, here’s what stood out.

Event Recap | AI-Amplified Development | Foxbox Digital

Thanks to everyone who joined us in Chicago for an open conversation about AI-amplified development and its role in our industry. These discussions are only becoming more important as the ground beneath software development keeps shifting. 

Looking forward to keeping the momentum going and seeing where it all leads next. 

Below is a recap of what stood out to us.


It’s clear a major cultural shift is underway: developers are learning to let go of treating code as precious. Instead of clinging to large codebases, the focus is moving toward the 

There’s a growing willingness to start over—to explore a partially working solution, then reframe or rebuild it with fresh context.

The long-term mindset: adapt and evolve with tools, but rely on your own ability to learn. Our friend Dustin Anderson had perhaps the quote of the day: 

“The model to stick with long-term is yourself.”

2. Prompting, Specificity, and Workflow Success 

AI tools like Claude Code is highly dependent on specificity and context. The better the prompt and problem framing, the better the results.

One group discussed using workflow tools (like NAM?) to compartmentalize steps and tune processes, improving outcomes from LLMs.

The importance of structure and breaking work into discrete, contextual pieces came up repeatedly.

There was discussion about potential intergenerational divides in how AI tools are adopted:

Older professionals may bring depth and experience but less familiarity with modern AI tools.

Younger folks may experiment more, but questions remain: Will they still need to learn to code? When does AI replace or augment that skill?

This led to a broader theme: When do we rely on expertise, structure, and process vs AI-driven agility?

One historical analogy was made to Visual Basic: non-programmers could build real apps—useful, if limited. We’re now facing the same blurred lines with AI.

4. Tool Fragmentation and Future-Proofing

Attendees raised concerns about tool fragmentation and uncertainty. It’s hard to say what will be around in 5 years.

Standards that scale from startup to enterprise

Balancing morale vs velocity is another leadership challenge—how to move fast while keeping teams grounded and motivated.

Claude Code impressed many and is being used creatively:


One attendee described using Cursor to write user stories, then Claude to build features.
Another mentioned Claude auto-generating commit history and invoices.

—where an AI agent orchestrates tooling.

One example involved crypto platforms using APIs like CoinGecko:
A new MCP server  was added to the agent interface, removing the need to directly interface with complex APIs. “That was a game changer." 

6. Trust, Quality, and System Reliability

Across industries, especially in healthcare and enterprise, the central challenge is 

One line summed it up: “It will always boil down to reliability and trust.”
Developers and engineers have to be accountable for code generated by AI.


How to scale processes across teams with different work styles

Some teams are exploring LLM-on-LLM supervision: using one large language model to monitor or verify the output of another.

For example: one LLM generates code, another flags issues or corrects it. This concept echoes the need for checks and balances, potentially reducing hallucinations and increasing reliability.


Tools like Claude Code’s code reviewer or Gemini code assist can be helpful for first pass reviews.


At what point does a product or idea need to transition from low-code or AI-built prototypes into a more “real” language or 

 around AI-assisted development before it scales too far without oversight?

Will non-developers continue to build increasingly complex tools—

From rethinking long-held engineering mindsets to experimenting with emerging tools like Claude Code, it’s clear the way we build is changing fast. These conversations are rooted in curiosity, healthy skepticism, and real-world application, and they’re essential as we navigate what comes next. 

Stay tuned for future gatherings as we continue to explore, question, and shape the future of AI-amplified development together.

AI Amplified Development: Beyond Vibe Coding - Event Recap

The power of predictability, helping healthcare leaders get more out of AI, and a worthwhile visit to Atlanta.

Foxbox Founder's Update - Spring 2025 | Foxbox Digital

As we navigate 2025, one thing is clear: stability and predictability have never been more valuable. In a world of economic uncertainty, companies that focus on consistent, proactive product maintenance are reaping the rewards.

Product Maintenance: A Strategic Advantage

We’ve always believed that product maintenance is more than just fixing bugs—it’s about ensuring reliability, improving performance, and keeping products aligned with evolving customer needs. Our clients are proving that a 

leads to higher user engagement, fewer costly disruptions, and a foundation for future innovation.

One of our healthcare clients recently revamped their product maintenance approach, shifting from a reactive model to a proactive one. The result? Faster feature releases, fewer production issues, and better provider adoption rates. It’s a reminder that 

 isn’t just about survival—it’s a competitive edge.

We’ve been weaving AI into nearly every aspect of our business, from how we build and maintain products to how we 

help our clients leverage AI responsibly.

Healthcare organizations, in particular, are looking for guidance on AI integration. Our team is helping clients navigate AI with confidence—whether that’s through predictive analytics, automation, or patient-facing AI solutions. It’s not just about innovation, it’s about creating practical, compliant, and ethical AI applications that drive real impact.

Healthcare: Engaging Providers & Patients

Healthcare remains a core focus at Foxbox, and we’re seeing a growing demand for better digital experiences for both providers and patients. Our team is tackling these challenges head-on, from reimagining provider workflows to building consumer-grade patient experiences.

One thing we’re particularly excited about is how  

 are enabling organizations to modernize their systems without being locked into a monolithic EHR vendor. Additionally, our 

 are helping healthcare clients explore how AI can enhance their operations, from administrative automation to personalized patient engagement.

I was lucky to spend some quality time with Elliott Torres and Brett Ryan in Atlanta. Special thanks to Bala Balakrishnan and Vio Balanoiu for helping carve us out some space to make the most of it. I worked closely with Bala, Vio, and of course Elliott at Anthem building 

. Always great to collaborate together the old fashioned way.

Speaking of reasonable things to do - we took a home-made airplane for a test drive. Thankfully it was built by a friend whose engineering brain I trust implicitly. Thanks 

As we look ahead, we’re doubling down on building products that are not just innovative, but also scalable, maintainable, and future-proof. If you’re curious to explore more of what that looks like for your business, let’s talk.