On why AI fails in production, what it actually takes to scale, and the work most companies avoid. 

Vishal Sanghavi Founder & CEO, JEH Aerospace | Former COO, Tata Group (Aerospace & Defence)

Vishal Sanghavi brings two decades of rich experience at the intersection of deep manufacturing, aerospace supply chains, and enterprise transformation. As former COO of Tata Group’s Aerospace & Defence vertical and now Founder & CEO of JEH Aerospace, he has a ground-level understanding of what it takes to operationalize complex technology in industrial contexts and in this conversation, he brings that lens to the questions every enterprise leader should be asking. 

Atul Arya Founder & CEO, Blackstraw.ai | Former VP – AL/ML, Nielsen

Atul Arya is the Founder and CEO of Blackstraw.ai, a global data and AI services firm helping Fortune 500 enterprises transform complex data into measurable business outcomes. With over 20 years of experience, he has led high-impact initiatives across finance, retail, healthcare, manufacturing, and logistics, consistently bridging business priorities with advanced technology. A hands-on entrepreneur and strategist, Atul brings deep expertise in AI, analytics, and digital transformation, building scalable solutions that deliver real-world impact. His background includes advanced degrees in engineering and management, along with leadership roles across startups and large enterprises. Beyond his professional work, Atul is passionate about mentoring future leaders and advancing practical, responsible AI adoption at scale.

1: From Nielsen to founding Blackstraw in 2018, what did you see inside a company that size that most people on the outside never get to see about why enterprise AI fails?

A: Working at Nielsen gave me the opportunity to see how AI operates at an enterprise level, across different markets, departments, and business needs. And one of the biggest things I learned was just how difficult it can be to take an AI model from development into actual day-to-day operations.

We had talent, access to strong data assets, and leadership that understood the potential of AI. But even with all those resources, getting models into production was a challenge. Teams were often starting from scratch because there weren’t established frameworks they could reuse or consistent standards for what a production-ready solution should look like. 

I started thinking about how much harder that process must be for companies without Nielsen’s resources. That was really what led me to start Blackstraw in 2018. I wanted to make it easier for businesses to build AI solutions they could actually put to work. 

It’s also where our tagline, “Simplifying AI,” comes from. We’ve used it since the beginning, and it still reflects what we’re trying to accomplish today.

2: Industry research suggests the vast majority of AI models never make it to production. From your experience working with Fortune 500 clients, what are the 2 to 3 hidden blockers that CEOs still don’t talk about enough?

A: From what I’ve noticed, some of the biggest obstacles have very little to do with the technology itself. Here are three issues I’ve repeatedly seen:

1. Data ownership: What ends up happening in large organizations is that the team managing the data and the people who actually need to use it work under completely different leadership. Getting everyone aligned can be difficult, especially when departments have their own priorities. I’ve seen projects die down simply because no one could agree on who should own what. 
2. Change management: Companies put a lot of effort into developing an AI model, but don’t always think about how employees are going to use it afterward. By the time the system is ready to launch, they’re just starting to address adoption and training. People need to understand how the technology fits into their work and feel comfortable relying on it before launch. That should be part of the development process from the beginning. 
3. Long-term model maintenance: Deploying a model is one thing, but keeping it running effectively requires ongoing attention. Who’s monitoring model drift? When does a model need to be retrained? Who’s responsible when performance starts to decline? These are questions that should have answers long before launch.

3: Blackstraw positions itself as a long-term implementation partner, not just a project vendor. What actually breaks between a successful pilot and month six in production, and what does it take to stay accountable through that?

A: A lot can change between the pilot and six months in production. What worked at a smaller scale is not holding up. The person who originally championed the project may have moved on to other priorities, and suddenly, no one’s sure who’s responsible for keeping everything running. 

I think that’s where having a long-term implementation partner makes a difference. At Blackstraw, our involvement doesn’t end when the system goes live. We continue working with clients to monitor performance, address issues, and make adjustments as their needs change. 

One thing I’ve learned is that you need to plan for inaccuracies just as carefully as you plan for improving accuracy. AI systems won’t always behave exactly as expected, especially when they’re operating in changing environments. Having a process for catching those issues and responding to them is key. 

We also build a rollback plan into our deployments from the beginning. If something isn’t working as intended, there needs to be a way to reverse course without disrupting the entire operation. That gives clients more confidence to implement AI in areas where they may have been hesitant before. 

For me, accountability means being there after the launch, when the system is being used by real people and the challenges you couldn’t predict during testing start to show up.

4: Can you walk us through a real client example where the end-to-end approach delivered measurable results?

A: There was this global frozen food manufacturer we worked with. They were processing thousands of sales orders every week through manual validation, pricing checks, and inventory confirmation across emails, portals, and documents. By the time the order was confirmed, there was already less time to respond to a major restaurant chain client. And those delays could end up costing the business. 

We deployed a multi-agent AI system that orchestrated the full order workflow across their CRM, ERP, and inventory systems simultaneously. Order processing time dropped by 85%, order-related errors fell by 40%, and stock-out risk came down by 25% through proactive inventory intelligence. 

That’s what we aim for with AI: solving problems customers have been bringing up for a while and making improvements that the commercial team can actually see in their day-to-day work.

5: Gartner predicted that 40% of enterprise applications would embed AI agents by the end of 2026. We are now in the middle of that window. CXOs are being asked to move fast on agentic AI. From what you are seeing in production deployments, where does it actually work today, and where should enterprises still be cautious?

A: I think agentic AI works best right now in areas where companies are handling a lot of repetitive tasks and the decisions follow established rules — things like processing orders, flagging issues, routing customer complaints, or managing inventory alerts. These are processes where you can define what the system should do and track the decisions it makes. 

We implemented agentic AI for a large US home improvement retailer that reduced complaint resolution time by 65% and automated more than 70% of complaint handling from start to finish. A big reason it worked was that we had clear rules for how the system should handle different situations, and we could track the decisions it made along the way. 

Where I think companies need to be more careful is when they’re giving AI the authority to make important decisions without human oversight. There are limitations in how businesses monitor and manage that level of independence, and some problems don’t become obvious until the system is actually running. 

For me, the question is how much responsibility you can reasonably give an AI agent. It’s great for handling repetitive work, but there are decisions that still require human judgment. Companies need to understand where that line is before giving these systems more independence.

6: AI readiness means different things to different organizations. From your experience, what is the one dimension that enterprises most often underestimate when they begin their AI journey?

A: Almost every organization focuses on data quality, and that’s the right instinct. But what consistently catches them off guard is how much the human side of readiness matters.

I’ve seen companies with excellent data infrastructure struggle to get even one department to act on an AI recommendation because the people using it simply didn’t trust the results yet. The analyst who built the model and the ops manager who was supposed to use it were never brought into the same room. The system wasn’t wrong often, but when it was, nobody explained why. And that was enough to erode the confidence of the people whose behavior actually needed to change.

Data gets you to the starting line. But building operator trust is what gets AI to actually change how decisions get made. From what I’ve seen, most companies don’t think about earning that trust until much later in the process, when it really should be a priority from the beginning.

7: AI is changing the nature of work across enterprise functions. What conversation should leaders be having with their teams right now, and where does resistance to adoption typically come from?

A: A lot of the resistance comes from middle management, and I can understand why. These are people who have spent years learning how the business operated, solving problems, and making decisions based on their experience. Now, they’re being told that AI can take over some of that work. I think it’s natural for them to question what that means for their role and whether their experience will still be valued. 

Leaders need to involve those managers in the process rather than simply introduce AI and expect them to use it. A regional manager, for example, could help identify when the system makes a mistake and provide feedback to improve it. When people have a say in how the technology works, they’re more likely to trust it and see its value.

8: Blackstraw works across CPG, retail, market research, aerospace, and more. How do you define production-ready AI in 2026, and what does it actually take to get there?

A: Just because an AI model has been deployed and is performing well doesn’t necessarily mean it’s ready for everyday business use. There’s a lot more that needs to be in place. Companies need to be able to monitor the data going into the system and catch problems early. They also need a plan for retraining the model when business conditions change, a process for handing decisions over to a person when the AI isn’t confident in its results, and someone who’s actually responsible for how the system performs. 

On top of that, there needs to be a clear record of the decisions it makes, both for internal reviews and compliance purposes. Without those things, a model might work well during testing but struggle once it’s being used across the business. 

We built a product intelligence platform for a global CPG manufacturer that monitors over 7 million SKUs across 1,300 retail websites, achieving 90% product recognition accuracy and 95% automation across key workflows, without adding to their operational headcount. That is what production actually looks like in practice.

The thing I’d want every CXO to internalize is this: governance is not a constraint on speed. It is the infrastructure that makes speed sustainable. Organizations that skip it move fast early and stall later. The ones that build it in from the start are the ones still running at scale two years in.

9: As someone who has built a global AI services and implementation company, what do you think most founders and business leaders still get wrong about building sustainable AI capabilities, as opposed to chasing the next model or tool?

A: The most common mistake is optimizing for breadth when depth is what actually matters. Founders who try to serve every industry with every AI capability end up with a portfolio of custom builds, no repeatability, and no platform leverage. The ones who pick two or three domains, go deep, and build reusable foundations are the ones who scale. There’s no shortcut to that.

The same applies to enterprise leaders evaluating AI vendors. The question isn’t who has the most impressive demo. It’s who has done this specific thing, in your specific context, more than once. Repeatability is the key here. Everything else is a pitch.

10: Three years from now, what separates the enterprises that actually won with AI from those still running pilots? And as a founder, what has building Blackstraw through this wave taught you about the relationship between people and technology?

A: The companies that will be ahead in three years are the ones who did the unglamorous work early: data governance, operator trust, feedback loops, retraining schedules. That foundation will show up in margins, in decision speed, and in capabilities their competitors simply cannot replicate quickly. 

Building Blackstraw has taught me just how important the people behind the technology really are. You can develop a great AI system, but its success still depends on the people building it and the employees who need to trust it enough to use it. That’s something we’ve tried to reflect in how we run Blackstraw, too. 

We’ve been Great Place to Work certified for four consecutive years, and we’ve had a profit-sharing program for more than four years, allowing employees who’ve been with us for at least two years to share in the company’s success. I believe that when people feel valued and invested in what they’re building, it shows in their work.

Technology will continue to evolve, and competitors can adopt many of the same tools. But building a team that works well together and takes real ownership of what it delivers takes time.

Editorial Note: This interview features Atul Arya, Founder & CEO of Blackstraw.ai, in conversation with Vishal Sanghavi, Founder & CEO of JEH Aerospace. The views and opinions expressed are those of the interview participant and do not necessarily reflect the editorial views of TheNextHint.

Image Disclaimer: The featured image is a digitally created editorial composite using reference photographs of both individuals. It does not represent an actual photograph of the interview.

The post Vishal Sanghavi, Founder & CEO, JEH Aerospace, interviewing Atul Arya, Founder & CEO, Blackstraw AI  appeared first on The Next Hint.