Tech Explained: Why 90% of AI models never make it to production: Blackstraw.ai.’s Atul Arya in Simple Terms

The true hurdle for businesses transitioning from AI experiments to full-scale deployment is not model creation, but rather their operationalisation across the enterprise. Atul Arya, founder and CEO of Blackstraw.ai, discusses with The Economic Times Digital the essential elements for deploying production-ready AI, emphasizing the importance of governance, data engineering, ongoing monitoring, and user adoption alongside algorithms. He also addresses why many companies find it difficult to transition from AI experimentation to achieving tangible business results. Edited excerpts:

Economic Times (ET): Blackstraw has experience in taking AI from experimentation to production at enterprise scale. What does truly production-grade AI mean to you, and why do most enterprises still struggle to get there?

Atul Arya (AA): Production-grade AI, to me, means a system that works in the real world and not just in a controlled environment with clean, curated data. It means models that are monitored, maintained, and built to handle the messiness of live enterprise data at scale. Most enterprises struggle because they treat deployment as a finish line. Which it is not. It is the starting point. We consistently see common failure points: data that is not production-ready, models that drift without anyone noticing, and teams that never fully adopt the system. At Blackstraw, we engineer for production from day one, with optimized pipelines, continuous monitoring, automated retraining, and governance built in. One of our clients saw $16 million in savings, with systems supporting over 16,000 users. That is what production-grade actually delivers.
ET: Enterprises often struggle to move from AI pilots to production. From your experience, what are the three most common failure points, and how does Blackstraw’s platform-led model address them?
AA: Over the years, we have seen three failure points consistently: data readiness, deployment discipline, and change management. Various studies have shown numbers claiming up to 90% of ML models failing to make it to production. These recurring issues happen because, firstly, most pilots are built on clean, curated datasets. Production means messy, real-world data at scale that is often fragmented across systems. If the data foundation isn’t there, the model will not survive contact with reality.

Secondly, enterprises treat deployment as a one-time event. Over time, models drift and data patterns change. Without continuous monitoring and retraining infrastructure, performance degrades fast and most organizations do not catch it until the damage is done.

Third, adoption. You can build a technically sound system, but if it does not fit into existing workflows or if teams resist using it, it dies on the vine. Making change management is absolutely crucial, it is not optional and it is where most pilots actually fail.
Our approach is built around these realities. We do not just build models, instead, we build the full operational stack: standardized pipelines, version control, monitoring systems, automated retraining workflows.ET: The term “AI-native organization” is used widely today. In practical terms, what distinguishes an AI-native enterprise from one that is merely deploying AI tools?
AA: The term “AI-native,” for me, in practical terms, comes down to this: is AI making decisions in the business, or just informing them? Today, most enterprises are deploying assistive AI tooling, the likes of dashboards, copilots, analytics that support existing workflows and surface current trends. That is useful, but it is not transformative.

An AI-native organization would rebuild how decisions actually get made. They would have models that trigger action, not just suggest. Something that enables exceptions to route automatically, and decisions to happen in minutes because the system is designed to act on AI outputs directly, not take days because someone has to review a recommendation first.

For a business to get there, it requires three shifts. First, decision architecture, defining what gets automated, what needs human oversight, and who’s accountable when AI drives outcomes. Second, operational integration, where AI is not a separate system; it is embedded in CRM, ERP, risk platforms, with guardrails and feedback loops built in. Third, continuous learning as a discipline, capturing outcomes, measuring business impact, and retraining models based on real-world performance, not just lab accuracy.

ET: Blackstraw works deeply across data engineering, ML, and automation. Which of these layers is proving to be the biggest bottleneck for enterprises today, and why?
AA: At Blackstraw, we have identified that data engineering remains a challenge. The data is always there, but few companies understand what it takes to build AI-ready datasets and the infrastructure to support them at scale. The journey from pilot data to production data is where most AI projects stall. Models work beautifully in controlled environments with curated datasets. But production means ingesting from hybrid sources, handling unstructured formats, maintaining availability across the enterprise. That gap is consistently underestimated.

The challenge shows up everywhere. Data collection and diversity, an AI-ready dataset needs to be balanced and representative, with images covering varied lighting and backgrounds, transactional data capturing edge cases. Most enterprises do not have systematic processes for this. They build models on whatever data happens to exist.

Infrastructure gaps are another constant. Our assessments consistently reveal that companies lack the foundational pieces: data governance policies, centralized lakes, ingestion pipelines. Without these, you cannot turn raw data into something models can actually use at scale.

Then there’s integration complexity. Data is available in silos across systems that were not designed to talk to each other. Building pipelines that clean, process, and harmonize this data while maintaining security and compliance requires deep engineering expertise that most teams simply do not have.

In our work across data deployments, the lesson is clear: data engineering is the foundation that determines whether everything else works. If that’s wrong, even the most sophisticated models won’t deliver business value. It simply won’t work.

ET: As AI adoption scales, governance, security, and reliability become critical. How are enterprise priorities shifting from experimentation to production-grade intelligence?
AA: As AI adoption scales, enterprise priorities are fundamentally shifting. In pilots, risk is theoretical. In production, you are accountable to regulators, customers, and the business. The EU AI Act, passed in 2024 with fines up to 35 million euros. Governance isn’t optional; it’s a compliance imperative. Enterprises now understand that the same systems promising competitive advantage can introduce real risks, such as data breaches, copyright violations, and misinformation, if left uncontrolled.

Yet execution lags intent. While 97% of organizations recognize they need AI governance frameworks, only 19% have implemented one. That’s where production AI stalls. Moving to production-grade intelligence requires fundamental changes in how teams build and deploy. Governance has to be embedded from the start – transparency, explainability, accountability built into pipelines and workflows, not retrofitted after the fact. Risk management must be proactive, with automated systems that discover, classify, and secure data, identifying violations before they become incidents. And security has to operate at scale – data loss prevention, access controls, audit trails, continuous monitoring as models interact with live data.

ET: Your partnerships with Microsoft Azure and Databricks are central to Blackstraw’s strategy. How do these collaborations strengthen your value proposition, and what kinds of enterprise problems are you jointly solving?
AA: Our longstanding partnerships with Microsoft Azure and Databricks is crucial to provide production-grade AI at a large scale. These are now more than just vendor relationships.

While Azure gives us enterprise-grade security, hybrid cloud flexibility, and global infrastructure, Databricks provides the lakehouse architecture that helps us unify data warehouses and lakes, critical when enterprises have data scattered across dozens of systems. Together, they enable us to create complete workflows, from data integration, model deployment, and ongoing monitoring on platforms that already meet regulatory standards and can scale securely.

The problems we are solving are the ones standalone tools cannot handle. Breaking down data silos, migrating fragmented data into unified lakehouses with governance, lineage tracking, and secure sharing built in. Building MLOps platforms that automate machine learning pipelines from data prep to deployment, so clients can operationalize AI at scale instead of managing one-off experiments. Implement LLM solutions with Azure OpenAI Services by fine-tuning them on business data or hosting open-source models like Llama2 on Azure GPU instances for secure and cost-effective deployments.

ET: How large is Blackstraw’s current revenue base, and what has growth looked like over the last 2-3 years?
AA: Since our founding in 2018, we’ve been fortunate to grow consistently. We started in Tampa, expanded to delivery centers in Halifax, Toronto and Chennai & Mumbai, and worked with Fortune 500 enterprises across North America, Europe, and global markets. The growth has been organic and client-driven, with long-term partnerships, not short-term projects.

At present, we are investing heavily in GenAI and Agentic AI capabilities, and our focus remains on sustainable expansion aligned with delivering real-world impact. We have been growing at a 20-30 CAGR for the last couple of years and are on track to a $100mn revenue by 2028 which also coincides with our 10-year anniversary.

ET: Which industries are currently driving the bulk of your revenue, such as BFSI, manufacturing, retail, or technology, and why are these sectors adopting AI faster?
AA: CPG, retail and market research have been our traditional powerhouse sectors. But over the last couple of years, we have seen significant revenue coming from financial services, utilities and logistics. These are industries where AI adoption is accelerating because the cost of inaction has become too high. In financial services, its risk management, fraud detection, and personalized customer experiences. Areas where real-time decisions directly impact the bottom line. Retail and manufacturing clients focus on supply chain optimization, demand forecasting, operational efficiency. Even marginal improvements translate to significant competitive advantages. Healthcare organizations are using AI for clinical decision support and operational workflows, driven by regulatory pressures and the need to improve patient outcomes.

ET: From a geographic perspective, where does most of your business come from today? How do demand patterns differ between India, North America, and other global markets?
AA: From a geographic perspective, the majority of our business today comes from Fortune 500 brands and large enterprises based in North America (US & Canada). These organizations are further along in operationalizing AI and are focused not just on experimentation, but on scaling AI responsibly across core business functions. We also work with established enterprises in Europe and Asia, and we’re seeing increasing traction in these markets as AI adoption matures globally.

Demand patterns differ meaningfully by region. In North America, conversations are centered on execution at scale, focusing across governance, accountability, measurable ROI, and embedding AI into decision workflows. In Europe, regulatory alignment and responsible AI frameworks are front and center. In India and other emerging markets, there is strong momentum around digital transformation, cost optimization, and leapfrogging legacy systems with AI-first architectures.

ET: hat is your go-to-market strategy: are you selling primarily through enterprise contracts, platform subscriptions, or long-term transformation partnerships?
AA: We believe in long-term transformation partnerships, while remaining open to one-off consulting engagements or transactional projects when they make sense. We work primarily through enterprise contracts with Fortune 500 clients who need to move beyond pilots to production-grade AI that delivers measurable business outcomes. These are multi-year relationships where we become an extension of the client’s team, deeply embedded in their data architecture, AI strategy, and operational workflows.

We focus on solving complex, high-impact problems that require data engineering, AI/ML expertise, and domain knowledge. The problems that cannot be solved with off-the-shelf platforms or generic consulting frameworks. Our clients come to us because they need a partner who can navigate the gap between ambitious AI goals and practical execution at scale.

We do not sell platform subscriptions or standardized services. We deliver custom solutions built around each enterprise’s data landscape, security requirements, and business priorities. This approach lets us build deep, strategic relationships that evolve as the client’s AI maturity grows. We are aligned with their long-term success, beyond project milestones.

ET: With GenAI accelerating AI adoption, how do you see enterprise requirements evolving over the next 12–18 months, from models to data infrastructure to talent?
AA: From what I am seeing, I reckon over the next 12 to 18 months, enterprise requirements will shift from model experimentation to production infrastructure and governance. The focus is already moving beyond which foundation models to use and more toward building the systems that actually matter: clean data pipelines, robust security layers, and integration with existing enterprise workflows. Enterprises are realizing that sustainable AI requires more than technology. It demands cross-functional teams who can bridge business strategy with technical execution, not just more data scientists. Agentic AI is accelerating this shift, forcing organizations to rethink decision-making processes and human-AI collaboration models.

We’re seeing clients prioritize scalable, secure solutions that deliver measurable business outcomes rather than impressive demos. The question isn’t “Can we build this?” anymore. It’s “Can we run this at scale with the right governance, cost controls, and adoption in place?”