By: Marcy Paulson
Scaling AI across a large enterprise sounds simple in a boardroom. So why does MIT report that 95% of generative AI pilots fail to deliver measurable business impact? And why does Gartner estimate that organizations will cancel more than 40% of agentic AI projects before the end of 2027?
Atul Arya, Founder and CEO of Blackstraw, says the gap between AI experimentation and AI at scale is an execution gap, and closing it requires engineering-first delivery and a lot of disciplined data work. That’s how he and his team earn trust as they help enterprises build and optimize AI solutions for measurable business impact.
Why AI Implementation At Scale Across A Large Enterprise Looks Nothing Like The Strategy Deck Version
In the strategy deck, AI rollout looks like a neat sequence. Everyone nods along with the linear steps in the boardroom: assess, design, build, deploy, and scale. Then, they walk into the workplace and discover that one business is a muddle of conflicting priorities and definitions. Even worse, it’s a technology landscape that reflects every merger and workaround from the last 15 years.
The deck assumes a company has clean data and a level of alignment across stakeholders that most large organizations only achieve right before a budget cycle closes. But when the model that looked great in the demo hits missing values, stale records, inconsistent naming conventions, and non-standard processes, it falls apart.
“The strategy usually isn’t wrong exactly,” Arya observes. “It’s optimized for a company that doesn’t exist, and that’s a problem when you’re trying to build one that does.”
What Atul Arya Learned About Scaling Enterprise AI Across Over 200 AI Implementations
Arya has seen this pattern play out repeatedly. Before founding Blackstraw in 2018, he led AI and machine learning teams at Nielsen and rose to Global VP of Innovation. With over two decades spanning technology, data, and AI, he launched Blackstraw to close the gap between AI experimentation and production so enterprises could move from pilots to measurable business impact through an engineering-first approach.
Across more than 200 enterprise AI implementations, a few lessons show up consistently. Arya first warns that “Off-the-shelf never fits as-is, no matter how good the vendor demo looks. Every implementation needs real customization for that company’s data and workflows.”
He also argues that the model is rarely the bottleneck. In his years of experience, data readiness and internal buy-in are what cause the logjams. Organizations can build impressive prototypes, but if upstream pipelines are fragile or business teams don’t trust what the model is saying, the initiative stalls.
One of Arya’s most valuable lessons has to do with focus. He notes that the programs that actually stick are the ones that prove measurable value fast in one narrow area instead of trying to transform the entire company in one sweep.
As an example, Atul points to a client that saw a 15% lift in sales lead conversion in the very first quarter simply by applying predictive analytics to one focused part of the business. It was an early win that did what no workshop could: created belief and earned the right to scale.
And last but not least, Arya learned that while big-bang AI programs make for great press releases, they’re the hardest to deliver. The most difficult part of the rollout is adoption at scale, under real constraints and accountability.
Blackstraw’s Engineering-First Execution Versus Strategy-Only Consulting Engagements For Scaling Enterprise AI
“A strategy document doesn’t survive contact with legacy infrastructure,” Arya says. “And by the time that document fails, the consultants who wrote it have moved on to the next client. I’ve read plenty of frameworks that read beautifully on slide three, assuming data would magically be clean, systems would talk to each other, and teams would fall in line. Then the consultants leave, and six months later it’s a binder nobody opens. Engineering-first means we’re still in the room when those assumptions break.”
Blackstraw’s engineering-first model exists because the real work begins when assumptions break. Engineering-first means staying in the room when systems don’t talk to each other and when a business unit pushes back on process change. Instead of measuring success against projections, the work is measured against what’s actually running in production, and those lessons are looped back into the architecture as the implementation evolves.
One approach ends with the recommendation. The other ends when it’s actually working, but stays involved to keep it that way. In one recent engagement, Blackstraw helped a client move over 800 data pipelines off a legacy platform without a single day of disrupted reporting.
Ultimately, scaling AI across a large enterprise is less like launching a product and more like upgrading the nervous system of a living organization. It requires patience and a willingness to do the careful work to build a strong foundation of pipelines, definitions, workflows, governance, and reliability. That’s not just a strategy deck; it’s engineering and the discipline to turn early wins into an enterprise capability that lasts.
