How do you move from an AI demo to real world impact?

Sidarth: The first step is to stop thinking of AI as a model and start thinking of it as a product. Begin with a clearly defined business problem. Only after that is precise should you identify the AI tasks needed to solve it. Do not start with the model. Start with the problem.

Scalability rests on data readiness. You need data that is consistent, complete, and reliable. With that foundation, design a modular and expandable architecture so the system can evolve as requirements change. Then tie model metrics to business KPIs. Define success in business terms and ensure model evaluation directly maps to those outcomes.

You also need a plan for continuous improvement and a clear adoption strategy. A high performing model does not guarantee adoption. Integration into real workflows matters, both internal and customer facing. Build feedback loops so users can influence the roadmap and so the system gets better in production. In short, think product lifecycle, not just model training.

How did you scale your AI powered product matching system at Staples?

Sidarth: We built an enterprise grade product matching application that has been in production for more than two years and supports quoting tools and multiple search systems across our organization. Its success comes from disciplined data design, automation, and user centered iteration.

From the start, we invested in reliable data and automated pipelines to keep data and outputs fresh. The matching flow begins with query understanding. Customer queries can be informal or non standard, so we normalize and extract meaning first. With a standardized representation, we use a hybrid search approach that combines a keyword method with a dense retrieval embedding model. We fine tuned the embedding model on high quality human labeled ground truth, which lifted relevance significantly. The hybrid design gives strong recall, but to reduce false positives we add a final validation step using a large language model to check alignment between the query intent and the candidate matches. That helps ensure we return a small set of high quality results.

Equally important, we followed a user centered process. We launched an initial version to a pilot group, ran AB tests, and iterated on features based on real usage. We tested from the user interface forward to ensure the system met operational constraints and integrated smoothly with business workflows. That end to end approach yielded a solution that is sustainable and widely adopted.

What hurdles did you face when scaling AI, and how did you address them?

Sidarth: Ambiguity is common in product matching. Queries can be short or vague. Query understanding helps by expanding tokens and terms into a richer vocabulary that aligns with our models and catalog. Another challenge is domain specific language. Off the shelf models often miss nuanced product terminology. Fine tuning on domain data makes a big difference.

There is also a constant tradeoff between latency and accuracy. More advanced models can be slower. Our answer is a multi stage retrieval and ranking pipeline. Do the heavy lifting with fast methods to reduce the candidate pool, then apply a more nuanced ranker to a small set of top candidates rather than millions of SKUs. This structure delivers both speed and quality.

Finally, there are organizational hurdles. Systems must meet service level agreements, integrate with existing tools, and be straightforward to adopt. We worked closely with stakeholders, validated from the UI down to the services, and kept a tight feedback loop to align technical choices with operational realities.

There is also a constant tradeoff between latency and accuracy. More advanced models can be slower. Our answer is a multi stage retrieval and ranking pipeline. Do the heavy lifting with fast methods to reduce the candidate pool, then apply a more nuanced ranker to a small set of top candidates rather than millions of SKUs. This structure delivers both speed and quality.

Finally, there are organizational hurdles. Systems must meet service level agreements, integrate with existing tools, and be straightforward to adopt. We worked closely with stakeholders, validated from the UI down to the services, and kept a tight feedback loop to align technical choices with operational realities.

How do scalable AI systems transform customer experiences?

Sidarth: Scalable AI brings consistency, speed, and quality. In a cross referencing or product matching scenario, customers see more precise matches earlier in the results. That saves time and supports faster decisions, which shows up in better conversion and satisfaction.

To get there, businesses should invest in reliable data pipelines and consistent, high quality data across systems. They should define clear business KPIs and measure them over time to understand the contribution of AI. Beyond that, set up evaluation pipelines and dashboards for continuous tracking. Capture user feedback signals and leverage them for model improvement.

Model drift is a reality as queries and domains evolve. Monitor for drift, set alarms, and retrain incrementally to stay aligned with the business context. Pilot smart, but design for production from day one. Even an early pilot should consider model versioning, stable data pipelines, and integration paths. If you think product from the beginning, you avoid a painful rewrite later.

What is your advice for leaders who want to scale AI projects?

Sidarth: Balance innovation with operational efficiency. You can build an excellent model that fails in practice if it does not meet latency or reliability requirements. Choose approaches that satisfy constraints while moving the frontier on quality. Often the right answer is a layered system that applies sophistication exactly where it yields the most return.

Change management is essential. Plan adoption in stages, from pilot to launch, and schedule ongoing feedback sessions so you can track how the product is used and where it needs tuning. Adoption is a process, not a one time event. When users feel heard and see improvements, they trust the system and rely on it, which is where real business impact comes from.

Also, keep your focus on measurable outcomes. Tie every improvement to a KPI, validate with experiments, and share results with stakeholders. That transparency builds credibility and secures the sponsorship needed for continued investment.

Conclusion

Sustained AI impact comes from product thinking, not model worship. Define the business problem, prepare reliable data, design modular systems, and connect model metrics to enterprise KPIs. Pair fast retrieval with targeted sophistication, monitor for drift, and build with adoption in mind from day one. With disciplined engineering and thoughtful change management, organizations can move past impressive demos and deliver AI products that scale, endure, and create lasting value.

Sidarth Thapar is a Manager of Applied AI at Staples, where he drives enterprise scale AI initiatives spanning semantic search, retrieval systems, product intelligence, recommendations, dynamic pricing, and generative AI.