4 months ago
39
As 2024 comes to a close, many enterprise and IT companies are moving their AI PoCs to production. It took them an average of two years to find use cases and offer products to enterprises. Given this delay, building all these solutions in-house now seems questionable, raising concerns among experts.
With the debate around the value of AI PoCs intensifying, industry leaders offer contrasting views on whether they represent a prudent investment or a waste of resources. Critics argue that PoCs rarely deliver scalable value, while proponents emphasise their role in validating innovative AI solutions.
Vin Vashishta, AI advisor and author of From Data to Profit, is among the staunch critics of PoCs. He questioned their purpose, stating: “What’s the point of an AI PoC other than to make consulting companies rich? PoCs don’t deliver revenue and can’t be scaled in production,” he said, adding that while the value of such PoCs is 0, he still hears people trying to justify them.
Credits: Vin VashishtaIf Not PoCs, Then What?
According to Vashishta, businesses should focus on simpler initiatives that build capabilities and deliver quantifiable results, rather than sinking money into PoCs that often lack direction or measurable outcomes.
Instead of PoCs, Vashishta advocates for alternative approaches such as leveraging vendor demos, trialling AI tools, and conducting educational sessions like seminars to introduce AI capabilities to businesses. He argued, “Every PoC objective can be achieved faster and for less with common-sense solutions. Don’t buy into the PoC money pit.”
This sentiment has largely been prevalent in the Indian IT ecosystem. Most of them started out building AI PoCs and products, trying to mimic the success of big-tech companies like Google and Microsoft, or startups like OpenAI and Anthropic. Sooner or later, these IT giants realised that it was easier to build these products using their technologies, as it was easier to transition from PoCs to products quickly.
In their latest earnings calls, Infosys, TCS, Wipro, HCLTech, and Tech Mahindra announced that their AI PoCs were moving into production. Most of them built them using Meta’s Llama and NVIDIA NIM frameworks and have also partnered with AWS and other cloud providers.
As for the enterprises, they are still trying to build such PoCs and move them to production, which according to Vashishta is a waste of time and resources.
The Case for PoCs
Not everyone agrees with the anti-PoC stance. Stefan Ojanen, an AI product leader and MLOps expert, defends PoCs as a critical step in deploying great AI models. “When working with bleeding-edge AI/ML tech, it’s fundamentally impossible to predict a solution’s efficacy without empirical validation in your context,” he said.
Meanwhile, Vijay Raaghavan, the head of enterprise innovation at Fractal, told AIM that this transition from PoCs to real-world applications has presented new challenges, particularly when it comes to measuring value, which is still the toughest part of driving investment in generative AI. He outlined a multi-layered approach: “Generative AI is not a plug-and-play solution. It requires the right data, hyperscale strategy, and long-term commitment.”
According to some of these voices, PoCs test AI solutions within the business’s specific environment, uncovering edge cases and architectural challenges. By identifying the limitations early on, they reduce the risk of failures during full-scale implementation. PoCs explore high-risk, high-reward opportunities, offering insights that boilerplate solutions from others might not provide.
“Dismissing PoCs as worthless is intellectually dishonest,” Ojanen asserted. “PoC is the first iteration. It’s a way to discover asymmetric advantages.”
A Middle Ground?
The tension between critics and supporters often centres around how PoCs are positioned within the broader AI strategy. Critics believe that if a PoC successfully demonstrates value, it ceases to be a PoC and becomes a product. “Models running in production environments prove value. PoCs are like college projects for business,” one argued.
On the other hand, proponents like Fabian Leon Ortega, co-founder and CTO at SunDevs, view PoCs as stepping stones to innovation. He pointed out that PoCs in customer support for telco companies have successfully evolved into production systems. “Thanks to the PoC, stakeholders could see the real value and potential of AI, applying the technology to their own business cases,” Ortega explained.
Vashishta’s criticism extends to the financial implications of PoCs. He contends that companies often treat PoCs as a substitute for robust architecture and design processes, leading to inefficiencies. “If you find a high-reward space, shouldn’t you explore it with a product customers will pay for rather than a PoC that doesn’t provide any new information?” he asked.
Advocates for PoCs say that they can be cost-effective when executed strategically. “I’d rather whip up an 8B model locally with sample data for a few thousand dollars than commit to a million-dollar project without validating assumptions first,” Ojanen remarked.
While PoCs can provide valuable insights and reduce uncertainty, their success hinges on clear objectives and alignment with business goals. Without these, they risk becoming a “money pit,” as Vashishta warns.
(Total 13 views)
