Jan 31, 2025 · 33m · big-technology
Implementing AI In The Real World — With Kyndryl's Antoine Shagoury
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Kyndryl Chief Technology Officer Antoine Shagoury joins Alex Kantrowitz to break down the technical, architectural, and organizational realities of deploying enterprise AI agents at scale. Moving past industry hype, Shagoury details how enterprises can overcome proof-of-concept failure rates through rigorous data governance, compute workload optimization, and targeted multi-agent workflows.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 22.6% of the talking time here. How this is scored →
Alex as informed peer 4.8 Guest teaching 5.6 Guest disagreement 1.3 Alex pushing back 2.4
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 3:09 Antoine rejects the simple automation rebrand premise
Antoine directly pushes back on the idea that agents are merely a cynical rebrand of traditional automation, arguing that the granularity of orchestration changes the impact radius fundamentally.
Hardest push from Alex ▶ 3:05 Alex challenges the reality of agentic hypeAlex confronts the guest with the skeptical tech perspective, questioning whether the agent wave is genuine technical progress or just recycled automation terminology.
Biggest teaching moment ▶ 21:40 Antoine breaks down mutable data cascades and hallucinationsAntoine systematically explains that model hallucinations often stem from brittle enterprise applications continually modifying source data rather than algorithmic flaws in the models themselves.
Alex holds their own ▶ 17:30 Alex anchors the conversation in Amazon negotiation systemsAlex demonstrates his investigative background by citing findings from his book Always Day One on how Amazon automated vendor negotiations, anticipating the current agentic paradigm.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Automation versus Agents and Telecom Code Deployment | 4 | 5 | 2 | 4 | Alex challenges whether agents are merely a rebranded term for standard automation. Antoine explains how the granularity of orchestration and state management differentiates modern agents, backing it up with a telecom code deployment case study. | |
| Kyndryl's Spinoff Heritage and the Proof-of-Concept Dilemma | 5 | 3 | 1 | 3 | Alex cites industry statistics that 80 to 90 percent of AI proofs of concept fail to make it to production. Antoine welcomes the framing and prepares to break down the divergence between market hype and operational realities. | |
| Analyzing Proof-of-Concept Failures and Setting Realistic Goals | 5 | 6 | 2 | 2 | Antoine educates Alex on why POCs stall, revealing that many business problems require simple programmatic fixes or data cleansing rather than novel generative AI models. Alex synthesizes this by noting enterprise leaders must slow down post-ChatGPT hype. | |
| Real-World Generative AI Successes in Telecom and Automotive | 4 | 6 | 1 | 2 | Alex asks which specific sectors represent the successful minority of AI deployments. Antoine outlines concrete success patterns in telecom customer basket personalization and automotive micro-models targeting supply chain quality. | |
| Multi-Agent Workflows and the Future of Automated Negotiation | 7 | 5 | 1 | 2 | Alex demonstrates substantial subject-matter expertise by citing reporting from his book Always Day One regarding Amazon automating vendor negotiations. Antoine agrees and details how multi-agent architectures (orchestration, opportunity, presentation) collaborate. | |
| Data Lineage, Tokenization, and Legacy Technical Debt | 4 | 7 | 1 | 2 | Antoine gives a masterclass on data lineage, tokenization, and technical debt, explaining that mutable source data and legacy application patches are the true source of model hallucinations. Alex reacts to the staggering scope of corporate data issues. | |
| Overcoming the Scaling Barrier through Workload Optimization | 4 | 6 | 1 | 2 | Antoine details the economics of scaling AI, explaining workload rightsizing across GPU tiers and memory profiles to avoid massive compute bills. He also notes a major shift where C-suite business leaders are driving AI procurement with little patience for failure. | |
| Balancing Commercial SaaS, Open Source, and Micro-Models | 5 | 6 | 2 | 3 | Alex pushes Antoine on whether enterprises should adopt off-the-shelf SaaS or build on open-source models. Antoine rejects a binary choice, explaining how companies blend embedded SaaS AI with targeted micro-models arbitrated across foundational models. | |
| Kyndryl's Engineering Collaboration with NVIDIA and Dell | 5 | 6 | 1 | 2 | Antoine details Kyndryl's deep engineering collaborations with NVIDIA and Dell on private sovereign AI infrastructure and NIM microservices. Alex jokes about the ubiquitous use of NVIDIA's 'accelerate' buzzword while acknowledging the tangible results. |
Statements from this episode (13)
Enterprise AI adoption cannot start directly with AI agents, says Shagoury
“Especially with agents, it's, you don't start there. It doesn't start with an agent. There's a whole different process that kind of gets us into that set of capabilities.”
Kyndryl uses AI agents to automate telco client code testing
“We've been able to use our AI agent framework to deploy not only code assist agents, but also into test and deployment agents. So how we can also shorten the time in which, where we see errors in coding to be returned back and be refined. So we've deployed tha…”
Enterprise AI proof-of-concept volume is increasing exponentially, says Shagoury
“The POCs have actually gone up exponentially. They haven't gone down. They haven't died down at all in that scenario.”
An 80 percent failure rate for enterprise AI proofs-of-concept is common
“And although, yeah, 80% failure rate is not uncommon or call it, you know, getting thrown on the shelf, right? First of a kind is last of a kind type of scenario.”
Most enterprise AI projects actually just require simple automation, says Shagoury
“And we often find many of the POCs turn into, they don't require complex AI, new model development and things in that space. They require simple automation. So they require more or more data, right? Or more programmatic changes in how the application is operat…”
Automakers use AI micro-models to analyze customer interactions and optimize supply chains
“Automotive a little bit differently. Interesting though. Big push on understanding personalization. So harvesting a lot of the surveys, the interactions, the chats, and really directing that through. So they've really started to leverage models and now micro m…”
Amazon pioneered using automated systems to negotiate with its fulfillment vendors
“When I was reporting on Amazon, I found that they were using automation systems to negotiate with the vendors who are supplying their fulfillment centers.”
Mutated source data in fragile legacy applications drives enterprise AI hallucinations
“Data has become so mutable in the environments for many business because the applications are fragile. So what they've done is they layer application services, but they don't touch them. And what they do is they keep modifying source data. And when you start t…”
Maintaining continual real-time AI model operations in production is very expensive
“So getting deep in analysis, deep in processing, continual real-time model operations become very expensive.”
Business leaders, not IT, now direct the majority of enterprise AI investment
“We're seeing large, the majority of investment being directed through business leaders and how we're driving into it.”
Enterprises have lost their appetite for building homegrown AI solutions
“There's a lot less appetite to start building homegrown solutions anymore.”
Enterprises will not replace the AI tools native to their SaaS platforms
“There's the embedded tooling. So like the things like the Salesforce conversations or the ServiceNow conversations, there's a lot of AI that's embedded within the workflows and the information being gathered. And a lot of that, you're not necessarily going to …”
Enterprise AI ROI requires targeted micro-models arbitrating between large foundational models
“Can you now devolve or basically create a micro model that's very targeted to your business? That basically was a recipe. So how do I look at what's happening within, you know, within a CHATGP versus Gemini? How do I start looking at Lama differently? And that…”