The model is the last mile, not the hard part. That is the core argument Manas Talukdar makes on this episode of Unscripted, and it comes from someone who has spent 19 years building the infrastructure underneath AI systems rather than the models themselves. While most of the industry argues about which large language model is best, Talukdar has spent his career on the data pipelines, context engineering, and system architecture that actually determine whether any of those models work in production. His view is blunt: most enterprise AI failures have nothing to do with the model at all.
Manas Talukdar has spent 19 years building the data backbone behind AI systems, starting in classical machine learning and data infrastructure in the process industry, then moving into the platform work behind one of the largest enterprise AI companies, and later into the training data systems behind modern large language models at a frontier AI lab. He recently left to found his own AI company with a co-founder, still in stealth. Across every stage of his career, from rules-based systems to classical ML to today's LLMs, his focus has stayed on the same problem: not what the model can reason about, but whether the data feeding it is structured well enough for that reasoning to be trustworthy.
Talukdar's framing is that AI is fundamentally a data processing problem, not an inference problem. The model provides reasoning ability, but on its own it is powerless without a stack feeding it properly constructed data, running pipelines, and drawing inferences from production systems. He is direct about the stakes of getting this wrong: a chatbot mixing up which company's CEO someone is, is a minor annoyance. The same kind of error occurring in a critical infrastructure environment, the space where Talukdar spent over a decade of his career, can be catastrophic. That distinction is why he sees the unglamorous work around the model, not the model itself, as where enterprise AI succeeds or fails.
According to Talukdar, the single biggest failure point in enterprise AI is long horizon problems, workflows that need to run reliably over extended periods rather than answer a single question. As a workflow gets longer, tool-calling accuracy tends to decrease, and no amount of context window expansion fully solves this. He points out that even as vendors advertise context windows in the millions of tokens, or talk about infinite context, there are mathematical limits to accuracy that persist regardless of window size. This is part of why AI memory has become its own category of startup, and why Talukdar spent time in his last role specifically working on the AI memory problem. He is candid that reliable agent fleets, ones that can run a workflow for two or three hours without accuracy degrading, largely do not exist in production yet. Pilots look strong. Sustained production is a different problem entirely.
Talukdar is clear that large language models are inherently probabilistic, not deterministic, and that no architectural trick changes that fact without an entirely new model architecture. What can be built around that probabilism is resilience: mixture of experts setups, multi-model architectures where one model verifies another's output, and domain-specific fine-tuning that improves accuracy within a narrower scope. He points to world models, an area researchers like Yann LeCun have championed, as one of the more promising research directions toward something closer to determinism, particularly in physical AI, but stresses this is still research-stage work. His expectation, based on how long it typically takes research to reach commercialization, is that the next two to three years will start producing real output from that work, in both physical AI and enterprise AI.
Two forces make this harder than it sounds. The first is economics: feeding an enterprise's full knowledge base, code, tribal knowledge, and call recordings into an LLM is expensive at scale, which is why the industry has shifted from what Talukdar calls token maxing to token optimization, with active academic research into compression techniques still working their way toward commercialization. The second is accuracy: an organization with ten petabytes of data trying to get an LLM to reason accurately across all of it, through retrieval or memory systems, is a genuinely hard problem, and most current AI memory startups are still solving the simpler version of it, retaining context across a single ongoing conversation, rather than the harder enterprise version spanning knowledge graphs, code bases, and multiple systems of record. On top of the technical difficulty sits a philosophical one: intellectual property exposure. Talukdar notes that most companies do not have Nvidia's leverage to negotiate dedicated, private model instances, which raises real questions about what happens to proprietary data once it's fed into a third-party model, a concern he says both Alex Karp of Palantir and Satya Nadella have raised publicly in different ways.
Partially, and Talukdar doesn't dodge this. He agrees that a lot of the mechanical work, building data ingestion pipelines, defining data models, basic cleansing, is already becoming commoditized. What he believes will hold value longer is judgment: deciding what's actually worth collecting out of, say, 40 years of tribal data, and architecting systems that compound in value over years. He's careful not to overstate the durability of even that advantage, noting that given the current pace of change, some of what isn't commoditized today could be within a couple of years.
Talukdar names three specific gaps between what's being promised and what's actually being delivered. The first is the long horizon and long-running workflow problem already discussed. The second is cost: the token pricing much of the current enterprise AI wave runs on is still substantially subsidized by venture capital, a dynamic he compares directly to the early ride-share pricing wars, and he doesn't expect that subsidy to last. The third is the feedback loop: building evaluation harnesses and online retraining into a system from the start, rather than treating them as an afterthought once something breaks in production. He also flags a related risk showing up inside enterprises right now: unstructured AI-assisted coding, or vibe coding, that skips architectural thinking around failure paths. Done well, by someone who understands the tradeoffs, he says an engineer can become dramatically more productive. Done without that understanding, it quietly builds tech debt and introduces bugs someone else eventually has to clean up.
Talukdar's answer has three parts: proprietary data, deeply integrated workflows, and operational discipline, meaning governance and system resilience built in from the start rather than layered on top. The common thread across all three is that AI has to be architected as core to how a company operates, not bolted on as a feature. He describes what that looks like concretely: a customer call recording gets ingested automatically, the system identifies that a new support ticket is needed, creates it, notifies the relevant engineer in Slack, and simultaneously submits a pull request in GitHub for review, all within minutes of the call ending, with no person manually triggering any of those steps. He's built versions of this internally before and says the pattern is well within reach for companies willing to build this way from the ground up, rather than adding AI features onto existing workflows after the fact.
What does Manas Talukdar mean when he says the model is the last mile? He means the model only provides reasoning ability, and that reasoning is worthless without the surrounding infrastructure, data pipelines, context management, and system architecture, that feeds it properly constructed, reliable data.
Why do enterprise AI projects fail on long horizon problems? Because tool-calling accuracy tends to decrease as a workflow gets longer, a mathematical limitation that persists even as context windows expand, making it difficult for AI systems to run complex, multi-hour workflows reliably in production.
Who is Manas Talukdar? An AI infrastructure builder with 19 years of experience spanning the process industry, enterprise AI platforms, and training data systems at a frontier AI lab, now building his own AI company.
Will data engineering get commoditized as models improve? Partially, according to Talukdar. Mechanical work like data ingestion pipelines and basic data modeling is already commoditizing, but the judgment of what data is actually worth collecting, and architectures that compound in value over years, are likely to hold value longer, though even that gap may narrow within a couple of years.
What will provide durable competitive advantage as AI models get commoditized? Proprietary data, deeply integrated workflows, and operational discipline, with AI built as core to how a company operates rather than added on as a feature.
Watch the full conversation and explore more episodes at Unscripted with Jeff Pedowitz.