Your AI model is the easy part. The rest is the job.

I sat through another AI demo a few weeks ago. Clean dashboard, a churn model with genuinely good numbers, a founder using the word “transformation” a lot. Then someone on the client side asked the only question that ever really matters: “Where does it pull the customer data from?”

The room went quiet. The data lived in a system from 2003 that exports a flat file once a night, and nobody still working there fully understood the export format. The person who wrote it had retired years ago.

That single question is why most “AI projects” never make it past the slideshow. And the latest data backs it up: a global study by Harvard Business Review Analytic Services with Cloudera, published in March 2026, found that only 7% of enterprises say their data is completely ready for AI, and 73% admit they struggle with AI data preparation. Gartner had been warning that 60% of AI projects lacking AI-ready data would be abandoned through 2026, and that prediction is materializing right now.

I’ve been building software for a long time, and I now run engineering at a company that spends most of its days wiring modern things into systems that were already old when I started. So here’s something that might land badly at a school full of people learning to build models: the model is the easy part. Almost nobody tells you that on the way in.

The part the courses get right

Let me be fair. Learning the math, the architectures, how a gradient actually moves through a network, that’s worth doing, and good programs teach it properly. A well-designed Data Science course helps learners build these foundations while also understanding how data flows through real business systems. If you can take messy data and get a model to predict something useful, you have a real skill.

But the model almost never lives alone. In a classroom it gets a clean CSV with a tidy target column. In a company it gets a 40-gigabyte table with three fields nobody can explain, a status column that means five different things depending on which decade the row was written in, and a hard rule that it cannot touch the production database between 8am and 8pm.

That distance between “works in the notebook” and “works in the company” is the whole game. And the distance is not made of AI. It’s made of plumbing.

The part nobody puts on the slide

Most companies older than ten years don’t run on a clean, modern stack. They run on layers. A core system from one era. A CRM bolted on in another. A reporting tool somebody’s cousin set up in 2012. And holding it all together, a pile of spreadsheets that has no business being load-bearing but absolutely is.

So when a board says “we need AI,” what they’re really asking is for a model built in 2026 to sit down and have a civil chat with a system from 1998. Nobody demos that conversation. It makes for a lousy photo op. But that’s where the money is, and where most projects quietly fall apart. Deloitte’s State of AI in the Enterprise 2026 puts numbers on it: technical infrastructure readiness sits at 43%, meaning companies have bought the tools, but data governance readiness is at 30% and talent readiness at 20%. The new brain is in the building. What’s missing is the plumbing that connects it to everything else.

A story I’ve seen more than once

A bank wants fraud detection. Reasonable goal. The data science is well understood, the team is sharp, the model performs in testing. Then reality walks in.

The transaction data sits in a mainframe that runs a batch job at 2am. The records are fixed-width, written in a style most people under forty have never touched. There’s no API. The documentation describes the system as it was in 2009, which is not the system you’re looking at now. There is, however, one very calm person near retirement who knows how all of it works, and everyone treats them like a national monument because that’s roughly what they are.

The model was finished in a few weeks. Getting clean, trustworthy, recent data into it took months. The hard part was never the intelligence. It was archaeology. A lot of what my teams at Luby do under the banner of “AI projects” is exactly this kind of careful surgery on systems that refuse to die.

It’s also worth noting what MIT’s research had already shown about who succeeds a finding that keeps being confirmed in 2026: companies that buy from specialized vendors and build partnerships see roughly 67% success rates, while purely internal builds succeed only a third as often. The failure mode isn’t lack of brainpower. It’s underestimating the plumbing.

What the boring work actually looks like

You spend a week just agreeing on what a “customer” is, because sales, finance, and support defined it differently and all three are convinced they’re right. You build a pipeline that has to survive a source system going down at random, sending duplicates, and occasionally inventing orders placed on the 30th of February. You decide what the system does when the model is not confident, because a wrong answer delivered with total certainty is worse than no answer at all. You figure out who gets a phone call at 3am when something drifts.

None of that is on a course syllabus. But it is the actual work. The model is maybe 20% of a real project. The rest is making sure the smart thing can be trusted, fed, watched, and explained to a human who has to answer for it. RAND researchers who interviewed 65 experienced AI practitioners found the same thing from another angle: the most common causes of failure aren’t technical at all, they’re misaligned incentives and tools built without the people who’d actually use them.

Why this should matter to you, specifically

The market is filling up fast with people who can train a model. “I can build a classifier” is quickly becoming the floor, not the ceiling. Sorry.

The real edge is being the person who can stand with one foot in each world. Someone who understands what a transformer is doing and why the company’s order data grows a phantom duplicate every February. Someone who can look at a thirty-year-old database schema without flinching and work out how to get its data somewhere a model can use it safely.

That person is rare. That person gets paid. And barely anyone is deliberately training to become them, because it isn’t glamorous and it doesn’t make a good portfolio screenshot.

So here’s a slightly strange piece of advice: alongside the new and shiny, spend a weekend with the old and ugly. Learn what an ETL pipeline has to survive in the wild. Find a genuinely awful dataset, inconsistent dates, a free-text field where someone typed their feelings into a column meant for numbers — and make it usable. That exercise will teach you more about applied AI than another tutorial on a model you’ll never deploy. The blog at the Boston Institute of Analyticsis a good place to follow both ends of this: the modeling and the messy engineering reality wrapped around it.

One last thing

The most valuable people in applied AI over the next few years won’t be the ones with the fanciest model. They’ll be the translators, the ones who can make something genuinely intelligent shake hands with something genuinely old, without breaking either side of the handshake.

A model that can’t reach trustworthy data is just a very expensive demo. A modest model wired correctly into the right systems, fed clean inputs, watched closely, and actually trusted by the people who use it, that’s the thing that quietly changes how a business runs.

The model is the easy part. Learning to land it inside the strange, stubborn, half-broken systems that real companies run on, that’s the real job. And honestly, it’s a good one.

Rodrigo Gardin runs technology at Luby, a software engineering firm that’s been around for over twenty years, working with companies across Brazil and the US to untangle and modernize the systems they depend on. More about what the team does at luby.com.br.

Submission details

• Meta Title: Your AI model is the easy part. The rest is the Job.
• Meta Description: Only 7% of enterprises have AI-ready data (HBR/Cloudera, 2026). A CTO’s honest take on why the model is easy and why integration with legacy systems is where projects live or die.
• Word count: ~1,375 words
• External links: Boston Institute of Analytics blog (contextual, “Why this should matter” section); Cloudera/HBR press release and Gartner can be linked at first mention if the publisher allows additional outbound links.
• Backlinks: luby.co (body, contextual) and luby.com.br (author bio).

Sources cited in the article

1. Cloudera × Harvard Business Review Analytic Services – Taming the Complexity of AI Data Readiness (March 2026): only 7% of enterprises say their data is completely ready for AI; 27% say it’s not very or not at all ready; 73% struggle with AI data preparation.
2. Deloitte – The State of AI in the Enterprise 2026: The Untapped Edge (survey of 3,235 leaders across 24 countries): technical infrastructure readiness at 43%, data governance at 30%, data management at 40%, talent at 20%. (Readiness breakdown reported via secondary coverage verify against the original report before submission.)
3. Gartner – Lack of AI-Ready Data Puts AI Projects at Risk (press release): prediction that 60% of AI projects lacking AI-ready data would be abandoned through 2026.
4. MIT NANDA – The GenAI Divide: State of AI in Business 2025 (Fortune coverage): vendor partnerships succeed ~67% vs ~1/3 for internal builds.
5. RAND Corporation – The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed(interviews with 65 AI practitioners): top failure causes are organizational, not technical.

Data Science Course in Mumbai | Data Science Course in Bengaluru | Data Science Course in Hyderabad | Data Science Course in Delhi | Data Science Course in Pune | Data Science Course in Kolkata | Data Science Course in Thane | Data Science Course in Chennai