They're failing because nobody wants to admit the process isn't ready.

Sobhan Khani is President and Partner at Plug and Play — the largest corporate innovation platform in the world. 550 corporate partners. 100,000+ startups in their database. 800 people across 70 cities. He has watched IoT, FinTech, crypto, and now AI cycle through the enterprise conversation from both sides of the table.

In this episode, Sobhan and Nick go into the real reasons AI pilots fail to reach production — and what the programs that actually ship do differently.

What you'll hear:

↳ Why McKinsey's stat — 60–70% of enterprises running AI agents, less than 5% seeing results — is a process problem, not a model problem ↳ The top-down vs. bottom-up debate: why both camps are getting results, and what that tells you about AI transformation complexity

↳ The undocumented workaround problem — every enterprise process has a real version and a documented version, and AI finds the gap on day one

↳ Why trying to fit AI into the workflow you already have typically doesn't work — and what Microsoft's Chief Scientist office said about it

↳ The internal champion pattern: what separates the deployments that survive the middle section from the pilots that quietly die

↳ The budget shift thesis: what happens to headcount and software spend when agents can scale without FTE cost

↳ The CTO rule from a seed-stage startup that every mid-market company should borrow before their next hire

This is not a conversation about AI potential. It's a conversation about why the gap between the demo and the deployment is still killing programs at Fortune 500 companies — and what the people actually navigating it are doing differently.

Justin La Julière has been the CIO and CISO at Delta Dental in New Jersey and Connecticut for over 25 years — running IT and security simultaneously in one of the most regulated environments you can operate in: dental insurance, PHI everywhere, state and federal compliance on every AI decision.

Chapters

• 00:00 Introduction and Transformation Journey
• 05:59 Rethinking Processes and Creating New Value
• 12:03 Compliance and Governance in AI Initiatives
• 18:01 Measuring AI Success and ROI
• 23:52 Navigating Compliance and Cultural Change
  

And he's one of the furthest along.

In this episode, Justin shares the exact architecture and cultural playbook he used to drive AI adoption inside a regulated insurance company — without waiting for perfect governance conditions, without blocking tools employees were already using anyway, and without the kind of top-down IT mandate that kills adoption before it starts.

What we get into:

We talk about how Justin's team built a PII/PHI filtering layer on the backend of their internal AI environment — so employees could put anything they wanted in the tool without having to classify their own data first. The cognitive load of "wait, is this PHI?" was killing adoption. Removing that question changed everything.

We talk about AI hours — a cadence Justin runs every three weeks where practitioners from across the business (not IT people) demo what they're actually building with AI. Two-thirds of the company shows up. Not because they were told to. Because someone from provider relations showed them something useful and they wanted in.

We talk about governance with real teeth: a framework that forces a two-page business case before any AI initiative gets resourced, and a discipline of killing projects early — before they become someone's baby — so they don't survive on organizational inertia long after they've stopped delivering value.

We talk about what Justin calls the "green glass risk" — the danger that your entire team only knows what's inside the building, and never brings back signal from outside. And why staying at the tip of the spear on new technology has been his personal operating principle for a quarter century.

Justin's read on where regulated industries are right now: the compliance and legal teams that are sitting on the sidelines aren't protecting the organization — they're falling behind it. The gap between companies that have started and companies that haven't is widening faster than most executives realize.

His single piece of advice for a mid-market healthcare or insurance company just getting started: don't try to sprint before you walk. Get safe tools in people's hands. Build cross-pollination. Let practitioners talk to practitioners. The technology will keep improving — culture is the only thing you actually have to build yourself.

Justin Lahullier is on LinkedIn and responds to messages. If you're a CIO, CISO, or technology leader in a regulated industry navigating any of this, he's worth reaching out to.

In Production is the podcast for technology and AI leaders who have moved past the theory — and are doing the hard work of deploying AI in environments where failure has real consequences. New episodes drop regularly. Subscribe wherever you listen.

They're not. They're making a data accountability decision — and most of them aren't ready for it.

Lance Harlan is the Data Governance Program Manager at Trustmark Bank and the author of the KISS Data Success Guide series. His career didn't follow a straight line into governance — he went from auto mechanics to the Marine Corps, through defense contracting, IT infrastructure, and systems administration before landing in strategy and data governance. That non-linear path is precisely what makes his perspective valuable. Every insight he brings has been earned, tested in practice, and stress-tested against the organizational reality of a financial institution that's been operating for over 100 years.

In this episode of In Production, Nick and Lance go deep on what enterprise AI governance actually looks like inside a regulated institution — not the framework version, the real version.

What we cover:

The gap between paper-based governance and how people actually work on Thursday afternoon — and why most programs are built in a vacuum that doesn't survive contact with reality.

Why the KISS mindset (Keep It Stupidly Simple) isn't a framework — it's a filter. If your governance program isn't easy to understand, easy to explain, and easy to teach, it won't be adopted.

How Lance structured data contracts at Trustmark to keep authority with the people who already had it — and why that's the only way to avoid governance becoming an enforcement burden that nobody follows.

The hard truth about AI governance in banking: you can check every box in ISO 42001, satisfy NIST AI RMF, align with OCC guidance — and still have an AI system making decisions that no one in your organization can fully explain. Compliance is not effectiveness.

The 4 questions every bank leader must answer before greenlighting any AI use case — starting with: "Is the data driving this decision actually under governance yet?"

Why AI doesn't create data problems — it amplifies them. Inconsistent data quality, unclear ownership, and lack of governance don't disappear when you add AI. They scale.

What separates the banks that will win with AI in 5 years from those that won't. The answer isn't speed, budget, or vendor selection. It's whether the organization started with ownership and accountability before they touched the first model.

Why "accountability without understanding is just documentation" — and the difference between knowing who owns a decision and whether that person actually understands what they own.

The line from this episode worth sitting with:

"Just because you implemented AI to make decisions, you lose the right to say we didn't know."

Guest:

Lance Harlan — Data Governance Program Manager, Trustmark Bank. Author of the KISS Data Success Guide series on South Tech and Medium. Lance writes to work through governance challenges in real-world institutional environments, publishes only after testing what works, and connects with practitioners on LinkedIn.

📎 Connect with Lance on LinkedIn 📚
https://www.linkedin.com/in/lancewharlan/

Leading with AI Agents by Reddy Mallidi — Lance's current read and recommendation for this episode.

In Production is for CTOs, CIOs, CISOs, CDOs, and technology leaders who have moved past the theory. Every episode is a conversation with someone who has deployed AI in the real world — regulated industries, complex infrastructure, real stakes. No keynote speakers. No consultants with opinions. Practitioners only.

They fail because they mistake capability for readiness.

Debasish Bhattacharjee, Engineering Leader who has built and scaled AI systems across Fortune 500 organizations including Oracle, IBM, Broadcom, and SAP, has shipped seven production AI systems across Fortune 500 organizations — systems that collectively drive over $65 million in annual savings. Not lab experiments. Real deployments across expense management, procurement, HR, and customer support. Built with teams of 12 to 15 engineers. Shipped under a quarter.

In this episode of In Production, Debasish breaks down exactly what separates the pilots that looked impressive from the systems that businesses actually trust — and why the gap almost always has nothing to do with the model.

What we cover:

↳ Why his first AI deployment was off by 400% — and why the model had nothing to do with it. The data was the monster. Cleanup took four months. The model was ready in six weeks.

↳ The question that paused a $3M AI roadmap three weeks before launch. One room. Capable executives. Complete silence. What that silence revealed about organizational readiness.

↳ Why governance fails most enterprises — and what it looks like when it's built correctly. The difference between a permission gate and a feedback system.

↳ The triple lens: advisor, builder, operator. What each teaches you that the other two can't — and why the operator lens is the one most organizations are missing.

↳ The hidden 20% that lives in people's heads. Why asking employees to document what they do doesn't work — and what actually surfaces the undocumented rules before they become production incidents.

↳ Shadow mode as a process audit tool. How running AI silent alongside human decisions for 2–4 weeks before any automation reveals the broken process underneath — and why that's where the real savings are. One procurement deployment: $2.1M recovered from process redesign alone, before the AI system drove another $7–8M on top.

↳ The metric that doesn't lie. Why human override rate — and how it changes over time — tells you more about your AI initiative than accuracy, uptime, or any demo metric.

↳ Vendor lock-in and the pivot conversation. How to walk into a CTO's office and say "we made the right call with the information we had — the market moved faster than our assumptions" — and why that framing gets respected.

↳ What Debasish would tell any leader who just got the mandate to take AI from pilot to production scale: start with shadow mode, not automation.

If you're a CTO, CIO, or engineering leader who has been handed an AI mandate and is trying to figure out what separates the programs that ship from the ones that quietly die — this conversation is the one.

Debasish is reachable on LinkedIn at linkedin.com/in/debasishtech.

He's now the Director of Nevada's Office of Project Management, overseeing Core NV: the State's $200M ERP transformation and one of the rare large-scale government tech projects that actually hit its phase one timeline.

In this episode, Brian gets specific about what makes public sector transformation succeed — and fail. Nevada tried this once before. It didn't work. This conversation is about what changed the second time.

Chapters

• 00:00 Introduction and Project Overview
• 06:36 Teamwork and Urgency in Project Execution
• 14:14 Phase One Completion and Maintenance
• 19:30 AI in ERP Transformation
• 28:00 Future State of Government and AI in ERP
  

We cover:

→ Why most ERP projects in the public sector fail before the technology is ever a factor — and the single leadership condition that determines whether a program survives contact with the organization

→ How Nevada broke a core rule of project management (skipping the business process analysis) and what they discovered when they went into discovery anyway: processes that existed nowhere on paper, passed down from predecessor to predecessor for decades

→ The decision that changed the program's trajectory — telling every agency in the State that the process would change to fit the system, not the other way around. And what happened when they tested whether any laws or regulations actually needed to change (they didn't)

→ What an integrated team actually looks like in practice — contractors, vendors, system implementers, and State employees treated as a single unit with shared accountability. Why Nevada's first attempt failed partly because it didn't do this

→ The honest mistake: no plan for production support between phase one go-live and end of phase two. What it cost them in phase two capacity, and why the State's biennial legislative cycle made it nearly impossible to course-correct quickly

→ Where Brian sees AI landing inside a modernized ERP environment — not as a chatbot, but as a policy enforcement layer embedded directly into financial and HR workflows. The use case: a brand-new employee on day one gets the same institutional knowledge guardrails as a 25-year veteran

→ What Brian would tell other states that no vendor will ever say to them — one piece of advice about the relationship between a government and its system implementer that most jurisdictions get completely wrong

Brian's perspective is different from most guests on this show. He's not from the private sector. He's not selling a platform. He's a public servant running a program with real consequences — for State workers, for Nevada's citizens, and for the future of how government uses technology.

If you work in enterprise transformation, regulated environments, or anywhere the question is how to get a large, risk-averse organization to change — this one's worth your time.

Guest: Brian Bowles, Director — Nevada Office of Project Management Find Brian: linkedin.com/in/brian-bowles | opm.nv.gov

He started in the 90s — building neural networks and genetic algorithms inside Goldman Sachs's financial trading division. He went on to launch and run a quantitative hedge fund for 11 years, managing long-short equity assets with AI at the core. That fund was eventually sold to a bank. He filed the first machine learning patent ever issued in the United States. In 2019, he started collaborating with researchers at OpenAI — back when the lab was still largely unknown outside a small circle of practitioners. Most recently, he served as Head of Applied AI globally and Chief Data Scientist at Carlisle, where he architected and built the firm's entire investment AI infrastructure.

Chapters

• 00:00 The Evolution of AI in Finance
• 07:10 AI's Impact on Private Markets
• 13:41 Responsibility and Accountability in AI
• 23:07 The Future of AI and Recursive Self-Improvement
• 28:07 Organizational Transformation with AI
  

Dean calls himself an applied AI researcher. Not a theorist. Not a consultant. Someone who has taken these models into real environments, tested them against real constraints, and found out what actually holds.

In this episode, we go deep on what that experience actually looks like.

Dean breaks down why the proof of concept phase is a trap — and why skipping straight from experimentation to production readiness is the only move that leads anywhere. He explains the capabilities gap he saw forming early: not between humans and AI, but between what these models can actually do and how organizations are using them. His read then, and now, is that we're not even close to the ceiling.

We talk about what it takes to deploy AI inside regulated financial environments — private equity specifically — where there is no portfolio effect to absorb mistakes, and the fiduciary accountability sits entirely with a human being. Dean walks through the architecture he built at Carlisle: citation-grounded outputs, observation checkpoints at every stage of the pipeline, and a footnote agent trained on forensic accounting that caught an ASC 606 violation buried on page 933 of a data room — in 4 hours, not 3 weeks.

We also get into the organizational side of transformation. Why end users have to be part of the build process from day one. Why most enterprises skip the observation layer and pay for it later. And why Dean sees recursive self-improvement — not AGI debates — as the real frontier executives should be thinking about right now.

If you work in asset management, financial services, or any regulated industry trying to move AI past the pilot stage, this conversation is the one.

Dean can be reached at dean@dsconsult.ai

In this episode, Nick and Rodrigo go deep on the problems that don't show up in the pilot.

Chapters

• 00:00 Introduction and Background
• 10:18 Teaching AI to Non-Tech People
• 22:00 Cognitive Debt and AI Engineering
• 27:07 Defining the Problem in AI Initiatives
  

They start with the sorting problem: when your AI surfaces hundreds of patterns in customer data, how do you decide what's actually relevant? Rodrigo walks through how Loadsmart built a conversational layer on top of their freight analytics product — not because the RAG wasn't working, but because users didn't know what to ask. The fix wasn't a better model. It was ranking insights by context extracted from the user.

Then the conversation shifts to governance — and gets uncomfortable.

Rodrigo introduces what he calls the recursive trust problem. When a human operator can override an AI recommendation, who do you trust? The model that optimizes for margin, or the salesperson who has a relationship with the carrier? Both can be wrong. And most enterprises have no mechanism to track which one is failing them, or when.

His answer: trust is a function of maturity, not architecture. You start by trusting the human more. Over time, if you build the right feedback loops, you transition to trusting the model. The problem is — most teams make that transition without realizing it, and without the instrumentation to manage it.

The episode closes on cognitive debt. Rodrigo is direct: every engineer using AI-assisted tooling is quietly becoming a manager of agents. Junior work is being delegated before junior engineers ever develop the instinct to do it themselves. Nobody knows yet what that means for the senior talent pipeline in five years. He doesn't pretend to have the answer — but he's watching it happen in real time at Loadsmart, and he's concerned.

If you're building AI into production systems — in logistics, in any data-intensive vertical — this is the conversation that gets into the specifics most people skip.

Topics covered: → How Loadsmart used dynamic pricing ML from day one — and why daily retraining was non-negotiable → The RAG product nobody used — and the architecture change that fixed it → The sorting problem: why finding the signal is easy, ranking it is the hard part → The human-override loop and what it reveals about enterprise AI governance → Trust maturity: when to trust the model more than the operator → Cognitive debt — what happens when engineers stop building and start delegating → Agile vs. waterfall for AI: why the old tradeoffs need to be rethought

Connect with Rodrigo Senra: linkedin.com/in/rodsenra

Now he's doing something most executives at his level won't: writing openly about why 95% of enterprise AI programs fail.

Not the technology. The strategy. The leadership. The things that break before the first model is ever touched.

In this episode, Nick sits down with Neil to unpack the 7-Pillar framework behind his book "Why AI Fails" — and the uncomfortable patterns he's watched play out across defense, aerospace, energy infrastructure, and the Fortune 500 companies he consults with today.

What we cover:

↳ The real reason 95% of enterprise AI pilots never reach production — and it has nothing to do with the model

↳ Why "AI all the things" is not a strategy, and what boards and CFOs are starting to ask when the ROI doesn't show up

↳ The context problem: why two years of knowledge trapped in one employee's brain will break your first deployment

↳ Governance theater vs. real governance — how to build enough of it to move safely without slowing the business down

↳ Why documenting your current process is the wrong first move, and what Neil recommends instead

↳ The 50/50 split between top-down AI mandates and bottom-up tactical adoption — and why both fail for the same reason

↳ What the 5% of successful enterprise AI programs actually did differently (hint: they said no to a lot of things)

↳ Why the best AI results show up in 6-8 weeks, not 3 quarters — if you pick the right first project

↳ The CFO conversation that kills most AI programs, and how to get ahead of it

The throughline: AI doesn't fail because the technology is immature. It fails because leadership principles that have existed for decades get abandoned the moment the word "AI" enters the room.

If you're a CIO, CTO, CDO, or CISO responsible for an AI program right now — or about to be — this conversation is the peer-level diagnostic you won't get from a vendor pitch or an analyst briefing.

About Neil D. Morris:

Former CIO of Maxar Technologies, Ball Aerospace, and Redaptive. 25+ years in technology leadership across defense, aerospace, and energy infrastructure. Author of "Why AI Fails" and creator of the 7-Pillar AI Readiness framework. Based in Colorado.

Find Neil:

→ LinkedIn: Neil D. Morris

→ whyaifails.com

About In Production:

The podcast for CTOs, CIOs, CISOs, CDOs, and technology founders who have moved past the AI demo and are doing the real work of getting it into production. Hosted by Nick Melnychuk.

New episodes weekly.

That scene is the entire AI-in-procurement problem in one frame.

John Cornetta joins me for new episode of In Production. 30 years in procurement across banking, financial services, apparel, and hospitality. He's been in the rooms where AI gets deployed on data that structurally can't support it — and he's watched the quiet, expensive failures that follow.

In this conversation we get into:

↳ Why "crap in, crap out" is 10x more dangerous once AI is in the loop

↳ The 100-year-old bank with a 2014 contract authority table its own C-suite didn't know existed

↳ How a killed source-to-pay business case still shipped — 3 of 4 pieces, over 24 months

↳ The posture shift AI forces on procurement: from reactive to position-first

↳ What every executive needs to fix BEFORE the first AI vendor walks through the door

If you lead procurement, finance, or any operational function staring down an AI mandate — this one's a field manual, not a pep talk.

John https://www.linkedin.com/in/johndcornetta/

MIT PhD in AI and craft theory. MBA in operations and entrepreneurship. Hands dirty at Google. Then — surprising everyone — he went into enterprise. To learn how the machine actually works from the inside.

In this episode, Somesh shares the single most important thing he's discovered across a startup, a PhD, a research lab, and a Fortune 500 deployment:

AI cannot automate a secret.

It can reason through defined parameters. It can process structured and unstructured data at scale. But if the underlying logic driving a decision was never written down — if it lives in the head of a 30-year supply chain expert who has never been asked to explain why they do what they do — the model has nothing to work with. Trillion-parameter model or not.

We get into:

• Why the missing SOP kills more enterprise AI deployments than bad technology ever will
• The process layer as the actual track AI runs on — and what happens when it doesn't exist
• How Somesh built a personal AI agent to optimize credit card points — and why it broke immediately, and what fixing it taught him about enterprise architecture
• Why you don't need the best model. You need the right alignment. (And the open-source tool he uses to get there without a 570B parameter overhead)
• The human-in-the-loop conversation that actually matters: how to take a 30-year domain expert and make them the architect of the system — not the person it replaces
• Start with the map, not the machine — his direct recommendation for any operations leader whose leadership just made AI a priority

If you're responsible for getting AI into production in a complex environment — manufacturing, enterprise ops, regulated industry — this episode is the one to share with your team.

Somesh Mohapatra | AI & Craft Theory PhD, MIT | Connect with him on LinkedIn. https://www.linkedin.com/in/pikulsomesh/

Takeaways

• AI as an Operating Model
• Trust and Governance are Key
• Value Realization from AI Programs

Chapters

• 00:00 Advice for New CDO or Analytics Leadership Role

Takeaways

• AI implementation challenges
• The impact of AI on different industries

Chapters

• 00:00 The Evolution of AI in Business
• 07:18 The Role of Technical Skills in AI
• 14:43 AI Adoption in Healthcare
• 21:36 Understanding Human Workflows
• 28:02 Navigating AI Hype Cycles

Takeaways

• AI and transformation in healthcare
• The role of governance in AI adoption

Chapters

• 00:00 The Role of AI in Healthcare Transformation
• 08:01 Balancing Speed and Governance in Healthcare
• 21:18 Addressing the Gap Between Process and Knowledge
• 30:03 The Failure of Transformation Programs

Takeaways

• Operationalizing AI through managed services
• Challenges in AI adoption and governance models

Chapters

• 00:00 Introduction and Background
• 05:40 Managed Services vs. Software Products
• 13:06 Governance Models in AI Implementation
• 19:33 Challenges in Live Contact Center Environments
• 25:53 Starting the AI Journey in Organizations

Takeaways

• Transformation in Different Industries
• The Role of Technology in Business Value
• Challenges of AI Adoption in Enterprises

Chapters

• 00:00 Introduction and Career Background
• 05:29 Transitioning to ALM and the Importance of Clarity
• 14:06 Innovation Theater and Enforcing Discipline
• 24:25 Factors for Successful AI Programs
• 31:21 Preparing for AI Adoption in PE-Backed Companies

Takeaways

• AI in Enterprise
• Data Foundation
• Human Element in AI Projects

Chapters

• 00:00 Introduction to Data Analytics
• 07:12 Formal AI Education and Experience
• 13:27 Mid-Market AI Implementation
• 18:58 Automation and Decision Making
• 25:00 AI Roadmaps and Education
• 30:28 Future of AI in Enterprise

Takeaways

• Implementing AI in production requires addressing edge cases and testing model evaluation.
• Organizational alignment and human-centric AI implementation are crucial for successful AI adoption.

Chapters

• 00:00 Introduction and Career Background
• 05:13 AI Adoption and Organizational Challenges
• 10:17 Prioritization and Resource Allocation
• 16:05 AI Implementation Challenges
• 22:28 Measuring AI Success