AI Manufacturing Innovation: What Really Works

I'm Kurt Schmidt, founder of Schmidt Consulting Group, and after 300-plus episodes of The Schmidt List interviewing operators, technologists, and growth leaders, one topic keeps surfacing with more urgency than almost anything else: how B2B services firms and industrial companies actually put AI to work in ways that matter. The conversation around AI manufacturing innovation has gotten louder, but the substance behind most of it is thin.

Recently I spent time talking through this with BJ, an enterprise AI practitioner working across battery manufacturing, healthcare, and some of the largest technology organizations in the world. The insights from that conversation cut through a lot of the noise, and I want to lay out what I think every B2B leader needs to understand about making AI work in complex, high-stakes environments.

The core idea is this: AI model selection is an HR problem masquerading as a technology decision. Get that framing wrong and you'll spend a lot of money being frustrated by stochastic machines that refuse to behave the way deterministic software always did.

Why Is AI Manufacturing Innovation So Hard to Execute?

Executing AI in industrial and B2B services environments is hard because the mental model most leaders bring to it is wrong from the start.

Before generative AI arrived in force around early 2022, solving a business problem with software meant building a deterministic system. You told a machine what to do, and you could guarantee the output. Bugs happened, but they were programmed in. The relationship between input and output was fixed. That's the mental model baked into decades of IT decision-making, vendor selection, and implementation planning.

Generative AI broke that contract entirely. These systems are stochastic, meaning they're statistically driven and non-deterministic. You can be confident you'll get a thematically coherent response, but you cannot guarantee what that response will be. This is a profound shift, and most enterprise IT and procurement processes haven't caught up.

I've worked with services firms that hired capable developers, gave them API access to OpenAI or Anthropic's Claude, and then watched the project stall for months because the team kept trying to force deterministic behavior out of a probabilistic system. Getting an API key and integrating it the way you'd integrate any other application is the entry point, but it's nowhere near enough. Unless you understand how these models were trained, why they were structured the way they were, and what the organizational culture of the company that built them looks like in the model's behavior, you'll burn enormous time trying to get these things to do your bidding.

The firms winning at AI manufacturing innovation right now, including some operating at the scale of Amazon and Google, are the ones that understand the full depth of what they're working with. That depth includes the model's disposition, its tendencies, its guardrails, and yes, its cultural fingerprint.

Should You Treat AI Model Selection Like an HR Decision?

Yes, and this reframe is one of the most practically useful things I've come across in the AI space. Selecting an AI model for enterprise use should follow something closer to a rigorous hiring process than a software evaluation.

Think about what you'd actually want to know before bringing someone into a high-stakes role. You'd look at their background, their track record, their previous work, the values they operate from, and how their personality fits the culture of the team they're joining. You'd want to understand what biases or experiences might shape how they interpret problems. You'd think carefully about which role enables their success rather than exposes their weaknesses.

AI models reward exactly this kind of thinking. Anthropic's Claude behaves very differently from OpenAI's GPT-4o, which behaves differently from Google's Gemini. These differences aren't random. They reflect the cultures, priorities, and training philosophies of the organizations that built them. Claude tends toward caution and precision. ChatGPT is verbose and will take more risks, which means more creative output but also more hallucination exposure. Gemini sits somewhere in the middle, with a stronger factual grounding in many technical domains. None of these are inherently better. The question is which disposition fits the work you need done.

This is why the "how do I trust it" question that executives keep asking me is ultimately a question about organizational culture and risk tolerance. The technology works. The real question is whether you've selected the right model for your context, configured it with enough organizational context to behave in alignment with your values, and put appropriate guardrails around the outputs that matter most.

For Fortune 500 companies, this also means you can't simply route sensitive business data through a public-facing API. Data governance, security controls, and purpose-built deployments are table stakes at that scale. The cost and complexity of doing this right is real, but it's also the reason firms with genuine AI manufacturing innovation expertise command significant fees.

How Does Big Data Infrastructure Connect to AI Manufacturing Innovation?

The big data era wasn't wasted, even though a lot of organizations felt like they never fully got the return they expected.

A decade ago, companies poured money into data lakes, platforms like Snowflake and AWS Redshift, and teams of data scientists and analysts. Many of them ended up with all their data in one place and a growing realization that they were still years away from making meaningful sense of it. The ROI felt distant. The headcount was expensive. The insights were incremental.

Generative AI changes the equation on all of that stored data. The infrastructure built during the big data era, the cleaned and centralized datasets, the warehousing architecture, the metadata management, turns out to be exactly the foundation that makes enterprise AI deployments faster and more reliable. Organizations that did the hard work of getting their data into a single accessible place are now positioned to feed that data into AI systems in ways that would have been prohibitively expensive to build from scratch before 2022.

This is why I tell B2B services firms and their clients to stop treating the big data investment as a sunk cost. It's a foundation. And the firms that are pragmatic about layering generative AI on top of real, well-organized enterprise data, rather than expecting the AI to compensate for data chaos, are the ones getting real results.

The honest caveat: this still requires significant investment. Agentic AI architectures, where models act autonomously across multi-step workflows rather than just answering single questions, are genuinely powerful but genuinely complex to implement well. If you're a smaller services firm under 50 people looking for pure productivity gains, starting with tools like GitHub Copilot for development work or a well-configured Claude or Gemini workspace is a better entry point than trying to build a custom agentic system from scratch. AI project management automation

What Prompting Strategies Actually Improve AI Output Quality?

Most people use AI the way they used Google search five years ago, and that's why they're disappointed with the results.

Good Google search behavior means fragmenting your query into high-signal keywords. Short, specific, keyword-dense inputs. That's exactly the wrong approach for a large language model. These systems were trained on human language, full sentences, conversational turns, contextual explanations. When you communicate with them the way you'd communicate with a knowledgeable colleague, the output quality improves substantially.

A few approaches that consistently work:

Frame your expertise level explicitly. If you're asking about something outside your domain, say so: "Explain this to me like I'm an accountant trying to understand an engineering problem." The model will translate the concept into your frame of reference. This is one of the genuinely underrated capabilities of generative AI, cross-domain translation that would otherwise require either deep expertise or an expensive specialist.

Iterate like you're managing a smart junior employee. The model won't get it right on the first pass every time. Don't abandon it when it misses. Give feedback. "That was close but I need it to be more concise" or "You got the structure right but the tone is off" produces better second and third outputs than starting over with a new prompt.

Match the reading level to the audience. Running finished content through a model with instructions to simplify it to an eighth-grade reading level is about scannability. When people are reading a slide deck or skimming a proposal, they're processing at that level. The content doesn't get less intelligent; it gets more accessible.

Use voice input when possible. I switched a significant portion of my AI interaction to speech-to-text and the quality of responses improved noticeably. Because these models were trained on human conversation, spoken input often produces more natural and contextually coherent outputs than typed queries. I use this regularly now when I'm working through a problem in real time.

The broader point is that treating AI like a search engine produces search-engine-quality results. Treating it like a capable but junior team member, one that needs context, feedback, and clear instructions, produces work you can actually use.

What's the Biggest Long-Term Risk of AI Manufacturing Innovation Deployed Broadly?

The productivity gains from AI are real and they're significant. But there's a risk I think gets far too little attention in the executive conversations I'm part of, and it goes by a name I find genuinely clarifying: convergence.

Here's the structure of the problem. Every large language model is, at its core, an averaging machine. It was trained on the aggregate of human knowledge and expression, and it produces outputs that are statistically central to that training distribution. That's useful. But if every organization is using the same two or three dominant models, optimizing for efficiency, and routing more and more creative and analytical work through those models, the outputs across industries start to converge.

Fifty percent of all people are, by definition, below average. It's arithmetic. The same statistical logic applies to AI-generated outputs. Breakthrough ideas come from the edges of the distribution.

The people and companies that have historically changed industries, Steve Jobs framing the personal computer as a tool to change the world rather than a hardware problem, or the kind of neurologically atypical thinker who approaches problems from directions that don't emerge from consensus thinking, these people matter because they diverge from the mean. They bring perspectives that exist outside the training data. The risk is that widespread reliance on averaging machines, deployed broadly over a decade or more, gradually reduces the diversity of thought that produces the next genuinely new idea. You can still have the brilliant outlier leader. But if their team, their analysis, their strategic options, and their communication are all being processed through the same AI systems as everyone else, the outlier's use is diminished.

The partial antidote is open-source AI development. Meta's decision to open-source LLaMA and Google's release of Gemma are meaningful moves in this direction, because open-source models let smaller teams, academic researchers, and domain-specific operators build models trained on diverse data with different values and different dispositions. A monoculture of AI is the risk. Diversity of models, tools, and training approaches is the counterweight. Per the Linux Foundation's 2024 generative AI report, 41% of the generative AI infrastructure organizations deploy is open source.

I was recently part of a conversation at a major university executive committee reviewing curriculum implications of AI, and this convergence problem was the thing that most needed to be named explicitly. It still doesn't get named enough. security and B2B strategy

What Does AI Mean for the Entry-Level Talent Pipeline?

This is the societal implication that keeps me thinking, and I want to be precise about it because the common framing misses the actual problem.

The popular version of the AI jobs story is that AI will replace workers. I don't think that's where the real disruption lands, at least not in the near term for experienced knowledge workers. The real disruption is structural, and it hits entry-level roles first.

At firms like Deloitte, KPMG, or any large consulting organization, the traditional model has been to hire cohorts of analysts directly from university programs, run them hard on spreadsheet analysis, data synthesis, and slide production for two to four years, and let that process develop the next generation of senior practitioners. You learn by doing the grunt work. That's not glamorous but it's how expertise compounds.

AI handles a substantial portion of that grunt work now. A single experienced partner with strong AI prompting skills can now produce in a day what previously required a team of four analysts working a week. The math on entry-level hiring changes immediately. You don't need a team of six interns anymore. You might need one, or none.

The problem is the lost developmental on-ramp. If the entry-level analyst role disappears or shrinks dramatically, what replaces the mechanism by which junior talent learns the work deeply enough to eventually lead? We haven't solved this. Universities haven't solved this. The firms benefiting most from the efficiency gains don't have strong incentives to solve it on their own.

Humans are resilient and we'll figure out new pathways. But it's a real transition cost that's mostly being ignored in the productivity celebration. workplace culture and strategy

---

Key Takeaways

• Treat AI model selection as an HR decision. Study the cultural disposition, training data background, and behavioral tendencies of models the way you'd evaluate a candidate for a high-stakes role.
• Stochastic systems require a different trust framework than deterministic software. Stop expecting guarantees and start managing tendencies.
• The big data infrastructure built over the last decade is now the foundation for serious enterprise AI deployment.
• Prompt in full sentences and treat the model like a capable junior collaborator. Iterate, give feedback, and use voice input when it's available.
• Convergence is the most underappreciated long-term AI risk. Diversity of models, open-source development, and human divergent thinking are the counterweights.
• The entry-level talent pipeline is the most immediate structural casualty of AI efficiency gains. The firms that figure out how to rebuild that developmental pathway will have a long-term advantage.

I covered this topic in depth on The Schmidt List, and the conversation is one you should listen to if you're building AI strategy for a services firm or advising clients who are.

The question I keep sitting with: if every team in your industry is running through the same two or three models, what does your firm's actual differentiation look like five years from now?