Question 1

How are AI agents different from RPA (robotic process automation)?

Accepted Answer

_Agents add reasoning where RPA's rule engine stops._

RPA is a rule engine. It automates structured, low-skilled tasks where every decision is essentially zero or one, which is why UiPath and Automation Anywhere built real businesses but hit a ceiling: a script either breaks or runs, and it cannot handle uncertainty or unstructured data. Agents add a language model on top, so they can read an email's style, cope with ambiguity, and self-correct. Cherry's portfolio company Octomind does this for front-end testing with a self-healing loop that notices when a test goes wrong and tries another path, something a rule engine cannot do. The catch: you still have to give an agent clear boundaries, or it goes rogue.

— [The Edge: “AI Agents: The Future of Work?”](https://www.lutzfinger.com/transcript/2024-12-12-ai-agents-the-future-of-work/) · _Podcast_

Question 2

How can AI support frontline workers who aren't sitting at computers all day?

Accepted Answer

_Think voice-first interfaces and knowledge retrieval, not just desktop tools._

Frontline workers move around and might not have time to sit at a laptop, but they could have a quick second to speak or type a question. You could support them with an AI interface that answers questions in their language, translates into technical knowledge, and searches knowledge databases. The key is making sure answers are right. Don't just put ChatGPT on your FAQs. You need human in the loop at the beginning, testing a lot to ensure frontline workers are happy with the output before rolling it out widely.

— [You asked, we answered](https://www.lutzfinger.com/transcript/2023-07-28-you-asked-we-answered/) · _The Edge_

Question 3

How can startups use large language models to improve their products?

Accepted Answer

_Simplify interfaces and gain user insights through language_

As a startup founder, think about which interface will be changed by language input. Whenever users needed to use their mouse or learn complex clicking and dragging, you suddenly have the ability to let them do it via language. This simplifies the interface dramatically. Additionally, language prompting gives you more feedback. Before, you'd create an interface and see how users clicked buttons and moved around without really knowing their intentions. Now with language interfaces, you actually see what they're searching for and asking for. This gives you valuable input for building your product.

— [AI as an interface](https://www.lutzfinger.com/transcript/2023-04-24-ai-as-an-interface/) · _The Edge_

Question 4

How does AI change healthcare?

Accepted Answer

_AI applications, value-based care, regulation._

> I just got back from the JP Morgan conference. A few learnings that apply to any industry, not just healthcare:Humans + AI will change the industry. Not AI alone.Cash flow matters. Not AI hype.The real AI revolution is operational.One of the most hyperbolic promises in healthcare in the last month was the claim that: "AI is going to design our drugs and will replace researchers." In my view, this is not true.

— [Healthcare's AI Lesson: Autocomplete Isn't Understanding](https://www.forbes.com/sites/lutzfinger/2026/01/16/healthcares-ai-lesson-autocomplete-isnt-understanding/) · _Forbes_

Question 5

How does AI change the way we interact with computers compared to previous interfaces?

Accepted Answer

_From punch cards to conversation: AI as natural interface_

AI enables me to communicate with machines using natural language instead of learning complex commands or navigating graphical interfaces. Previously, we evolved from punch cards to MS-DOS commands to graphical interfaces with mice and clicking. Each step made things more intuitive, but I still had to search and learn the system's rules. Now with large language models, I can simply describe what I want in my own words. The AI understands my intent and produces output, even if I have no idea how it works technically. This is revolutionary because humans naturally communicate through language.

— [AI as an interface](https://www.lutzfinger.com/transcript/2023-04-24-ai-as-an-interface/) · _The Edge_

Question 6

How does Catalog ensure certain workflows always follow the same steps when using LLMs?

Accepted Answer

_A rules engine ensures critical workflows always execute consistently._

We built a workflow engine underneath our LLMs. The LLM generates a workflow on the fly every time, but users can design deterministic journeys inside our system using workflows (like Zapier but internal). If a request matches a workflow definition, it triggers that workflow every time instead of a dynamically generated journey. It's no code, so users can do it themselves. First, the LLM does a workflow lookup to see if there's something you always want to do, then for the rest it goes into free flow logic.

— [Large Language Models: one interface to rule them all?](https://www.lutzfinger.com/transcript/2023-08-15-large-language-models-one-interface-to-rule-them-all/) · _The Edge_

Question 7

How does Catalog prevent large language models from hallucinating when retrieving enterprise information?

Accepted Answer

_Stopping hallucinations by feeding LLMs the right information first._

The way to avoid hallucination is to stop the model from reaching into the model layers for information. If you give it structured or unstructured information in raw format and then ask it to summarize or list, it gives you exactly what you provided. Without that context, it makes things up because it's doing probabilistic retrieval from its training. This is retrieval augmented generation. You get the information first, augment the prompt with it, then generate the response based on that specific context.

— [Large Language Models: one interface to rule them all?](https://www.lutzfinger.com/transcript/2023-08-15-large-language-models-one-interface-to-rule-them-all/) · _The Edge_

Question 8

How does Catalog use large language models to simplify enterprise software access?

Accepted Answer

_One interface to access all your enterprise tools at once._

Catalog uses LLMs to create a unified interface for all enterprise software. Instead of navigating multiple systems, employees enter a single prompt. The LLM understands the query, identifies which data repository holds the answer (like Workday for payslips), converts the request into an API call, performs deterministic retrieval, and returns the result. We don't use LLMs for data retrieval itself, only for reasoning and synthesis. This compresses workflows from multiple surfaces and ten to fifteen clicks down to one query and two or three clicks.

— [Large Language Models: one interface to rule them all?](https://www.lutzfinger.com/transcript/2023-08-15-large-language-models-one-interface-to-rule-them-all/) · _The Edge_

Question 9

How does LegalOS use LLMs to make legal knowledge more accessible?

Accepted Answer

_LLMs translate legal knowledge into something everyone can use._

LegalOS originally built a no-code visual diagramming tool that lets legal experts encode their knowledge in a structured, machine-readable way. It's like a programming language for law. In recent months, we've pivoted to use LLMs much more heavily to make that legal knowledge accessible. LLMs provide a higher level of abstraction for automations. They let us map open-ended legal requests to structured tasks without requiring someone to manually encode every single rule. The LLM acts as an interface or translator, so humans don't have to understand legalese directly but can still interact with legal text.

— [Will LLMs scale and revolutionize the legal space?](https://www.lutzfinger.com/transcript/2023-06-15-will-llms-scale-and-revolutionize-the-legal-space/) · _The Edge_

Question 10

How has natural language processing changed since 2015, and why does that matter for legal tech?

Accepted Answer

_From counting words to real understanding in less than a decade._

I quit my PhD in 2015 and co-founded a natural language processing company. Back then, the state of the art was basically counting words in a glorified way. You had an idea of what the user might be talking about, but you couldn't map it to something with certainty. You had to collect your own data and train specialized models to do one narrow task. LLMs changed that. You can use foundational models out of the box without fine-tuning. They solve many tasks already. Understanding text and mapping open-world requests to closed sets of tasks became way more accessible.

— [Will LLMs scale and revolutionize the legal space?](https://www.lutzfinger.com/transcript/2023-06-15-will-llms-scale-and-revolutionize-the-legal-space/) · _The Edge_

Question 11

If agents are the future, why won't one giant model just do everything?

Accepted Answer

_The moat is the workflow, not a bigger model._

Because the durable approach is the opposite: smaller, faster, specialized models orchestrated together, not one ever-bigger general model. Even OpenAI orchestrates the long tail of consumer search with smaller models, because a big model would be far too slow. For founders this is the key warning: if your agent's next step is something generic, like deciding whether a calculator is needed, that is the long tail, and the big models will absorb it, just as Google eventually swallowed the specialized search engines. The defensible business is one where the founder knows the specific stepwise workflow of a domain and chains the right models and rules to serve it.

— [The Edge: “AI Agents: The Future of Work?”](https://www.lutzfinger.com/transcript/2024-12-12-ai-agents-the-future-of-work/) · _Podcast_

Question 12

Should you trust OpenAI blindly?

Accepted Answer

_Vendor risk and architectural diversification._

> Stargate's $500B investment — message to the competitors — meant to intimidate potential challengers with overwhelming financial backing. AI is extremely dangerous; only a select few should be allowed to develop it — OpenAI's storyline to use safety concerns as justification for a regulatory setup that creates barriers to entry and gives them more monopolistic power. From now on, models should respect IP rights — OpenAI ignored IP…

— [Is OpenAI A Victim — DeepSeek And IP Rights — A Playbook For Leaders](https://www.forbes.com/sites/lutzfinger/2025/01/29/is-openai-a-victim---deepseek-and-ip-rights---a-playbook-for-leaders/) · _Forbes_

Question 13

What are the four approaches to addressing bias in AI models?

Accepted Answer

_Four practical methods to reduce AI bias in production systems._

First, identify the bias by testing features that shouldn't have impact, like running loan applications 100 times to see gender differences. Second, help users understand potential bias through UX, like asking specifications when someone requests an image of a nurse. Third, correct the data through prompting, prompt tuning, or fine-tuning with better examples. Fourth, create guardrails or constitutions that let AI critique itself based on rules. Each approach has limitations, and changing underlying biased datasets in large models remains hard.

— [Bias in, bias out: Addressing the impact of bias in AI](https://www.lutzfinger.com/transcript/2023-06-07-bias-in-bias-out-addressing-the-impact-of-bias-in-ai/) · _The Edge_

Question 14

What is a data product?

Accepted Answer

_Definition and how to build one._

> Now, to be a good product developer, you need to think about self-referential loops within your products. How to get signal, how to get feedback from your users so your AI, or your data science, can leverage that information to make the product work better. So that moved a little bit within it. The innovation, and the inspiration about, and the curiosity about how to use data in novel manners, that's still data science, because that's…

— [Unit 3 Full Transcript](https://www.lutzfinger.com/ecornell/) · _eCornell course_

Question 15

What is an AI agent, really?

Accepted Answer

_Not a sci-fi character, a workflow that chains steps._

An agent is not a sci-fi character and it is not one giant model. It is a piece of code that takes a defined task, breaks it into steps, and lets you act on software through language. The core idea is chaining: instead of asking a model to "improve revenue," you break the goal into step one, step two, step three, exactly the way a manager hands work to a junior employee. Tools like LangChain popularized this. So an agent is less a breakthrough model than a workflow that combines models, tools, and sometimes plain deterministic code, each doing one step.

— [The Edge: “AI Agents: The Future of Work?”](https://www.lutzfinger.com/transcript/2024-12-12-ai-agents-the-future-of-work/) · _Podcast_

Question 16

What is the hallucination problem with large language models and why does it matter for businesses?

Accepted Answer

_LLMs invent answers when they don't know, risking business credibility_

Large language models will make things up when they don't know the answer. I demonstrated this by asking ChatGPT to summarize a completely fake Cornell website that doesn't exist. It generated a summary saying my course was the most amazing and made very good logical arguments, but it was completely fake. For businesses, this is a huge problem. If a customer asks a shop for something it doesn't have, the model might just make it up and promise red shoes with blinking laces. That's why proprietary data matters. You need to ground the model on your local data.

— [AI as an interface](https://www.lutzfinger.com/transcript/2023-04-24-ai-as-an-interface/) · _The Edge_

Question 17

What is the killer app of AI?

Accepted Answer

_Where the durable AI value will accumulate._

> Both are more about hype than reality. In 2025, there will be no AGI but large language models will find their "killer app." What are AGI and the Singularity? Artificial General Intelligence: an advanced AI that can think, learn, and solve problems across a wide range of tasks, just like a human. The Singularity: the idea of AI surpassing human intelligence, improving itself endlessly, and causing massive, unpredictable changes in…

— [No AGI But A 'Killer App' — 2025 AI Prediction](https://www.forbes.com/sites/lutzfinger/2025/01/03/no-agi-but-a-killer-app2025-ai-prediction-110/) · _Forbes_

Question 18

What is the missing moat in AI?

Accepted Answer

_Eval data as the durable moat._

> What wins the AI race? UX?

— [The Missing Moat In AI: Your Eval Data](https://www.forbes.com/sites/lutzfinger/2026/05/26/the-missing-moat-in-ai-your-eval-data/) · _Forbes_

Question 19

What patterns make a workflow 'agentic' rather than just a wrapper around an LLM?

Accepted Answer

_Reflection, self-correction, knowledge expansion, coordination._

A few recurring patterns. Reflection: the system plans, asks "is this approach good, does this make sense, can you double-check," and only then executes, the way a reasoning model breaks a request into several approaches, picks one, and verifies the answer. Self-correction: it looks at the result, notices it went wrong, and tries a different route. Knowledge expansion: when the answer is not in its memory, it searches the web for current information. Coordination: it splits a job into sub-tasks and routes each to the right tool, for example handing a math question to a calculator instead of guessing. Wrapping a chat box around an LLM is none of these.

— [The Edge: “AI Agents: The Future of Work?”](https://www.lutzfinger.com/transcript/2024-12-12-ai-agents-the-future-of-work/) · _Podcast_

Question 20

What practical business use cases exist for language model chaining beyond autonomous agents?

Accepted Answer

_Real business applications from databases to booking to summarization_

There are many business cases beyond the autonomous agent hype. Question answering over a database is one: use a language interface to access the database, figure out the information, distill it, and report back. Chatbots interacting with booking systems need to understand how to access the system and remember personal preferences. Document summarization is another: if I take all seven Harry Potter books, I can't paste them into OpenAI because it's a finite machine. Instead, summarize page by page, store that in memory, then do iterative summarization to get the final result. These applications are very practical.

— [Connecting LLMs to the world: The possibilities of the feedback loop](https://www.lutzfinger.com/transcript/2023-05-02-connecting-llms-to-the-world-the-possibilities-of-the-feedback-loop/) · _The Edge_

Question 21

What problem did transformers solve for Ultimate AI's multi-language support?

Accepted Answer

_One transformer model replaced 100 language-specific models per customer_

Before transformers, we had different models for different languages. Supporting 100 languages meant running 100 different models per customer, which was painful for model operations. When we switched to transformers like BART about three and a half years ago, we could use one polyglot model for all languages. You could train the bot in English and ask questions in Finnish, and it worked reasonably well. It also helped our clustering pipeline show customers all their questions across languages semantically grouped together. You could see Finnish, Spanish, German, and English examples all meaning the same thing without worrying about the language.

— [How Ultimate evolved from hundreds of supervised models to UltimateGPT](https://www.lutzfinger.com/transcript/2023-05-12-how-ultimate-evolved-from-hundreds-of-supervised-models-to-ultimategpt/) · _The Edge_

Question 22

What problems with RAGs still need to be solved before widespread adoption?

Accepted Answer

_The observability and control challenges that remain in RAG systems._

RAGs don't completely stop hallucination yet. The model uses both the trained corpus and new data, but we don't have a very good solution to manage that difference. Who decides when new information should override training data? Is the LLM taking that decision or the RAG? We lack observability and control. For example, I created a document saying I'm Captain America, but ChatGPT rejected it as fake. What if that was actually true breaking news? Amazing startups will help create this observability, solve access control, and manage the difference between what the corpus was trained on and corporate datasets.

— [Are we one step closer to using LLMs in practice?](https://www.lutzfinger.com/transcript/2023-07-21-are-we-one-step-closer-to-using-llms-in-practice/) · _The Edge_

Question 23

What was Ultimate AI's key insight that made customer onboarding faster than competitors?

Accepted Answer

_Automated clustering replaced manual training data creation for faster setup_

We couldn't afford AI trainers manually creating training data, so we built a clustering pipeline that analyzed historical customer conversations already sitting in CRMs like Salesforce. This let us automatically find the 100 or 200 most common questions and show customers all 30,000 actual ways users asked questions like where's my order. Competitors were using FAQ pages and having people manually write 50 to 100 variations, which didn't reflect real customer language. Our approach gave faster time to value because we used actual conversation data instead of imagined variations, and customers could see their real top questions before even talking to us.

— [How Ultimate evolved from hundreds of supervised models to UltimateGPT](https://www.lutzfinger.com/transcript/2023-05-12-how-ultimate-evolved-from-hundreds-of-supervised-models-to-ultimategpt/) · _The Edge_

Question 24

Where does Ultimate AI see its competitive defensibility beyond just AI models?

Accepted Answer

_User experience and distribution trump data or model advantages_

I don't think there's a silver bullet or secret IP that guarantees you always win. On the data side, we stream live conversations from big companies in a secured way. We have hundreds of millions of proprietary conversations that competitors starting today don't have. But more important than data moats or technical moats is user experience and distribution. I love products that take 10 minutes to learn but years to master, like Shopify or Excel. If you build something easy to start but with depth, solid brand, and good distribution, that matters much more than some little advantage in the back office models.

— [How Ultimate evolved from hundreds of supervised models to UltimateGPT](https://www.lutzfinger.com/transcript/2023-05-12-how-ultimate-evolved-from-hundreds-of-supervised-models-to-ultimategpt/) · _The Edge_

Question 25

Why are AI workflows the real moat?

Accepted Answer

_How redesigning work around AI beats model upgrades._

> SaaS can build the new AI-driven workflows on top of its user data and customer surfaces. But workflows are brittle. One new prompt, one model update, and the whole chain quietly breaks. I watch this every time I get students [coding agents in my Cornell workshops](https://ecornell.cornell.edu/courses/artificial-intelligence/building-enterprise-ready-ai-agents-from-prototypes-to-production/).

— [The Missing Moat In AI: Your Eval Data](https://www.forbes.com/sites/lutzfinger/2026/05/26/the-missing-moat-in-ai-your-eval-data/) · _Forbes_

Question 26

Why can't I just tell an agent to plan my honeymoon or buy things for me?

Accepted Answer

_Match the agent to the stakes._

Technically you can wire it up, but it will not be trusted yet, and for good reason. The blocker is not hallucination, which is largely a solved technical problem, it is stakes and trust: people do not want an agent that might book the wrong flight or charge the wrong thing to their card. The honest pattern is to match the agent to the stakes. Lutz started by imagining a health "nurse in your pocket," found the stakes too high to be reliable, and moved to lower-stakes commerce like helping someone find the right ring or coffee machine, where an agentic workflow that cleans the data and searches step by step adds value without the risk. High precision with an agent is possible; you scope it to where a mistake is survivable.

— [The Edge: “AI Agents: The Future of Work?”](https://www.lutzfinger.com/transcript/2024-12-12-ai-agents-the-future-of-work/) · _Podcast_

Question 27

Why did Ultimate AI initially focus on helping human agents instead of automating customer conversations directly?

Accepted Answer

_Agent assistance was a learning ground before full automation_

We turned the problem around. Instead of having consumers talk directly to bots, we gave recommendation tools to human agents. Agents had more tolerance for failures because they could ignore incorrect suggestions and still do their jobs, but when recommendations were correct they got value and could train the system. This was less stressful for building products. We could iterate and fail without ruining the customer experience. The technology wasn't good enough yet for full automation, so this approach let us learn gradually. Once models improved, we could then turn it back to direct user interaction.

— [How Ultimate evolved from hundreds of supervised models to UltimateGPT](https://www.lutzfinger.com/transcript/2023-05-12-how-ultimate-evolved-from-hundreds-of-supervised-models-to-ultimategpt/) · _The Edge_

Question 28

Why do AI image generators require increasingly complex prompts to get the results I want?

Accepted Answer

_Image generation gets complex fast, but workarounds exist_

When you create images with AI, you write something and get somewhat okay results, but then you iterate and the prompts become more and more complex. I've seen prompts that include things like hyper realistic, specific photographer names, lighting details, and many other technical terms all strung together. The interface is better than drawing with a mouse, especially if you can't draw, but it's not ideal yet. However, you can actually turn this around. Using tools like CLIP Interrogator on Hugging Face, you can upload a photo and retrieve the language needed to generate such an image.

— [AI as an interface](https://www.lutzfinger.com/transcript/2023-04-24-ai-as-an-interface/) · _The Edge_

Question 29

Why does 'good enough' beat perfect for AI products?

Accepted Answer

_Shipping iteratively vs chasing perfection._

> The core problem is that AI, by design, is average. It's spectacular at pattern recognition and synthesis, but it produces the most statistically likely output — not the most surprising, not the most brave, not the most emotionally resonant. ### What Made Steve Jobs Leadership Special Let's look at Apple in the iPod era.

— [AI Cuts Creative Teams: What Companies Must Learn About AI Leadership](https://www.forbes.com/sites/lutzfinger/2026/02/13/ai-cuts-creative-teams-what-companies-must-learn-about-ai-leadership/) · _Forbes_

Question 30

Why does Catalog change the user interface depending on the type of request?

Accepted Answer

_Dynamic interfaces that adapt to what you actually need to see._

We maintain about twenty to twenty-five abstract views, not thousands of deterministic flows like traditional enterprise tools. If it's a search, we render search results. If you ask for a summary, we render a summary view. If you ask for revenue from Q2 2023, we render a graph. The LLM dynamically adapts the interface to the use case. This is one level of abstraction higher than traditional UIs, which build very bespoke workflow-level interfaces. Ninety percent of use cases fall within the first five views.

— [Large Language Models: one interface to rule them all?](https://www.lutzfinger.com/transcript/2023-08-15-large-language-models-one-interface-to-rule-them-all/) · _The Edge_

Question 31

Why does e-commerce need fine-tuned AI?

Accepted Answer

_Why general-purpose LLMs aren't enough._

> To explore this question, I looked at one industry: e-commerce. In my online certificate programs, I've explored AI use cases across many industries such as healthcare, media, coffee, finance and e-commerce. While healthcare tends to be slowest to adopt AI, e-commerce consistently leads.

— [ShopTalk 2025: Just 1% Of Vendors Use Generative AI To The Max](https://www.forbes.com/sites/lutzfinger/2025/03/21/state-of-genai--only-1---shoptalk-2025-in-numbers/) · _Forbes_

Q&A, AI Product

How are AI agents different from RPA (robotic process automation)?

How can AI support frontline workers who aren't sitting at computers all day?

How can startups use large language models to improve their products?

How does AI change healthcare?

How does AI change the way we interact with computers compared to previous interfaces?

How does Catalog ensure certain workflows always follow the same steps when using LLMs?

How does Catalog prevent large language models from hallucinating when retrieving enterprise information?

How does Catalog use large language models to simplify enterprise software access?

How does LegalOS use LLMs to make legal knowledge more accessible?

How has natural language processing changed since 2015, and why does that matter for legal tech?

If agents are the future, why won't one giant model just do everything?

Should you trust OpenAI blindly?

What are the four approaches to addressing bias in AI models?

What is a data product?

What is an AI agent, really?

What is the hallucination problem with large language models and why does it matter for businesses?

What is the killer app of AI?

What is the missing moat in AI?

What patterns make a workflow 'agentic' rather than just a wrapper around an LLM?

What practical business use cases exist for language model chaining beyond autonomous agents?

What problem did transformers solve for Ultimate AI's multi-language support?

What problems with RAGs still need to be solved before widespread adoption?

What was Ultimate AI's key insight that made customer onboarding faster than competitors?

Where does Ultimate AI see its competitive defensibility beyond just AI models?

Why are AI workflows the real moat?

Why can't I just tell an agent to plan my honeymoon or buy things for me?

Why did Ultimate AI initially focus on helping human agents instead of automating customer conversations directly?

Why do AI image generators require increasingly complex prompts to get the results I want?

Why does 'good enough' beat perfect for AI products?

Why does Catalog change the user interface depending on the type of request?

Why does e-commerce need fine-tuned AI?