The Resilient Entrepreneur, Edition #135
Hi there
I hope you had a great week!
Here are the topics in today's edition:
- AI in Your Product: A Token Guzzler and a Security Risk
- AI in Your SaaS Product: Easy to Build, Hard to Define Use Cases
Please reach out with comments, questions, or suggestions for articles!
Talk soon,
Tom
TACTICS FOR RESILIENT ENTREPRENEURS
AI in Your Product: A Token Guzzler and a Security Risk
Do you really need a chatbot in your software, or is an FAQ page good enough? To answer this question, consider token costs and security risks.
“Does your product have AI?” That’s probably the most common question software entrepreneurs hear nowadays.
Is it a sensible question?
Think about it. AI is just a tool. A very powerful tool, but still a tool. It’s a bit like people asking you, “Does your tool have a database?” or “Is your company using electricity?”
This article is not about questioning or glorifying AI. It’s about two often-overlooked aspects of AI: Token costs and AI security.
It’s Gonna Cost You
LLMs consume tokens to process users’ prompts and generate answers. The model is simple: the more text you send to the LLM, the more tokens it consumes. The more comprehensive the LLM’s answers are, the more tokens it consumes.
Because most cloud AI providers such as ChatGPT, Claude, and Gemini still offer free tiers, users think AI is free. Wrong. Those AI providers pay for the free-tier users’ tokens, hoping they will move to a paid tier for better performance and more functionality.
This isn’t the whole story. When you integrate AI capabilities into your software product, there is no such thing as a free tier. You need to connect your product to your AI provider’s console. And this console measures your product’s token consumption from the very first token. The console will also generate a neat invoice for you at the end of every month.
If your users use the AI features more than you expected, your invoice will go through the roof. If your cloud AI provider increases token prices, there is not much you can do.
What alternatives do you have? If you don’t want to use cloud AI providers, you can always set up an on-premises AI stack. You might not get all the gimmicks of the frontier AI models, but for many use cases a second-tier AI stack is good enough. But never forget, operating and maintaining an on-premises AI stack will also cost you a ton of money.
People Will Try to Outsmart Your AI
Let’s assume you found a way to manage the token costs and your AI features are live in your product. Most probably, your product will feature a chatbot where users can interact with your product in a similar way they know from ChatGPT, Claude, or Gemini.
The chatbot interface is more dangerous than it might look. Like in real life, users can talk to your product in their own words. That’s a huge usability advantage, but it also risks letting malicious users extract information from AI that is not intended for outsiders.
Why is this possible? Always remember that AI answers are statistical rather than deterministic. This means that no two answers to the same question will ever be exactly the same (a bit like humans, really). In most cases, that’s not a problem, but it opens a security backdoor: You never know for sure how the AI will answer a user’s question. And if a user tricks your AI into providing unintended answers (for example, asking for a dump of your entire database), you didn’t just add token costs to your product, but also a rather big security loophole.
The mechanism to protect your product from your AI model is called guardrails. These are instructions to your AI model on what kinds of questions and answers it should never process. Guardrails provide some security, but hackers and fraudsters will find loopholes in new guardrails as soon as they appear.
With all its advantages, AI will always be vulnerable to attacks on its guardrails. Why? Because it’s designed as a statistical answer generator and it was trained on vast amounts of data. There are simply too many possible answers from an AI model to test them all. And that’s the real security loophole: We’re relying on production systems that weren’t fully tested, and there is no possibility to fully test them even if we wanted to.
Conclusion
Do you really need a chatbot in your software? Or is a simple FAQ page and a link to your support portal enough?
This question cannot be answered once and for all, but you should consider token costs and security aspects before deciding whether to integrate AI into your product.
STRATEGIES FOR RESILIENT ENTREPRENEURS
AI in Your SaaS Product: Easy to Build, Hard to Define Use Cases
Technically, it’s easy to integrate AI into your SaaS product. From a use case perspective, it’s not. Here is a real-world story.
Everybody wants to integrate AI into their software products. At Yonder, the B2B SaaS company I co-founded, we are no different from other software companies. But compared to B2C SaaS customers, B2B customers care very much about data sovereignty these days. And naturally, these days, they also care about AI sovereignty.
In our last annual user conference, we made AI sovereignty and AI integration the cornerstone topics of all the presentations and discussions.
Since the typical attendees at our user conference were business representatives and not IT dudes, we had to frame the discussion accordingly. Nevertheless, we decided to include a super-technical session to show our guests how complex AI sovereignty has become.
Here is an overview of the three key sessions we provided during the user conference.
Easy: What is an MCP server?
As part of our product roadmap update, I explained in simple terms what a Model Context Protocol (MCP) server is.
The Model Context Protocol (MCP) provides a new way for different software applications to communicate with each other. Instead of implementing APIs for every service, an MCP server is a standardized bridge between your SaaS product and other products. And that bridge is AI: Any AI model can talk to any MCP server the same way. The only thing you have to do is expose tools for your SaaS product on your MCP server, and other applications can talk to your product through AI.
I love using analogies, especially for explaining technical concepts to non-tech people. I used the war between the U.S. and Iran as an analogy. The Pakistanis are the MCP server. They shuttle between the Americans and the Iranians with text messages, incorporating the envoys’ answers into their own message to the other party.
As in diplomacy, everything in AI is based on text and spoken language. And both diplomats and MCP servers incorporate the answers from other people or software tools into their own answers.
So far, so easy.
Hard: What is AI sovereignty?
After a break, we cold-showered the attendees with a super-technical session delivered by Microsoft, our core hosting provider. The speaker threw around lots of tech expressions: Sovereign public clouds, sovereign private clouds, data residency, data boundaries, bring-your-own-key, confidential computing, foundry models, local foundry.
All clear?
Our audience was equally puzzled as you might be right now, but they took away one key learning: Maybe AI sovereignty isn’t as easy as it sounds.
Time for the lunch break.
Hardest: Now, what do you want to do?
After lunch, we embarked on the hardest discussion: What exactly do you want to do with your AI tool stack and the Yonder software?
It sounds almost trivial. But when you expose tools for your SaaS product on an MCP server, you need to plan carefully to ensure your customers use the integration only in the way you want them to. That’s because each user will “speak” to your SaaS product in their own words, rather than using a deterministic user interface with fixed buttons and inputs. This will keep doors wide open for all sorts of edge cases that are hard to debug.
Our discussions went back and forth, but by the end of the session we agreed on a very small set of tools our customers would need on the MCP server to fulfill the diverse use cases they described.
And that’s why the addressable universe for your SaaS product will multiply with an MCP server hosting just a few curated tools.
About Me
I’m a tech entrepreneur, active reserve officer, and father of three — writing about entrepreneurship, leadership, and crisis management from hard-won experience. No AI, no fluff, no promos. Just plain-text insights for people building and leading under pressure.
When I’m not solving problems, I find clarity in the mountains around Zermatt.
If this was useful, here’s how to get more:
📌 All my articles, no paywall — read everything in one place. Visit the blog.
📌 Buy me a coffee—it keeps the writing going. Thank you.
