/tidal-point-20/images/tidalpoint_logo2024_hori_blk_med.png){kind=logo}
CONNECT
In the 1970's the "Internet Protocol" or IP was conceived. Combined with efforts to connect computer systems at research organizations in similar timeframes, the Internet was born. Gradually connectivity speeds increased, more people connected to it, and Internet-based applications were created (like Voice calling and Social Media), but they all fundamentally rely on "IP" (or more precisely in most cases, "TCP/IP"). A complete revolution in terms of how the world connected and conducted business.
The same way the Internet transformed connectivity between computers, AI is transforming intelligence between people and machines. But just as the early Internet needed applications like web browsers, email clients, and search engines to make its power tangible, AI also needs applications—purpose-built software layers that translate raw intelligence into real-world utility. That’s what we mean by an AI application.
An AI application is software that integrates artificial intelligence models—like large language models (LLMs), computer vision systems, or recommendation engines—into specific workflows or problem domains. Think of it as the difference between a search engine and the TCP/IP protocol that underpins its operation. The protocol moves data; the search engine makes it useful. Similarly, models like GPT-4o or Claude 3.5 provide general and often times wonderous capabilities, but AI applications deliver tailored outcomes by layering on domain context, constraints, user interfaces, and feedback loops.
The raw intelligence of a general-purpose model is powerful, but it’s not sufficient. Businesses need outcomes, not open-ended conversations. For instance, in cyber risk management—our domain at Tidal Point—the questions aren't generic. They’re framed by frameworks like SOC 2 or NIST, depend on integrations with infrastructure like AWS or Microsoft 365, and involve high-stakes decisions. An AI application in this space doesn’t just generate risk reports—it understands what’s material, what’s missing, and how to act on it.
This need becomes clearer when you look at what LLMs ("AI") don’t know: your internal policies, your industry’s acronyms, or what’s actually helpful in a given scenario. A model like ChatGPT might suggest controls for cloud risk, but it doesn’t know your S3 bucket configuration or the last time you rotated IAM keys. AI applications connect this gap by embedding intelligence into context, offering trustworthy actionability, and flagging inconsistencies that only make sense when the software is grounded in the user’s domain.
What's available off the shelf is quite amazing, and so it’s tempting to think that generic AI solutions will solve every problem. But that’s like saying raw Internet protocols would replace all applications. The OSI model is a technical reference to describe how different pieces of Internet technology fit together. This model may have outlined the theoretical layers of communication, but it was the creation of things like web browsers (the "Application Layer") that made it usable. It may be an imperfect analogy, but in AI foundation models are like the Transport and Network Layers—crucial, but probably invisible to most users. Applications are where value is created.
Generic chat models lack domain expertise, can’t self-diagnose errors reliably, and can (albeit less frequently) hallucinate. Dave, our CTO, likes to call LLMs "confident liars". The more specialized or regulated the environment, the more dangerous this becomes. That’s why notable AI companies (like Cohere, OpenAI, Anthropic, and Google DeepMind) increasingly focus on fine-tuning, plugin ecosystems, retrieval-augmented generation (RAG), and tool use—acknowledging that raw models must be steered.
AI applications are where intelligence meets infrastructure. Importantly, the connect to your own data and knowledge, but they also integrate with your tools, can follow your business logic, and evolve with user feedback. In security and risk, that might mean automatically correlating data from Jira, Slack, and your SIEM to surface issues that matter now. In sales, it might mean generating deal insights based on CRM and call transcripts. In all cases, it’s the application—not the model—that ensures reliability, traceability, and relevance.
We’re just at the beginning of this era. The Internet wasn’t born with Netflix and Google Maps; those took years of iteration, investment, and user feedback. Similarly, the most impactful AI applications of this generation are still being built. And like the Internet, the value of AI won't just come from the protocols or the models—but from the experiences crafted around them.
When we say “AI application,” we’re not just talking about throwing a chatbot into a product. We’re talking about deeply embedding intelligence into systems of work, with the right mix of models, data, and design. It’s not just about making AI accessible—it’s about making it useful, trustworthy, and transformational. That’s the mission we’ve embraced at Tidal Point with Risk Assist, and it's where the future of software is heading.