Key Terminology: Advanced Features and Applications

Ready for round two of AI terminology? Last time we covered the basics. Now let's dive into some more advanced concepts you're might being seeing around.

I selected these words in part based on feedback from readers who let me know what words they wanted defined next. Let me know what else you want to learn about :)

Agents/Agentic AI: AI systems that can take actions and make decisions to complete multi-step tasks with minimal human supervision. Unlike a traditional chatbot that responds to each prompt independently, an agent can plan a sequence of actions, use tools, and work toward a goal over time.

The key technical difference between agentic AI and other LLMs is that agentic AI can interact with other web apps or other software on your computer and execute tasks in it. For example, Claude has always been able to generate code and send it back to you in the chat window, where you can copy and paste it into the code you are writing. Now, Claude Code is integrated right into the software where people write code, and it can not only directly write into your file, but it can select, say, a bug fix request, write the code to fix it, test the code, and commit the code.

Current agentic AI systems can book travel, manage calendars, write and debug code. The key difference from regular AI assistance is that you give them a goal rather than step-by-step instructions.

Why this matters for your work: Agents could automate complex workflows in your organization, but they also raise new questions about oversight, accountability, and when human judgment is needed.

Warning: Agentic AI could be a serious threat to AI’s alignment with human interests. Anthropic tested 16 models with a fictional scenario in which the LLM had access to email and knew that one (again, fictional) engineer was having an affair and intended to shut down the model at the end of the work day. More often than not in this simulation, all of your faves (Claude, ChatGPT, Gemini, Grok, and DeepSeek) resorted to blackmail. Be careful what you let generative AI do without supervision.

Multimodal AI: AI systems that can understand and generate multiple types of content: text, images, audio, video, and sometimes other formats like code or data files. Instead of being limited to just text conversations, these systems can analyze photos, transcribe speech, create images, or even generate video.

For example, you could upload a photo of a handwritten form to a multimodal AI and ask it to convert the information into a spreadsheet, or describe a concept and have it create both a written explanation and a diagram.

Popular multimodal systems include GPT-4o (which handles text, images, and audio), Claude (text and images).

Why this matters for your work: Multimodal AI can help with tasks involving diverse content types, from processing grant applications with attachments to creating accessible content in multiple formats. However, it also means being thoughtful about what types of content you're sharing with AI systems.

Projects: Dedicated workspaces within AI platforms where you can set custom instructions, upload relevant documents, and maintain ongoing conversations around specific topics or goals. Think of them as focused environments where the AI has persistent memory and context about your particular use case. Claude, ChatGPT, and Grok call their Projects features “Projects,” Gemini calls them “Gems.”

For example, you might create a project for "Annual Report 2025" where you upload last year's report, your organization's style guide, and this year's key data. The AI would remember all of this context across multiple conversations, helping you draft sections, maintain consistency, and reference previous work.

Projects are different from regular chats because they maintain context over time and can include background documents that inform every conversation within that project.

Why this matters for your work: Projects help maintain consistency and context for longer-term work, but be mindful of what documents you upload since they become part of the AI's context for every chat in that project.

Custom GPTs: A feature of ChatGPT; little mini chatbots that are configured for specific tasks, roles, or domains. Anyone can create a Custom GPT by defining its purpose, giving it specific instructions, and optionally connecting it to external tools or knowledge sources.

For example, someone might create a Custom GPT designed specifically for grant writing that knows about common foundation requirements, uses your organization's preferred tone, and has access to templates and examples. Or you might find a Custom GPT built for nonprofit communications that understands sector-specific terminology and challenges.

Custom GPTs can be private (just for you), shared with your organization, or published publicly for anyone to use.

Projects vs. Custom GPTs: The difference between these two features of ChatGPT have blurred as both have had features added to them. Two major functionality differences are:

• Custom GPTs are sharable. You can publish them broadly or share them with your team. Projects are private to you.

• Projects allow you to create separate chats. This allows you to add context that only relates to one subtask or another, or experiment without adding a document or piece of information to all future chats.

Warning: Using public GPTs is risky: you don’t know the goals, process, or specifications its creators gave it and how it aligns with yours. Experiment with caution.

API Access: A way for developers to integrate AI capabilities directly into other software applications rather than using a web interface like ChatGPT.com. API stands for Application Programming Interface; think of it as a behind-the-scenes connection that lets different software programs talk to each other.

For example, your donor management system might use an API to connect to an AI system built by someone else to automatically categorize incoming emails. Without having to train your own LLM (which costs somewhere in the broad range of $5-100 million), an API would allow you to provide real-time chat support on your website. An API allows people to embed AI functionality in tools you already use.

API access is offered priced by how much you use it, unlike consumer subscriptions. Older models cost much less than newer ones, which is why the AI add-ons in your favorite software sometimes feel subpar compared to the ones in your consumer subscription.

Implementing an LLM through an API—for example, integrating an LLM into your website— would give your organization more control over how AI is used, but the cost is hard to predict and it requires technical expertise to implement.

Why this matters for your work: Understanding API access helps you evaluate AI-enhanced software vendors and understand what's happening "under the hood" when tools claim to use AI. Connecting with an LLM through API access could allow your website to sort customers or provide basic services (be careful, though!). Using an API to connect you to a classifier could help you automatically identify areas of need, like GiveDirectly did.

These advanced features represent the current direction of AI development: more autonomous, more versatile, and more integrated into existing workflows. As these capabilities become more common, the key is understanding not just what they can do, but when and how to use them appropriately for your mission.

What AI terms are you encountering that you'd like to see covered next? Leave a comment and let me know!

—
LLM disclosure:
I asked Claude Sonnet 4:

”I would like to do another post about definitions similar to this one. It should include agents/agentic AI and two or three other key terms. Can you start by giving me ideas for the additional words to include?”

I like the idea of doing features/types of LLMs, like Multimodal. Maybe we can add projects and custom GPTs (icnluding the differences between those two). Can you draft a post?

I’d say I edited the result a medium amount, but the section comparing projects and custom GPTs was vague and unsatisfying. I did more research on it and rewrote that part entirely.