Glossary: Custom AI Development Terms Every Firm Leader Should Know
Agents, orchestration, MCP connectors, deployment, handoff - defined plainly for the people who need to make decisions about building, not implement the build themselves.
The vocabulary around AI development has proliferated faster than the plain-language explanations. If you are leading a professional services firm and evaluating whether to build something, you will encounter these terms in proposals, vendor conversations, and technical discussions. This glossary is for you - someone who needs to understand the concepts well enough to make good decisions, without needing to implement any of it yourself.
Terms are ordered roughly from foundational to more specific.
Agent
An AI system that can take actions, not just produce text. A basic AI interaction is a conversation - you ask, it answers. An agent goes further: it can browse the web, read files, run code, call external tools, and take sequences of steps to complete a task. The key distinction is that an agent operates; it does not just respond. When people talk about "building an AI tool" for a specific firm workflow, they are usually describing some form of agent.
Multi-agent workflow
A setup where multiple AI agents work in sequence or parallel, each handling a specialized part of a larger task. Think of it like a team with roles: one agent monitors incoming data, another analyzes what it finds, a third formats the output for a specific audience. Multi-agent workflows tend to appear in more complex builds where no single prompt or agent handles everything cleanly. They produce better results on complex tasks because each agent can be optimized for its specific role, but they require more design and testing to get right.
Orchestration
The logic that coordinates which agent does what, and in what order. If a multi-agent workflow is the team, orchestration is the conductor - the system that decides when to pass a task from one agent to the next, what happens if something fails, and how the overall output is assembled. In practice, orchestration is often where the real engineering work lives in a complex build. Getting it right is what makes a multi-agent system reliable rather than fragile.
MCP connector
A standardized way to connect an AI system to an external data source or tool. MCP stands for Model Context Protocol - a specification that defines how AI systems and external services communicate. In practical terms, an MCP connector is what lets an AI read from your Google Drive, pull from your CRM, or write to Notion. Without connectors, an AI can only work with information you paste directly into a conversation. Connectors are what make a tool feel like part of your actual work environment rather than a separate application.
RAG (retrieval-augmented generation)
A technique where an AI retrieves relevant information from a document library before generating a response. Rather than relying only on what the model learned during training, a RAG system searches your firm's actual documents - past proposals, research, client files, internal policies - and uses what it finds to inform its output. This is why a RAG-enabled tool gives you answers that reference your specific data rather than generic information. It is the most common technique for making AI output specific to your firm's context.
Fine-tuning
Training a model on your firm's data so it learns your patterns - your voice, your terminology, your way of structuring documents. Fine-tuning modifies the model itself, rather than just giving it documents to search. It is more expensive and time-consuming than RAG, requires a substantial volume of high-quality examples to work well, and is usually overkill for professional services applications. Most firms do not need it. When a vendor suggests fine-tuning as a first step, that is worth questioning.
Deployment
Putting the built system into an environment where your team can actually use it. Deployment is the step between "this works in development" and "this is a tool your team uses in their actual workflow." It involves decisions about where the system lives (a web app, a desktop tool, an integration with software you already use), who has access, and how it handles real-world usage volumes. A tool that has not been deployed is a demo. Deployment is what makes it real.
Handoff
Transferring ownership of the built system from the builder to you. A real handoff includes the code, the documentation, and access to the environment where the system runs. If you end the relationship with the builder, the tool keeps working and you can maintain or modify it. A handoff that only gives you a running system without the underlying code or documentation is not really a handoff - it is a dependency. When evaluating a development partner, ask specifically what the handoff includes. Ownership of what gets built is a first-order question.
Discovery
The phase of a build engagement spent understanding the problem before designing a solution. Good discovery involves mapping the workflow the tool will support, identifying the edge cases that matter, understanding what quality looks like from the senior person's perspective, and specifying what the system needs to produce. Discovery is not a formality. The quality of the discovery document is usually the best predictor of whether the build works. Engagements that skip discovery - moving straight to building based on a brief conversation - almost always require significant rework later.
Context window
How much text an AI model can hold in working memory at once. If you hand a model a 200-page document, but its context window only holds the equivalent of 50 pages, it cannot process the whole document in one pass. This is why very long documents sometimes need to be processed in chunks, or why RAG (which retrieves relevant sections rather than loading entire documents) is often more effective than simply pasting everything in. Context windows have grown significantly over the past two years - what was a hard constraint in 2022 is much less of one now - but it is still a relevant consideration for document-heavy workflows.
Token
The unit AI models use to measure text. Roughly three-quarters of a word, on average - so a thousand words is approximately 1,300 tokens. Costs for AI API usage are measured in tokens (per thousand or per million), as are model limits. You do not need to think about tokens in daily use, but when evaluating build costs or usage-based pricing, token counts are the relevant unit. A tool that processes long documents at scale will use substantially more tokens than one that handles short inputs.
Prompt template
A reusable, structured prompt designed for a specific task. Rather than writing a new prompt from scratch each time you want AI help with a particular kind of work, a prompt template provides the structure - the instructions, context, and format - that you fill in with the specifics of a given instance. Prompt templates are the building blocks of more structured AI tools. A well-designed template produces consistent output across different users running the same task. Building a library of them is one of the most durable investments a firm can make in its AI practice.
Skill
A packaged, reusable AI workflow any team member can run with consistent results. More structured than a prompt template, a Skill encodes not just the instructions but often the data sources, the steps, and the output format for a specific task type. The goal is that a junior associate and a senior partner running the same Skill on the same input produce outputs that are similar in quality and format. Skills are what turns individual AI fluency into firm-level capability - when a Skill is well-built, it carries some of the senior person's judgment in its design.
These terms appear in different combinations depending on what is being built. A simple tool might be a prompt template with a RAG connection to your document library. A more complex one might be a multi-agent workflow with orchestration logic, several MCP connectors, and a custom deployment. Understanding the vocabulary helps you ask better questions of whoever you are working with - and catch proposals that use impressive terminology to describe something simpler than it sounds.
If you are trying to orient yourself to the full picture - what levels of AI practice exist and how they connect - the curriculum explainer maps that out. When you are ready to think about what to build, the Apparatus custom development practice starts with a structured discovery process.