AI Enablement Fundamentals for OFA 1

This appendix collects and expands on the AI-Assisted Work guidance introduced in every module into a single reference.

• 1.What AI assistants are.
• 2.What they are good at.
• 3.Where they fail.
• 4.How to use them responsibly in OFA work.
• 5.It is intentionally written at a foundational level.
• 6.A deeper treatment of AI enablement for more advanced OFA work is provided in the companion OFA 2 curriculum and the Level 3-5 curriculum.

• 1.The AI writing and chat assistants referenced throughout this curriculum, such as Microsoft Copilot or similar approved tools, are built on large language models.
• 2.Large language models are systems trained on very large amounts of text that predict likely, plausible-sounding continuations of a prompt.
• 3.This means an AI assistant is very good at producing fluent, well-organized language.
• 4.However, it does not "know" facts the way a person who was in a meeting knows what was said.
• 5.It generates text that sounds right based on patterns, which is a fundamentally different thing from remembering or verifying what actually happened.

• 1.An OFA does not need to become a data scientist.
• 2.A new OFA will increasingly encounter projects that touch artificial intelligence and machine learning beyond just writing assistants.
• 3.It helps to recognize the basic vocabulary rather than treating every mention of "AI" as the same thing.
• 4.Machine learning, broadly, is the practice of training a computer program to recognize patterns in data rather than following an explicitly written set of rules for every situation.
• 5.It is commonly grouped into a few basic categories:

• 6.A foundational OFA does not need to be able to build or train any of these.
• 7.The goal is simply to recognize which category a project or tool falls into, since that shapes what the tool can realistically be expected to do and what kind of testing or validation it will need.

"Chatbot" is a broad term covering at least two meaningfully different kinds of systems, and confusing them is a common early mistake.

• 1.Rule-based chatbots follow a predefined decision tree·if a user selects or types one of a fixed set of expected inputs, the bot follows a scripted path to a scripted response.
• 2.These are predictable and easy to test exhaustively.
• 3.They are brittle outside their scripted paths - an unexpected question typically produces a dead end or a generic fallback message.

• 1.AI-powered chatbots use a language model or a trained intent-classification model to interpret a wider range of free-text input and generate or select a more flexible response.
• 2.These handle unexpected phrasing far better than rule-based chatbots.
• 3.They are harder to test exhaustively and can produce a confidently wrong answer in the same way a general AI writing assistant can.

• 1.Many real chatbot deployments are actually hybrid·a rule-based structure for well-known, high-stakes paths, with an AI-powered layer handling open-ended questions that fall outside the scripted paths.
• 2.For example, a rule-based structure might handle providing an emergency contact number.

• 1.For an OFA supporting a chatbot-related requirement or test effort, the practical implication is this:
• 2.Requirements and test cases for a rule-based chatbot can specify exact inputs and exact expected outputs, the same way a traditional test case does.
• 3.Requirements and test cases for an AI-powered chatbot instead need to specify a range of acceptable responses and explicit boundaries on what the bot must never say, since exact-output testing is not realistic for a generative system.
• 4.For example, the bot must never provide specific pastoral or medical advice.

• 1.Turning rough, disorganized notes into a clean, structured first draft.
• 2.Suggesting candidate open questions to ask before a stakeholder conversation.
• 3.Summarizing a long document or transcript into a shorter overview for a human to verify.
• 4.Drafting a first-pass comparison of two options, two requirements, or two arguments.
• 5.Explaining an unfamiliar term or acronym in plain language as a starting point for further confirmation.

The single most important risk for a new OFA to understand is "hallucination": an AI assistant can produce a confident, specific, plausible-sounding statement that is simply false.

• 1.Inventing a decision that was never made.
• 2.Inventing a person who was never in a meeting.
• 3.Inventing a system behavior that does not exist.
• 4.Because the output reads as confidently as accurate output, it is easy to miss unless the OFA deliberately checks it.
• 5.AI assistants also have no direct access to internal systems, real-time data, or anything that happened in a specific meeting unless that information was explicitly provided in the prompt.
• 6.An AI tool cannot look up what actually happened.
• 7.It can only work with what it is told.

• 1.Confidential member, donor, financial, and personnel information must never be entered into an AI tool that has not been explicitly approved for that category of data.
• 2.When in doubt about whether a tool is approved for a given type of information, the correct action is to ask a supervisor before using it, not to assume it is fine.
• 3.This applies even to seemingly low-risk uses, such as pasting a stakeholder's full name and phone number into a general-purpose AI assistant to "clean up" a note.
• 4.The safer habit is to remove or generalize identifying details before using an unapproved tool, or to use only tools specifically approved for that data.

A good prompt to an AI assistant generally does three things:

• 1.Gives the assistant the actual source material to work from (notes, a transcript, a document) rather than asking it to work from memory of a conversation it was never part of.
• 2.States the desired output format explicitly (a three-part summary, a table, a specific template).
• 3.Explicitly instructs the assistant not to add information that was not present in the source, and to flag anything unclear as a question rather than guessing.

The OFA 2 column below is a preview of what comes next, not a current expectation for an OFA 1.

• 1.An OFA used an AI assistant to summarize a lengthy stakeholder discussion from an automatically generated transcript, then distributed the summary the same day without reading the full transcript.
• 2.The summary confidently stated that the stakeholder had agreed to a specific rollout date.
• 3.In fact, the transcript showed the stakeholder had proposed that date only tentatively, pending confirmation from their own leadership - a distinction the AI summary had smoothed over.
• 4.The rollout was announced internally based on the summary.
• 5.It had to be walked back once the stakeholder's actual, more tentative position came to light, creating avoidable confusion and a credibility cost for the OFA.

• 1.After generating the AI summary, the OFA should have skimmed the source transcript specifically for anything decision-critical - dates, commitments, and numbers - before distributing anything.
• 2.This single habit, checking decision-critical details against the source regardless of how confident an AI summary sounds, would have caught the issue before it caused any real-world impact.
• 3.This case is the clearest illustration in this curriculum of why every AI-Assisted Work section repeats the same instruction: verify before you distribute.