Practical AI with Kendr

Understand what an AI answer is

Start with the decision, not the technology

By the end of this course, you will build a repeatable workflow that turns a small evidence pack into a decision brief. You will write prompts, inspect evidence, test failures, and package the result so another person can use it. You do not need programming experience. You do need to be willing to check the work: fluent text is a starting point for review, not proof that a task has been completed correctly.

Our running example is Cedar Learning, a fictional company preparing an onboarding workshop. Its operations lead wants a recommendation: should the team run a small pilot next month? The exercise documents, people, dates, and prices are invented for this course. You can use them without uploading confidential work material. Download the workbook provided with this lesson and keep your prompts, observations, and revisions in it as you progress.

Generation, retrieval, and action are different jobs

A language model generates a response from the information available to it and patterns learned during training. It can produce a convincing sentence about a document it has not seen. It can also misunderstand an instruction, miss a qualification, or present a likely interpretation as a fact. Your job is to make the task inspectable: identify which source should support each consequential claim and check that the response actually uses it.

Retrieval adds selected source material to the task. That material might come from a file you supplied, a knowledge base, or a research workflow. Retrieval improves access to evidence; it does not guarantee that the right passage was found or interpreted correctly. A source about last year's workshop cannot, by itself, establish next month's venue availability.

An action changes something outside the answer: sending a message, editing a file, or scheduling an event. A draft saying "the venue has been booked" is not evidence that a booking occurred. Confirm an action through the tool's result or the destination system. Throughout this course, activities produce drafts and recommendations. They do not authorize purchases, messages, or bookings.

Worked example: a polished answer with three hidden problems

Suppose you provide only this note:

Cedar is considering a workshop for 24 new customers.
The proposed date is 12 November, but the room is not confirmed.
The operations lead wants a recommendation by Friday.

An illustrative weak response says: "The 12 November workshop is confirmed for 24 attendees. Everyone will receive a recording. Book the larger room to meet strong demand."

The first sentence upgrades a proposal to a commitment. The second invents a recording policy. The third turns 24 intended customers into evidence of demand and assumes room options. None of those changes is justified by the note. The problem is not primarily tone; it is the relationship between the answer and its evidence.

A stronger response would say: "Cedar is considering a workshop for 24 new customers on 12 November. Room availability and recording arrangements are unknown. Before recommending the date, confirm the venue and clarify whether a recording is required." This answer is less decisive because the evidence supports less certainty. That is appropriate.

Your role in the learning loop

For each activity, record four things: what you asked, what information you supplied, what the response got wrong, and how you checked it. Do not collect only the best output. An unsupported claim is useful evidence about where the workflow needs improvement. Later modules will turn those observations into a test set and a release checklist.

The readings, exercises, and workbook are available through the Academy course. Running models, research, or knowledge-base jobs in Kendr may use credits under your account's normal usage rules. Those live runs are optional practice. You can complete each reasoning exercise using the supplied material and illustrative outputs without buying credits. Neither a video nor a paid tool run is a certificate requirement.

Guided exercise

1. Label each weak-response claim above as supported, contradicted, or not established.
2. Write a two-sentence replacement that preserves the proposed status of the workshop.
3. List two questions you would ask before taking any real-world action.
4. Save the result in the workbook's claim ledger. Keep the original weak answer beside your revision.