Generative AI Fundamentals
Understand generative AI, retrieval and business uses through a document example. Check missing facts, unsupported claims and privacy before relying on output.
Generative AI creates outputs such as text, images, audio or code from patterns learned during training and the input supplied at use time. An application may also retrieve documents or call tools. Plausible output is not evidence that a statement is true, so factual work needs sources and checks.
Generative AI produces outputs such as text, images, audio or code in response to input. It learns patterns during training and uses those patterns when generating. The output can be useful without being a reliable statement of fact.
This is one type of AI capability. Classification, prediction and generation can coexist in the same application. It is misleading to divide all older AI into prediction and all current AI into generation.
Distinguish the model from the application
A language model can draft text from the context it receives. An application around it may search the web, retrieve a business document, run a calculation or call another service.
That means "nothing is retrieved" is not a sound description of every generative AI system. Ask which tools are actually available and what source the result used. A fluent answer alone does not reveal that.
Image-generation systems also use different technical approaches. Avoid treating one method, such as diffusion, as a definition covering every generator.
Four useful jobs to test
| Job | Example | What needs checking |
|---|---|---|
| Draft | Write an internal update from approved notes | Added facts and the intended audience |
| Transform | Turn a policy into a checklist | Missing exceptions or changed meaning |
| Extract | Put invoice fields into a table | Each value and its source |
| Answer from sources | Find a policy's reporting procedure | Current version and supporting passage |
These are candidate uses, not a claim that the system performs them correctly without review. Some tasks may be better handled by a template, a deterministic rule or an ordinary search.
A fictional source-grounded example
Assume a fictional policy says: "Travel requests require manager approval. International travel also requires security review."
Ask for a checklist using only that text. An acceptable draft should preserve both requirements and the condition that triggers the second one. If it adds a thirty-day deadline, that deadline is unsupported.
Now remove the international-travel sentence. The model should not keep the requirement because it appeared in an earlier example. This tests whether the answer follows the supplied source.
If the actual procedure is unclear, the correct next step is to find the responsible policy owner, not ask the model to fill the gap.
Write a brief that makes errors visible
State the task, source material, desired output and important boundaries. Ask the system to identify missing information and cite the passage behind each material factual claim.
For the fictional policy:
Build a request checklist using only the supplied policy. Preserve conditions and exceptions. Give the source passage for each requirement. Mark anything the text does not establish as unresolved. Do not invent approval steps or deadlines.
Check the result yourself. Asking for citations does not guarantee they exist or support the statement. Open the actual document and verify the wording and version.
Understand the important failure modes
The NIST Generative AI Profile addresses risks including confabulation and information-related harms. It is a risk-management resource, not a promise that a particular product avoids those failures.
For a business task, watch for unsupported details, omitted exceptions, incorrect calculations, exposed information and actions outside the requested scope. Better instructions can reduce some failures, but cannot guarantee correctness or fix every limitation.
Introduce it with a bounded trial
Choose one task with permitted, non-sensitive examples and an identifiable reviewer. Compare the assisted result with the current process, including time spent checking and correcting it.
Review the service's data handling before supplying confidential records. Decide who can authorize an external action. A generated email draft and an email sent to a customer are different outcomes.
Keep the trial results and stop conditions. Expand only when the observed result meets the task's requirements. There is no need to declare that every department must use generative AI before demonstrating one useful application.
Get the next one.
Field notes from real deployments, written the way this one was. No schedule promised and nothing sold to you. Unsubscribe whenever.
Frequently asked questions
Does generative AI always answer only from training data?
No. An application may supply documents, retrieve sources or call tools. Check what the actual system used and verify the supporting material.
Does a detailed prompt guarantee a correct answer?
No. Clear instructions help define the task, but outputs still need source, calculation and completeness checks appropriate to the work.
Where should a business start?
Test one bounded task with permitted examples, clear acceptance criteria and a responsible reviewer. Count correction and review effort before expanding.
Related guides
Need help implementing this?
Prometheus does AI consulting and implementation, remote, with onsite days available in Milwaukee. We set it up, train your team, and make sure it works.
let's talk