How to get good at AI at work when nobody trained you
Choose a low-risk task, define a good result and check the output against evidence. Practice with public or invented material before considering approved workplace use.
Learning Cadence editorial, , updated , 6 min readHow we work
A practical AI skill is knowing when a tool is useful, what information it may receive and how to check its output. A general course can teach foundations; a bounded task lets you practice applying them.
The exercise below uses invented notes and a known answer. It is an original teaching example, not output from a provider test or evidence that AI improves productivity.
The short answer
Practice one task with a known answer and check every material claim.
Start with an approved tool and public or invented input. Define the result, test a small task and compare the complete output with your source and requirements. Record errors and review effort. A fluent explanation or one correct fact does not establish that the rest is reliable. Ask a qualified person to review work you cannot assess.
The method
Six steps, in order.
Use a small, checkable task and keep the human decision and data boundaries explicit.
01
Choose a low-risk task and permitted input
Use an approved tool and public, synthetic or otherwise authorized material. Do not paste customer information, confidential work, credentials or personal data into an unapproved service.
02
Write a quality check before prompting
State the purpose, audience, required facts, forbidden additions and format. Decide who can judge the output and what would make it unusable.
03
Give the tool clear instructions
Supply the allowed source and ask it to identify missing information. Clear instructions help define the task but do not guarantee compliance or accuracy.
04
Check the complete result against evidence
Verify material facts, calculations and citations independently. Check omissions as well as invented content. An explanation can itself be wrong; one correct fact does not validate another.
05
Compare the whole workflow
Include preparation, generation, checking and correction when comparing time or effort with your normal method. A quick first draft is not the same as a usable result.
06
Record limits before widening use
Keep a permitted record of the task, inputs, errors and checks. Try relevant different cases before expanding use. One successful example is not evidence of general reliability.
In practice
A worked prompt with an answer key.
Invented public workshop notes: "Tuesday, 2pm. Topic: choosing a community-garden location. Compare Site A and Site B. No location decision has been made. Facilitator: Lee."
Practice instruction: "Turn only these notes into a three-item agenda. Preserve the time and facilitator. Do not choose a site or invent a budget. Mark missing details as unknown."
A valid result can list: 1. Confirm the location criteria. 2. Compare Site A and Site B. 3. Identify missing information and agree the next step. It must retain Tuesday at 2pm and Lee as facilitator. Criteria are a proposed agenda discussion, not facts already decided.
An answer saying "Approve Site A with a $5,000 budget" fails twice: neither a decision nor a budget appears in the source. Correct time and facilitator do not make those additions trustworthy. You can perform this check without running an AI tool.
In practice
When the answer is outside your expertise.
NIST's Generative AI Profile describes confident false content as a risk and recommends checking generated sources and citations. Fluency, a detailed explanation and a plausible citation are not independent evidence.
Use primary sources and qualified review where the consequences require them. If a material claim cannot be checked, label the uncertainty and do not rely on it for a consequential decision. Learning the subject and using a tool can complement each other; neither removes the need for verification.
In practice
Turn the exercise into a learning goal.
We make Cadence. Its courses on AI at work provide short lessons and practice, including thinking about accuracy and privacy. They do not certify a workplace workflow or grant permission to share information.
A useful goal is: "Learn to evaluate a draft agenda against the source notes." Apply the same answer check to your own safe practice example before moving to harder work.
Put it into practice
Start it today, one lesson at a time.
If you want to keep going, Cadence turns what brought you here into a course with practice and saved progress. Learn at your own pace. Cadence is our app.
Choose a low-risk task, use an approved tool with permitted input and define what a correct result contains. Practice checking the result, including omissions and unsupported additions.
Does checking one fact make an AI answer trustworthy?
No. Check every material claim you intend to rely on. A correct date does not validate an invented budget, recommendation or citation elsewhere.
Should I ask AI to explain its reasoning?
An explanation can help you inspect a proposed approach, but it can also contain errors. Check the steps and evidence independently rather than treating the explanation as proof.
How can I tell whether AI saves time?
Compare complete workflows, including preparation, review and correction, at an acceptable quality level. A single fast draft does not establish a reliable saving.
Do I need a general AI course?
A course may help with foundations. Task practice adds a concrete way to apply and check them. Choose learning that fits your knowledge, responsibilities and permitted tools.
See it in concrete terms
Explore example courses for your goal.
These Cadence courses cover AI at work and the data judgment underneath it, one lesson at a time.
