Implementation Playbooks4 min read

AI Training for Employees: Check the “Nothing Found” Answer

AI training for employees should teach a coverage check for “nothing found” results: scope, source coverage, variants, sampling and residual uncertainty.

Bokili EditorialPublished October 28, 2026· Verified October 28, 2026

An employee checks an AI all-clear result against source coverage and finds one missed section

AI training for employees should include a result that sounds reassuring: “Nothing found.” That answer can be correct, incomplete or simply unsupported. The learner’s job is not to make the model search again at random. It is to show what was checked, what was not checked and how much uncertainty remains.

A positive claim can be checked against a source. An absence claim is harder. “No retention clause found” may mean the clause is absent, the wrong documents were supplied, an appendix was skipped, or the model searched for one phrase while the contract used another. Training needs a specific method for that difference.

AI training for employees needs a coverage check

NIST’s Generative AI Profile identifies confident false or erroneous content as a risk. The NIST AI RMF Measure playbook recommends assessing generated output against ground truth, using human domain expertise and documenting boundaries and limitations. For an employee checking a negative result, that means evidence of coverage matters as much as the final sentence.

The UK Government Data Quality Framework makes a useful distinction: completeness asks whether all required records and important values are present, while accuracy asks whether values match reality. A complete source pack can still contain errors, and an accurate page cannot prove that a missing appendix was checked. The two questions must stay separate.

	Unsupported all-clear	Coverage-based result
Scope	Implied by the answer	Named documents, sections and date range
Search method	One vague request	Terms, variants and required fields
Manual check	None	A small sample or critical section checked by a person
Conclusion	Nothing found	Nothing found within the stated scope; limits recorded

Use the COVER method for a “nothing found” answer

COVER

Clarify the scope

Name the documents, versions, sections, dates and question. Do not let “all files” remain undefined.

Observe coverage

Require a coverage record showing which sources and sections were actually examined, including unreadable or excluded material.

Vary the search

Check synonyms, abbreviations, related concepts and one counterexample. A phrase may be absent while the obligation is present.

Examine a sample

Manually inspect one critical section or a risk-based sample against the original source. Use a qualified reviewer where consequences are high.

Report residual uncertainty

State what the result does not prove, which material was unavailable and what would trigger escalation or a second review.

COVER turns a vague instruction to “verify the output” into an observable skill. It does not guarantee that every absence claim is true. It gives the employee and reviewer a transparent basis for deciding whether the check is sufficient for the task.

Worked example: the missing retention term

Use a synthetic supplier pack in training: a main agreement, a pricing schedule, a security appendix and a service description. Ask the learner to find every statement about how long customer data is retained after termination.

Reading is a start. Practice makes it stick.

Start learning

The AI replies that no retention term appears. The learner should not accept or reject that answer yet. First, they record that all four documents were supplied. Then they ask for the sections examined, search for “retain”, “delete”, “erase”, “archive” and “survival”, and manually inspect the termination and security sections.

Turn the result into evidence

1
Write the test question
Find every clause that sets or implies how long customer data remains after the service ends. Quote the clause and source location.
2
Require a coverage table
List each document and section checked, plus any file or page the tool could not read.
3
Run variant searches
Search for deletion, return, archive, backup, survival and post-termination language, not only the word retention.
4
Check one critical area manually
Open the original security appendix and termination section. Compare the model’s claim with the source.
5
Write a bounded conclusion
State either the located clause or: “No term found in the checked sources; backup schedule unavailable; legal review required before reliance.”

In this exercise, the clause appears under “data return and erasure” in the security appendix. The model’s first answer was not necessarily fabricated; the search was too narrow. The teachable skill is showing why the first absence claim was weak and what evidence improved it.

A negative result is not proof of absence

It is a claim about the coverage and quality of the search. Record both before anyone relies on it.

Design the exercise for the employee’s real context

European Commission AI-literacy guidance says training should reflect people’s knowledge, experience and the context and risk of the system they use. A low-consequence content search may need a light sample. A legal, safety, employment or customer-impacting decision needs approved processes and qualified review. The UK Government AI Playbook also stresses meaningful human control and validation at the right stages.

Score the learner’s coverage check

The scope names specific sources and versions.
Unreadable or unavailable material is recorded.
Search terms include relevant variants.
At least one critical source area is checked manually.
The conclusion states its limits.
The learner knows when and where to escalate.

Connect the skill to a wider learning path

Start by auditing the input pack, then write the expected answer and red flag. When several tools agree, remember that agreement is not independent evidence. Bokili’s training for HR and L&D can turn these checks into short role-based practice.

Run a ten-minute absence check

Choose a short, approved practice pack with one known hidden item.
Ask the AI whether the item appears.
Record the supplied sources and sections checked.
Search two alternative terms and inspect one section manually.
Write a conclusion that includes the remaining uncertainty.

Employees do not need to distrust every negative result. They need to know what would justify trust. Train the coverage record, not just the final answer.