Responsible AI Practice4 min read

AI Training for Employees: Check the Numbers Before You Use Them

AI training for employees should teach the UNIT check: label the unit and period, expose the base, then recompute the total.

Bokili EditorialPublished October 30, 2026· Verified October 30, 2026

A reviewer checks an AI-generated document against units, time periods, calculation bases and recomputed totals.

AI training for employees should include a short number check before any AI-generated figure enters a brief, email, spreadsheet or decision. A fluent sentence can hide a monthly figure presented as annual, a percentage without its base, a total that excludes tax, or a calculation that no one has repeated.

Teach one visible habit: label the unit, name the period, expose the calculation base and recompute the total. The goal is not to turn every employee into an analyst. It is to stop unlabelled numbers from borrowing credibility from polished prose.

AI training for employees needs a number check

The UK Government Analysis Function advises checking formulas, units and whether outputs fall within sensible ranges. Its analytical standard separates verification—was the analysis carried out correctly?—from validation—was the right analysis used? The updated AQuA Book adds that assurance should be proportionate to the financial, legal, operational and reputational risk of the intended use.

These principles translate well to ordinary AI-assisted work. Employees often do not need to inspect a model. They do need to know what a number represents, where it came from, whether the comparison is like-for-like and who should review it before action.

	Fluent number	Decision-ready number
Unit	“Costs fell by 12”	“Costs fell by €12,000”
Period	“Revenue rose 8%”	“Revenue rose 8% quarter on quarter”
Basis	“Defect rate is 3%”	“3% of 420 inspected units”
Total	“Option B saves €18,000”	“€18,000 before tax and implementation cost; recomputed from the source rows”

Use the UNIT check before a number is used

UNIT

U — Unit

State the currency, quantity, percentage, time, weight or other measure. Do not let a bare number travel.

N — Named interval

Name the day, month, quarter, year or comparison window. Check that both sides use the same period.

I — Input base

Expose the numerator, denominator, population and important inclusions or exclusions behind the figure.

T — Total

Repeat the arithmetic from the source values, then compare the result with the AI-generated claim.

Worked example: two supplier quotes

An employee asks an AI tool to compare two software-support quotes. The answer says Supplier B is “18% cheaper” and recommends it. The source pack looks simple, but Supplier A quotes an annual fee including implementation while Supplier B quotes a monthly fee, excludes implementation and uses a different user count. The percentage is arithmetically neat and operationally useless.

Reading is a start. Practice makes it stick.

Start learning

Turn the comparison into a checked number

1
Write the units
Convert both prices into the same currency and identify whether figures are per user, per month or per contract.
2
Align the interval
Convert recurring charges to the same period and separate one-off costs.
3
Match the base
Use the same user count, service scope and tax treatment for both options.
4
Recompute
Calculate each comparable total from the source rows without copying the AI’s percentage.
5
State the limit
Report what remains excluded, uncertain or subject to confirmation before recommending an option.

After the check, Supplier B may still be cheaper. The difference is that the recommendation now rests on a comparable annual total, a named user base and visible exclusions. If one quote remains incomplete, the correct result is a question for the supplier, not a more confident percentage.

Arithmetic can be correct and the comparison still wrong

A recomputed percentage does not repair mismatched units, periods or populations. Align the basis before calculating.

Make assurance proportional to the use

A rough internal estimate may need a quick UNIT check by the author. A customer price, staffing decision or financial commitment may need an independent reviewer and the original source file. Training should teach both the check and the escalation rule. The European Commission’s AI-literacy guidance emphasises measures suited to staff knowledge, experience and the context in which AI is used.

Release a number only when

Its unit is explicit.
Its period and comparison window are explicit.
Its numerator, denominator or population is identifiable.
Important exclusions and assumptions are visible.
The arithmetic has been repeated from the source values.
The review level matches the consequence of using it.

Connect number checking to a wider learning path

First label forecasts and scenarios correctly. Then audit the input pack and use the “nothing found” coverage check when absence is the claim. Bokili’s training for HR and L&D turns these checks into short, role-specific practice.

Run a ten-minute UNIT drill

Choose a fictional AI-written paragraph containing two numbers.
Add the unit and time period to each number.
Write the base or denominator behind one percentage.
Recompute one total from three source rows.
Name the reviewer required before the number is used.

Numerical AI literacy begins before advanced analysis. It begins when an employee refuses to pass on a number that has no unit, period, base or reproducible total. That small pause turns a plausible figure into a reviewable one.

Practise the check on real roles

Bokili helps employees rehearse compact verification habits on work-relevant cases, with the task, evidence bar and feedback adapted to their role.