Every HR question has two dates.

Picture the first Monday of April. The HR director sends the board pack: 4,212 employees on 31 March. Three weeks later, the same report run for the same date says 4,198. Nobody deleted anyone. Fourteen terminations from a site closure were keyed in mid-April with effective dates in March. Both figures were correct when they were produced, and only one of them is in the minutes.

Two clocks in every record

HR systems are effective-dated. A pay rise, a transfer or a termination is stored with the date it applies from, and the system keeps the history so it can say what was true on any day. Separately, every change has the moment it was entered. The two drift apart all the time: a pay rise agreed in May and backdated to January, a termination keyed a week after the last working day, a transfer already in the system for next month that is not yet true.

The database world settled the vocabulary in the SQL:2011 standard. Valid time is “the time period during which a row is regarded as correctly reflecting reality”; transaction time is “the time period during which a row is committed to or recorded in the database”. A table that keeps both is bitemporal, and the idea goes back further, to Johnston and Weis's work on managing time in relational databases. HR systems implement the first clock well and expose it through an as-of date. The second is usually there too, as an audit field, and almost never reaches a report.

Where it goes wrong

1. Headcount that changes after the fact. A report for 31 March run on 2 April and again on 20 April gives two answers when a termination was back-dated in between.
2. Payroll that pays the wrong version. A retroactive change entered after the payroll cut-off needs a recalculation, and nothing flags it unless someone compares entry dates with the cut-off. A 2022 survey of 508 payroll and HR managers put average payroll accuracy at about 80%, at roughly 291 dollars per error.
3. Future-dated records in the wrong place. A transfer that takes effect next month is invisible to a query for today, and a planning question about next month needs it.
4. Questions with no date at all. “How many people are in Operations?” means today to one person and the end of last quarter to another.

Why an AI assistant makes it sharper

A person running a report picks the date in a prompt, even if they never think about it. An assistant turning a question into a query has to choose one, and the easy choice is today. Ask it about attrition last quarter and it may count the leavers correctly while measuring the headcount they left from on today's date. The answer looks precise and mixes two moments in time.

How we design briefs around it

1. Resolve the date first. Every question becomes an effective date or a range before any data is read, and when the question does not say, the brief states the date it assumed.
2. Show both dates. Every figure carries the date it is effective for and the moment it was read.
3. Explain the drift. When a scheduled brief disagrees with its previous run for the same date, it points to the records entered since.

What HRIS teams can check this week

1. Run last quarter's board-pack headcount again for the same date and compare. The difference is your back-dating rate.
2. List the changes entered after each payroll cut-off with an effective date before it.
3. Check whether your reporting tool, and any AI assistant you are testing, sets an as-of date or silently uses today.
4. Agree one rule for the board pack: figures are frozen at a stated entry date, and later corrections are reported as corrections.