WHAT TO FIX IN YOUR HR RECORDS BEFORE YOU BUY AN AI TOOL

An AI assistant reads whatever your HR system holds. Three records decide whether its answers are useful: your job architecture, your reporting lines, and your leave history.

Emmanuel Duncan · 30 August 2026 · 5 min read

Every HR software vendor now demonstrates an assistant. You type a question in plain language, and a paragraph comes back with a figure in it. The demonstrations are impressive, and they are run against data the vendor prepared.

Your data was not prepared. It was accumulated, by different people, under different naming habits, across several system migrations, with gaps where someone was on leave during a handover. The assistant will read it exactly as it is.

This is worth saying plainly, because the failure mode is quiet. A tool reading poor records does not return an error. It returns a confident sentence containing a wrong number, and the person who asked has no way to see the difference.

Three records carry most of the risk. Each can be assessed in a week.

Record One: Your Job Architecture

Job architecture is the structure connecting a person to what they do. In most HR systems this shows up as some combination of job title, role, grade, and department.

The common problem is that title has been allowed to do all the work. Titles are negotiated at hire, adjusted as informal recognition, and copied from whatever the previous person was called. Over a few years this produces an organisation where "Officer", "Coordinator" and "Associate" describe the same work in three departments, and where two people on the same grade doing the same job hold titles four words apart.

An assistant asked how many people work in procurement will answer from whatever field it can match on. If titles are the only reliable field, the answer will be wrong in a way that looks precise.

How to check it in a week: export every distinct job title with a headcount against each. Sort by count. Look at the tail, the titles held by one person only. In a healthy structure that tail is short. If a third of your titles are held by one person each, your architecture is titles, and you do not have a role layer at all.

The fix is not a full job evaluation exercise. It is a mapping table, one row per title, assigning each to a role family and a level. That table can be built in a spreadsheet and loaded as a lookup, and it is the single highest-return piece of data work in most HR functions.

Record Two: Your Reporting Lines

The org hierarchy is the field that lets any tool answer questions about teams, spans, and rollups. It is also the field most likely to be quietly broken.

Four things break it. Vacancies, where a departed manager stays in the system as the parent of a team. Acting arrangements, where someone covers a role without the record changing. Dotted lines, where a person's real supervisor differs from the recorded one. And loops, where a chain of managers eventually points back at itself, usually after a restructure.

A tool asked about a department's headcount will follow the recorded chain. If the chain is broken, the number will be missing people, or counting them twice, and nothing in the answer will indicate that.

How to check it in a week: run the hierarchy as a tree from the chief executive down and count how many employees are reachable. Compare that count to your total headcount. The difference is the number of people currently orphaned in your structure. Then look for any manager record whose own manager is themselves or is inactive.

This check takes an hour once someone can query the system, and in my experience it almost always finds something.

Record Three: Leave and Absence History

Leave records are the closest thing HR has to a continuous time series about people, which makes them valuable and makes their flaws consequential.

Three patterns to look for. Categories that have drifted, where "sick leave", "medical", and "illness" all exist and are used interchangeably by different administrators. Retrospective entry, where leave is recorded weeks after it happened, so the dates in the system are not the dates it occurred. And silent gaps, where a unit that manages leave on paper contributes nothing to the system at all, making its people look permanently present.

An assistant asked which department has the highest absence will confidently name the department that records leave most diligently.

How to check it in a week: count leave records per department per month for the last year. A department with sustained zeros is not a department with no absence. It is a department outside your system.

What the Tool Cannot Supply

There is a fourth category worth naming, because no amount of record cleaning will produce it.

Systems hold what happened. They do not hold why. They record that a role took four months to fill, and not that the first two candidates withdrew after the salary conversation. They record a transfer, and not that it followed a conflict. The reason sits in the memory of the people who were there, and when those people leave, it leaves with them.

This matters for AI specifically because reasoning is what these tools appear to offer. Asked why turnover rose in a unit, an assistant will construct a plausible explanation from the fields available to it. Plausible is not the same as correct, and a plausible explanation is harder to challenge than an obviously wrong one.

The practical response is to be clear about which questions your data can settle and which ones still require a conversation. That boundary should be explicit before a tool arrives, because afterwards it becomes very difficult to draw.

A Four-week Readiness Check

Week one, the title tail and the role mapping table.

Week two, the hierarchy reachability count and a list of orphaned records.

Week three, leave records by department by month, and a list of units contributing nothing.

Week four, write down the ten questions your leadership actually asks about the workforce, and mark which of them the cleaned records could answer.

That last step is the one people skip, and it is the one that determines whether the tool is worth buying. If eight of the ten questions turn out to be about why rather than what, an assistant will not help you, and the money is better spent on the diagnostic work that answers them.

The order matters. Fix the records, then buy.

Emmanuel Duncan is the founder of Aletheia Lunara Consult, a professional education and consulting practice working with organisations across Ghana and beyond on people analytics, workforce research and organisational capability.