THREE HR TASKS WHERE AI SAVES REAL TIME

Three HR tasks where an AI assistant saves time that stays saved, and the conditions attached to each. Two more where the saving usually turns out to be false.

Emmanuel Duncan · 7 September 2026 · 6 min read

Most HR teams have now tried a general purpose AI assistant on something. The results tend to split into two groups. A few tasks where the saving is obvious and repeatable, and a much larger group where the output looked impressive, took as long to check as it would have taken to write, and quietly stopped being used after a fortnight.

What separates the two groups is the shape of the task rather than the sophistication of the tool. The ones that hold up share three properties: the input is text you already hold, the output is a draft that a person will edit anyway, and a mistake is visible to whoever checks it.

Three tasks fit that description well. What follows is where I have seen the saving hold up, and the conditions attached to each.

Task One: Turning Long Documents Into Working Summaries

HR teams handle a lot of text that has to be read closely once and referred to often afterwards. Collective agreements, policy manuals, tender documents, statutory instruments, consultant reports, minutes of disciplinary hearings.

The reliable use is producing a structured summary of a document you have in front of you, with a clear instruction about what to pull out. Every obligation with a deadline. Every clause that mentions notice periods. Every commitment the organisation made, with the paragraph number.

The saving is real because the alternative is a person reading forty pages with a highlighter, and because the output is checkable. If the summary cites paragraph fourteen, you can look at paragraph fourteen.

The condition is that the document must be supplied to the tool rather than relied on from its training. A model asked about the content of a specific agreement without being given it will produce something plausible and wrong. That failure is the single most common one I see, and it is entirely avoidable by pasting the text in.

The second condition is that anything consequential gets verified against the source before it is acted on. Treat the summary as a well organised index rather than as an answer.

Task Two: First Drafts of Repetitive Written Work

Job descriptions, interview question sets, competency statements, induction checklists, letters that follow a standard shape, training outlines, policy sections written to a house pattern.

These are tasks where the blank page costs more than the editing. A team producing eleven job descriptions for a new department will spend most of its time on the first two and then copy. An assistant that produces a structured first version of all eleven, from the role titles and a description of the department, moves the work to editing immediately.

The saving is largest where a good template already exists, because the instruction can include it. It is smallest where the document is genuinely novel, which is also where the risk of accepting a generic draft is highest.

The condition is that the draft is treated as a draft. A job description that no manager has argued with is a job description nobody will recruit against. I would use the time saved on the writing to spend more of it on the conversation with the hiring manager, which is the part that determines whether the document is any use.

Task Three: Coding Open Text

Exit interview comments, survey free text, training feedback, grievance themes. Organisations collect a great deal of this and analyse very little of it, because reading four hundred comments and grouping them is a day of work that nobody has.

Classification against categories you define is a task these tools do well, and it is checkable by sampling. Define the categories yourself from a first reading, give the tool the list with a short definition of each, ask it to assign every comment to one and to flag anything that fits none, then check a random fifty by hand.

The saving is genuine, and it converts a dataset that was being ignored into one that gets reported.

Three conditions. Define the categories from your own reading, and do not ask the tool to invent them, because categories produced without context tend to be bland and to hide the specific complaint that mattered. Sample check every batch rather than the first one only. And strip names and identifying details before the text goes anywhere, which matters more here than in the other two tasks because the content is often sensitive and was given under an expectation of confidence.

Where the Saving Usually Turns Out to Be False

Two patterns worth naming, because both are common and both consume months.

Candidate screening and ranking. The time saving looks enormous and the risk sits in a place that is hard to see. A tool ranking applicants reproduces whatever pattern it was fitted to, and the organisation carries the consequences of a decision it cannot explain. If you use anything here, use it to check that applications are complete and to structure information for a human reader, and keep the judgement with a person who can be asked why.

Anything presented as an answer about your own organisation. Predictions of who will leave, engagement scores generated from text, sentiment measures on internal communication. These produce numbers with no visible working, and a number that cannot be traced back to records will not survive its first challenge in a leadership meeting.

The Conditions That Apply to All Three

Confidential material needs a tool with terms the organisation has actually read, and a rule about what may be pasted into it. Employee names, identification numbers, health information and disciplinary detail should be removed before anything leaves your systems. Where you cannot be sure of the terms, restrict use to documents that are already public.

Output that reaches a decision needs a named person who checked it. Not a policy saying output should be checked, a name.

And it is worth measuring the saving instead of assuming it. Time three of the job descriptions the old way and three the new way. The result is sometimes smaller than expected, and knowing that is more useful than a general impression.

Where to Start

Pick task one and a document your team is genuinely struggling to get through this month. Run it, check the citations, and note how long both the run and the checking took.

That single exercise tells you more about where these tools fit in your team than any vendor demonstration, and it costs an afternoon.

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.