the job description is the wrong unit of analysis
job titles average away the variation that matters, and the decisions enterprises now face need a smaller unit than the job.
A regional managing director asks a question in a Tuesday operating review. Attrition in the mid-market sales team looks high. Is it worse than last year, is it concentrated anywhere, and are we losing the people we wanted to keep.
The HR business partner writes it down. It reaches HR operations on Thursday. HR operations discovers that "mid-market sales" is not a field in any system, and that three definitions of regretted attrition are in active use. The request moves to people analytics, which has a backlog. An answer lands eleven days later, with a caveat about the definition.
By then the operating review has happened twice more. A decision was made in the second one, on instinct, because waiting was not an option.
This gap between a question and its answer is the queue. It is the most expensive thing in most HR functions, and it is almost never measured.
the queue is a ratio problem before it is anything else
Insight222's People Analytics Trends research covers 372 organisations employing more than 20 million people across over 180 countries. It found that people analytics team growth has stabilised at roughly one practitioner per 2,500 employees.
Sit with that ratio. In a 25,000-person enterprise, ten people are responsible for answering every workforce question that anyone in the organisation might ask, alongside the recurring reporting, the data engineering, the compliance submissions and the board pack.
The queue is not a sign of a badly run function. It is arithmetic. Demand for workforce answers scales with the number of managers. Supply scales with a headcount line that is negotiated annually and rarely wins.
why self-service BI did not remove it
The promise of self-service analytics was that this ratio would stop mattering. Give managers a dashboard, and the queue disappears. Two decades of evidence say otherwise. Gartner has published research under the title "Why BI and Analytics Adoption Remains Low and How to Expand Its Reach", which is a candid summary of an industry-wide result.
Three structural reasons explain it.
Dashboards answer anticipated questions. A dashboard is a set of questions somebody decided in advance were worth asking, rendered as filters. Real questions from operating leaders are unanticipated and composite. "Are we losing the people we wanted to keep" requires joining performance, tenure, internal mobility and compensation, in a segment that was not a segment when the dashboard was built.
The hard part is definitional, not technical. Most of the eleven days in the example above went on establishing what "mid-market" and "regretted" mean, not on running a query. A self-service tool hands the definitional problem to the least-equipped person in the chain and calls it empowerment.
The audience is small and getting smaller. Gartner research published in October 2025 found only 8 percent of HR leaders believe their managers have the skills to use AI effectively. The same applies to analytics tools. A tool that requires competence the population does not have will be used by the people who already had the competence, which is the analytics team.
three latencies, one of them technical
It helps to break the queue into its parts, because they have different fixes. The time between a question and a usable answer is the sum of three latencies.
- retrieval latency. The time to find and join the data. This is the part that tooling genuinely addresses, and the part most investment targets.
- definitional latency. The time to agree what the question means in the language of the data. This is not solved by tools. It is solved by a governed model of the workforce with settled definitions of role, segment, performance and movement.
- trust latency. The time between receiving an answer and being willing to act on it. This is invisible in every dashboard usage report, and in many organisations it is the largest of the three.
Most analytics programmes attack retrieval latency, which is usually the smallest component. Definitional latency is where the days go. Trust latency is where the value goes, because an answer that arrives on time and is not acted on has produced nothing.
the cost you never see
Decision latency is the visible cost. The invisible one is larger: the questions that are never asked.
Leaders learn the response time of a system quickly. Once a manager knows that a workforce question takes two weeks, they stop asking the small ones. The small ones are exactly where early signal lives. Nobody files a ticket to check whether a hunch about one team's engagement is right, so the hunch is either acted on unverified or dropped.
You cannot count suppressed questions directly, but you can detect the pattern. If almost all requests to your analytics team come from a small number of senior sponsors, and almost none come from front-line managers, the queue has already trained the organisation not to ask.
what changes when an agent answers directly
The interesting shift is not that a language model can answer questions. It is what happens to each of the three latencies when the answering system sits on a governed model of work rather than on raw tables.
Retrieval latency collapses, which is the least interesting part. Definitional latency collapses only if the definitions are already settled in the underlying model, which is a modelling investment, not an AI one. Trust latency is the one that determines whether any of this matters.
Trust latency falls only when every answer carries its evidence: the population it was computed over, the definition applied, the date range, the exclusions, and a path back to the source records. An answer without that is a confident sentence, and a confident sentence from a system is worth less than a hedged one from a colleague, because it cannot be interrogated.
This is where most deployments are failing right now. Gartner surveyed 114 HR leaders in October 2025 and found 88 percent said their organisations had not realised significant business value from AI tools. That result is consistent with a large number of systems that answer fluently and cannot show their work.
the question ledger
Before buying anything, run this for four weeks. It costs one spreadsheet and about ten minutes a day.
Log every question asked of HR data, from any source: meetings, email, chat, ticket queue. Record six fields.
- Who asked, and at what level.
- The question, in the asker's own words.
- Time from question to first usable answer.
- How much of that time went to definition rather than retrieval.
- Whether the answer changed a decision.
- Whether the asker had to come back and re-ask.
Then sort every entry into one of four types.
- recurring and defined. Asked repeatedly, with settled meaning. These belong in a scheduled report, and if they are still being asked by hand, that is a distribution failure.
- novel but composable. New questions answerable from existing definitions by combining them differently. This is the largest category in most ledgers and the one that direct answering genuinely addresses.
- definitionally contested. The delay is an unresolved disagreement about meaning. No tool fixes these. Governance does.
- genuinely novel. Requires new data, new instrumentation or real research. These should stay in a human queue, and protecting analyst time for them is the point of automating the rest.
The ratio between the second and fourth categories tells you what direct answering is worth in your organisation. It also tells you what it will not solve, which is the more useful half of the finding.
what to do with the result
If the ledger is dominated by type two, the constraint is access, and an agent that answers in natural language against a governed model will remove real cost. This is the problem entomo built enzo for, answering questions the way a colleague would and attaching the citation to every response rather than offering it on request.
If the ledger is dominated by type three, buying an answering layer will make things worse, because it will produce fast answers to questions the organisation has not agreed on. Fix the definitions first.
Either way, the ledger is worth four weeks. The queue is currently invisible in every budget document, which is why it never gets funded and never goes away.