Town halls have a structural problem, and most internal communications professionals know exactly what it is. The questions that get asked are rarely the questions most employees need answered. They come from whoever felt bold enough to type into the pre-event Q&A portal, whoever grabbed the microphone at the last minute, or whoever HR flagged as a reliable representative voice. None of those selection mechanisms reliably surfaces what the organization as a whole actually wants to know. Pre-event question clustering addresses this gap, but only if the collection and analysis process is set up to support it.
The Submission Volume Problem
Before a company-wide town hall, a typical internal comms team might collect anywhere from 150 to 800 pre-submitted questions, depending on organization size and how actively the submission was promoted. The format varies: Slido exports, Google Form responses, email prompts asking employees to send questions in advance. Once those submissions land, the team faces a familiar challenge with high-volume open text. What do you actually do with it?
The default approach is to read through the submissions and select the best ones. In practice, "best" means clearest and most concisely worded. That selection process introduces a meaningful bias: employees who write polished, well-structured questions move up the queue. Employees who describe the same concern in three run-on sentences, or who write in a second language, or who need a few hundred characters to give a concern its proper context, tend not to make the cut. The result is a town hall agenda built on the articulate minority's formulation of concerns that are often widely shared.
A clustering approach starts from the full submission set. Rather than filtering by quality of expression, it groups submissions by semantic similarity: what concern is this question pointing at, regardless of how it was phrased? That shift in framing changes what the agenda can be built from.
What Clustering Does to a Question Pool
When you cluster 400 pre-submitted town hall questions by semantic similarity, you don't end up with 400 themes. You end up with somewhere between eight and twenty distinct concern areas, depending on how broad or narrow the original question set is. Some clusters will be large. Suppose a 3,000-person organization runs a pre-event submission round, and 180 of the 400 questions resolve to some variant of "what is the plan for hybrid work policy going forward." Those 180 submissions are not identical. The surface language varies considerably: some are specific, some are abstract, some include personal frustration, some are logistical. But the underlying concern is the same.
That cluster output does two things at once. First, it tells you what is actually on employees' minds, with a frequency signal that isn't distorted by phrasing quality. Second, it gives you a principled basis for organizing the agenda. You are no longer choosing five questions from 400 based on editorial judgment. You are looking at the distribution of concern across the submission set and structuring the agenda around the themes that concentrate the most submissions.
The threshold question is worth thinking through carefully. A cluster containing 180 of 400 submissions is unambiguous: it goes on the agenda. A cluster of 12 submissions might represent a narrowly shared concern, or it might represent a concern specific to one team that deserves acknowledgment even if it doesn't warrant a full agenda slot. Cluster size gives you the distribution; human judgment still decides what to do with it.
Weighting Questions by Department and Concern Type
Not all clusters represent organization-wide concerns. Some themes are heavily concentrated in one department, one location, or one tenure band. A cluster of 40 questions about the performance review calibration process might be almost entirely concentrated among employees who joined in the last 18 months, with very little representation from senior staff who have lived through multiple review cycles. A cluster about a specific operational workflow might belong almost entirely to one business unit.
When you can break clusters down by the demographic profile of the employees who submitted them, it changes how you decide where each theme belongs. A concern shared across seniority levels, departments, and tenure bands should get substantive time in the main town hall, because it is an organizational concern in the full sense. A concern that is intense for one segment but barely present for others might be better handled in a targeted follow-up: a manager briefing, a dedicated Q&A session for that team, or a written communication that goes specifically to that group.
This weighting logic is not about dismissing minority concerns. It is about matching the venue to the concern. A concern felt by 40 people in one department deserves a response, but that response is more useful when it comes through the right channel, with the right specificity, rather than as a general statement from the stage that doesn't quite address what those 40 people actually asked.
Building the Agenda From Themes, Not Individual Questions
The practical output of a pre-event clustering process is a ranked list of themes with submission counts and, if demographic data is available, breakdowns by audience segment. An internal comms team that uses this as their agenda-building input produces something structurally different from a team that selected five polished questions.
Instead of "here are five questions we will address in order," the agenda becomes: "the three concerns most broadly distributed across the organization are X, Y, and Z; we will spend the first forty minutes there. Concern W is pressing for a specific group and will be addressed in a follow-up communication directed to that team." That structure also changes what leadership needs to prepare. Rather than crafting point-by-point answers to specific questions, they prepare substantive responses to themes, with enough depth and breadth that employees who asked the concern in a different form will recognize their question being engaged with.
One underrated benefit of this approach is that it changes the post-event experience for employees who submitted questions that weren't read aloud. If the town hall addressed the three themes that accounted for 70 percent of submissions, most employees will feel that the event engaged with what they cared about, even if their specific wording never appeared on screen.
What This Approach Does Not Do
It is worth being direct about the limits here. Clustering pre-event submissions does not remove the need for human judgment from the agenda-building process. A clustering output is input material, not a finished agenda. Internal comms professionals still decide how to sequence topics, how to handle politically sensitive themes, how to frame questions in a way that invites genuine dialogue rather than scripted reassurance, and how to allocate time across competing concerns. The clustering tells you what matters to employees. It does not tell you how to talk about it.
Clustering also does not guarantee that employees feel heard after the event. Whether employees feel heard is almost entirely a function of what leadership says and how they say it. A town hall built on a well-clustered question set can still be a bad town hall if the answers are evasive, incomplete, or condescending. The goal of this process is to ensure that the input going into agenda preparation is representative of actual organizational concern. What happens from there is a communication challenge that no text analysis tool can solve.
The Logistics: Collection Windows and Question Format
Pre-event question collection needs a minimum processing window: at least five to seven business days before the event to allow time for clustering, review, theme verification, and agenda revision. Collections that close 48 hours before the town hall rarely leave enough time for the analysis to influence preparation in a meaningful way.
The question format also matters. Open text with a soft character minimum, something like 30 to 50 characters, tends to produce better clustering material than short-form or yes/no submissions. Questions that include a sentence or two of context give the clustering process more to work with. "What is the WFH policy" clusters differently from "I'm in a role that requires some in-office presence but my team lead seems to expect full in-office attendance and I'm not sure what the official policy is." Both concern the same theme; the longer version gives both the clustering algorithm and the leadership preparation process more useful signal.
For organizations running town halls quarterly or more frequently, consistent pre-event collection builds a longitudinal record of concern. Not just what employees asked at this town hall, but how the thematic landscape has shifted across events. A concern that dominated six months ago and no longer appears with force may have been resolved, or may have shifted in character. A theme that appears for the first time and grows steadily across two or three events deserves attention before it becomes a crisis. That pattern visibility is something a single event's submission pool cannot provide on its own.