How AI Is Turning Employee Conversations Into Actionable People Analytics

People analytics runs on the parts of working life that were built to be counted. Headcount, tenure, salary band, promotion date, and a five-point answer to a question about whether someone feels recognised at work all load straight into a model, and none of them contains a word anyone said.

The words sit somewhere else. A manager and one of their reports meet every fortnight and spend half an hour on workload, a stalled project, and a promotion that keeps slipping. The surviving record is four lines in a shared doc, written afterwards by whoever remembered. Gallup has found that employees whose managers hold regular meetings with them are almost three times as likely to be engaged as employees whose managers do not, which puts a great deal of organisational weight on a conversation nobody writes down properly.

Software has started to close that gap. A tool for 1:1 meeting insights records the conversation, produces a transcript with speaker names attached, and runs language models over it to extract what was discussed, what was agreed, and what carried over from the meeting before. What comes out is a structured record of an exchange that used to leave almost no trace, and it is the input people analytics has been short of.

What a transcript holds that a survey does not

A survey asks someone to compress three months into a number. A transcript keeps the specifics: the name of the system that keeps breaking, the date the reorganisation was announced, the sentence in which someone said they had stopped applying for internal roles.

A few things can be pulled out of that material with reasonable confidence:

• Topics raised, and how often the same topic returns across a quarter
• Commitments made by either side, and whether they reappear unresolved in later meetings
• The split of speaking time between manager and report
• The ratio of questions to statements from the manager
• Whether career, development and progression come up at all, and in which meetings they do not

Those are all questions about behaviour and subject matter, which sit close to the surface of the text. Sentiment, morale and intent sit further from it, and the scores models produce for them move around a lot depending on the transcript and the prompt. A sentiment score can be a useful flag for a human to look at a conversation. Presented in a report as a measurement of how a team feels, it invites an argument the numbers cannot win.

Getting the conversation into a usable state

The modelling is the tractable half of this. Extracting topics and commitments from a decent transcript is a task current models handle well, and the prompts to do it can be written in an afternoon.

Capture is where the work sits. One-to-ones happen on Zoom, Microsoft Teams and Google Meet, across a desk, and on the phone. Each of those platforms comes with its own API, permission model, admin consent flow and release schedule, and each one changes on a timetable nobody outside that company controls. A recording that covers only the platform your organisation standardised on will miss the contractor on Meet, the manager who dials in, and the conversation that happened in a room.

Recall.ai, an API for meeting recording that is used by over 3,000 companies, handles that part through a single interface. It captures recordings, transcripts and metadata from platforms like Zoom, Microsoft Teams, Google Meet, Webex, as well as in-person meetings. This saves engineers from having to write and maintain an integration for each platform separately. Pricing is usage-based at $0.50 per recording hour or less, scaling down with volume, with no platform fee, no minimum and no commitment.

Transcription quality matters more here than it does in most text problems, because speaker attribution decides which person every number belongs to. If a transcript assigns the manager’s line to the report, the talk-time ratio for that meeting is wrong, and so is every aggregate built from it.

Questions the data starts to answer

Once 1:1s exist as structured records, some questions that used to be unanswerable become ordinary reporting.

Which topics recur in teams with high voluntary attrition and rarely appear in teams with low attrition? Which managers hold career conversations quarterly and which have not had one in eight months? Where do commitments get made and then quietly disappear? Which reorganisation announcements produced questions that were still being asked six weeks later?

Gallup’s long-running finding that managers account for at least 70% of the variance in employee engagement scores has been quoted in HR strategy decks for years without anyone being able to say which managers, or what specifically those managers do differently. Conversation data narrows that. Talk-time splits and question ratios are crude measures, but they are observable, per-manager, and repeatable month to month, which is more than an annual engagement score offers.

The output that gets used is usually a list of the teams where a named topic is climbing, handed to someone who can go and ask about it.

The governance question, answered before the build

Recording internal conversations sits under employment law, works council agreements and data protection rules that vary by country, and the requirements are stricter in the EU and the UK than most US-built tooling assumes. Consent, retention periods, who can see transcripts, whether a report can review what was captured, and what happens when someone objects all need answers before the first recording, because retrofitting them means deleting data you have already analysed.

A defensible default looks like this: aggregate reporting only above a minimum team size, no individual scores, transcripts visible to both participants, and a documented retention window measured in months. Writing those rules down before the first recording also gives you something to show a works council, which is a conversation that tends to happen whether or not anyone planned for it.

Start with one question worth answering, record only the meetings that bear on it, and see whether the answer changes anything a manager does. That is a smaller project than a people analytics platform, and it is the one that tells you whether the larger version is worth funding.