People analytics has a hidden data deficit. AI will expose it.

Most mature people analytics functions are operating from a data deficit they cannot see from inside their own instruments. The deficit is not in the stack. It is in what the stack was never built to capture.

If you lead a people analytics function, you already know the shape of your data. HRIS records. ATS pipelines. LMS completions. Engagement scores. Performance ratings. Maybe a workforce planning model that runs against finance. These are the signals your function has spent a decade learning to harvest, clean, and present to the business.

Here is the uncomfortable part. Almost none of that data describes how work actually gets done.

The missing context layer

Who decides what gets escalated. Where the handoffs break between recruiting and onboarding. Which manager is quietly absorbing the workload of two open requisitions. How a benefits exception actually moves from a Slack message to a resolved case. These signals, the ones that describe how the organization functions as a system, are almost never encoded in your tools. They live in the heads of the people doing the work, in DM threads, in unwritten norms, in the meetings nobody recorded.

This is what Nonaka and Takeuchi called tacit knowledge in The Knowledge-Creating Company nearly thirty years ago. It is the body of undocumented knowledge that makes the organization actually work. And it is the exact context layer AI needs to do anything useful inside HR.

Generative models do not fail in HR contexts because the models are weak. They fail because they are asked to operate on a process description that bears no resemblance to how the process actually runs. The org chart says one thing. The HRIS workflow says another. The work itself is a third thing entirely, and nobody has written it down. Gartner has reported for years that the majority of generative AI projects stall at or after proof of concept. The post-mortems point almost uniformly to the same root cause: insufficient business context.

The diagnostic question your stack cannot answer

Readiness Diagnostic:

If a generative model were given access to every system your HR function operates, could it produce an accurate description of how a single end-to-end process actually runs today, including the exceptions, the workarounds, and the people who quietly hold it together?

If the honest answer is no, you do not have an AI readiness problem at the model layer. You have a workforce knowledge problem at the data layer. And no procurement cycle, vendor demo, or governance committee will close that gap on its own.

Why this is a people analytics problem, not an IT problem

There is a temptation to push this to the technology function. Resist it. The technology function will instrument the systems they already own. They will not interview a benefits administrator about the seven exceptions she handles every Friday that never touch a ticket.

Across our engagements, the organizations that have ceded this ground are the ones now scrambling to re-enter the AI conversation through the CHRO's office rather than leading it. People analytics spent the last decade, as Insight222's annual research has consistently documented, building a seat at the strategic table. The context layer for AI is the next claim on that seat. Whoever owns the description of how work actually gets done will be the function the CHRO turns to when the board asks where AI is going to land first. The rest will be asked to validate someone else's answer.

Second diagnostic: Who in your organization is currently accountable for producing a structured, queryable description of how HR work actually gets done, not how it is supposed to get done? If the answer is no one, that vacancy is your strategic opening.

What a real readiness posture looks like

A defensible AI readiness posture treats workforce knowledge as a data asset to be captured with the same rigor your function applies to engagement or attrition data. In our ISD (Ikona Systems Diagnostic) engagements, that means 40 to 60 structured interviews across the HR organization, each one building on the last, processed through a five-layer intelligence refinery: raw capture, transcription, editorial cleaning, structured fact extraction, and anonymization for downstream use. The output is a queryable knowledge store the function owns, not a slide deck that ages out in a quarter.

This approach has limits worth naming. Structured interview synthesis captures the work as practitioners describe it, which means it inherits their blind spots. It is most defensible when paired with system telemetry, not as a replacement for it. But for the context layer specifically, the layer your stack has never reached, structured capture is the only method that produces evidence at the fidelity AI requires.

Six months after the initial engagement, clients are still activating new intelligence modules against the same knowledge store, answering questions that were never part of the original scope. That is what a compounding workforce knowledge asset looks like in practice.

If you want to name the gap before someone else does, that conversation starts with a structured AI readiness assessment. We'd welcome it.

Sources

• Nonaka, I. & Takeuchi, H. The Knowledge-Creating Company. Oxford University Press. [Link](https://global.oup.com/