AI value baseline: capture it before agents go live

Why the AI proof burden landed on people analytics

Gartner's CHRO priorities release on October 2, 2025 framed the 2026 agenda around realizing AI value and driving performance amid uncertainty. That phrasing is the tell. Realizing value is not a procurement task or an IT milestone; it is a measurement claim, and measurement claims about work and workers route to people analytics by default. A companion regional release on 2026 HR priority trends, issued in Australia in November 2025, carried the same theme, which suggests this is a global reframing rather than a US budget-cycle artifact.

Josh Bersin made a related point in "The Changing Role of Data and People Analytics in HR," published roughly three weeks ago: the analytics function is being redefined, not simply enlarged. The distinction matters operationally. An enlarged function gets more dashboards. A redefined function gets asked to arbitrate whether an enterprise investment worked, which is a different standard of evidence and a different relationship with finance.

The uncomfortable part is that the evidence standard arrives before the evidence does. Budget is moving first. S&P Global Market Intelligence's February 2026 HR technology forecast names employee experience, people analytics and talent intelligence among the drivers of market growth, meaning spend is accelerating ahead of the governance that would make the spend defensible.

The comparison period is not a report you can run later. If it was never instrumented, it is simply gone.

The comparison period is a perishable asset

Most HR metrics can be reconstructed after the fact. Time to fill, case volume and cost per hire can usually be pulled from history whenever someone asks. That reconstructability is exactly what makes teams complacent, and it quietly fails the moment AI enters the process, because what changes is not only the number but the definition underneath it.

When an agent triages tier-one employee relations cases, does an auto-resolved inquiry count as a case? When a sourcing copilot drafts outreach, does the requisition's clock start at approval or at first human review? Teams that answer these questions after deployment will answer them in a context where a favorable answer is also a convenient one. That is not fraud, it is drift, and it is enough to disqualify the analysis in front of a skeptical finance partner.

SHRM's reporting that the people analytics software market is maturing and being reshaped by AI compounds the risk. Consolidating vendors means metric definitions and data pipelines are in motion at the same time the workflows are. If you migrate platforms mid-measurement without a frozen baseline, you will never be able to separate the effect of the AI from the effect of the new system of record.

What a 30-day baseline capture actually contains

Keep the scope deliberately narrow. Pick three to five process metrics that the AI is explicitly supposed to move, and write down why each one was chosen. Useful candidates: cycle time for a defined HR service, cost per unit of that service, span supported per HR business partner or manager, time to fill for a named job family, and case resolution time. If a metric cannot be tied to a specific promise made during the business case, leave it out. A short list that survives scrutiny beats a wide one that collapses under it.

For each metric, record four things: the written definition including inclusions and exclusions, the current value with a timestamp, a named data owner who is accountable for the figure, and the source system the number came from. A definition that lives only in someone's head is not a baseline. Put it in a document with a version number and circulate it to finance before deployment, not after, so that the terms of the eventual argument are agreed while nobody has a stake in the outcome.

Then record two context variables that teams routinely forget: headcount and total hours in the affected function as of the baseline date. Organizations restructure constantly. If you do not timestamp the org as it stood, every subsequent efficiency gain will be credited to the AI by default, including the ones produced by a hiring freeze, an offshoring decision or a reorganization that was already in motion.

Governance that makes the baseline hold up

Treat the baseline document as a controlled artifact. Changes to a metric definition after the capture date should require a logged amendment with a reason, not a silent edit to a spreadsheet. If a definition genuinely has to change, keep both series running so the original comparison stays intact. This is standard practice in finance reporting and it is the habit people analytics needs to borrow as it moves closer to the investment-justification table.

Decide in advance what the measurement window will be and what counts as a result. A 90-day post-deployment read on a process with seasonal volume will tell you about the season, not the software. Agreeing the window before you know the answer is what separates an analysis from an advocacy exercise. AIHR's 2026 workforce analytics trends roundup is a reasonable secondary frame for teams building this out, but the discipline is simpler than any framework: say what you will measure, say when, say who owns it, then do not move the goalposts.

The alternative is the position many HR functions will find themselves in by late 2026. Asked to justify an AI investment, with no instrumented comparison period, they will have to reach for vendor-supplied impact claims. Borrowing a supplier's numbers to defend your own spend is a credibility loss HR cannot afford to take twice.