People analytics and review-cycle AI: what stuck, what stayed partial
Three applied workstreams in a People division's AI programme. Staff learned to reach HR data through APIs themselves; the ATS data layer stayed partial and an earlier executive calendar stayed a design.
Context
A listed payments company with more than a thousand employees across many markets asked us to help its People (HR) division work with AI: about thirty professionals, most without a technical background. Assessments, one-to-one sessions and role-based workshops were followed by applied workstreams over the following year.
This case covers three of those workstreams: people analytics and executive reporting, the performance-review cycle, and talent acquisition. Their results are uneven, which is why we publish them together. One condition shaped all three: build capability inside the team, on the tools it already had (Google Workspace with Gemini and Apps Script).
The challenge
The assessments found four recurring frictions:
- Reporting assembled by hand, from HRIS downloads to slides. The monthly executive People report reached leadership around mid-month, when part of it was no longer news.
- A twice-yearly engagement survey needing close to a hundred reports, produced by several people copying figures into each one.
- A quarterly review cycle with about a thousand written feedback records to assess. An existing agent handled one record at a time; full-set attempts stalled after a few rows.
- Recruiters reformatting interview notes across tools, and a sourcing set-up more complex than the hiring volume justified.
The team was not starting from zero: an analyst already maintained BI dashboards and a daily HRIS feed. The gap was that non-engineers could not reach data through APIs themselves.
Our approach
- 1Coach, do not build forOn a real task of their own, the practitioner wrote and debugged Apps Script with Gemini as technical coach; we explained the logic.
- 2Package what repeatsRecurring patterns became reusable agents with their own instructions and reference material.
- 3Simplify before automatingWhere a workflow was more complex than its volume justified, we recommended removing steps before adding AI.
- 4Revisit with evidenceAt twelve months we checked every item against records and labelled it working, partial, artefact only or design.
What was built and designed
People analytics and executive reporting
The analytics lead described a working set-up in which a selected HRIS report lands in Google Sheets through the API on a scheduled refresh. A People operations specialist built a direct HRIS API workflow with our coaching and confirmed it met its goal, about five weeks after the first workshop. In one working session with the lead we built and tested, on a real request, a technical-coach agent that returns complete Apps Script code with a beginner's step-by-step guide.
The analytics lead later confirmed that an extraction of quarterly review data from the performance platform returned every required field; a second dataset on the same platform still needed its own script. The ATS integration stayed partial: offer data was retrieved, but missing fields, a much larger candidate dataset and validation with recruiters were unresolved in our last records.
With a People business partner we mapped the route from source data to leadership review and designed an earlier calendar: data on day one, internal review around day four, executive discussion around day five. We have no record that it was implemented.
Performance-review cycle
The feedback analysis moved into Google Sheets, where the AI function applies the review framework to each record and criterion, for example separating specific, actionable feedback from ambiguous comments. We coached the specialist on the rubric and shared a formula-builder agent that turns an analytical requirement into a formula. The specialist reported the formula working; inconsistent output formats, occasional model refusals and manual triggering remained open.
The review-cycle agents were the client's own; we advised on coverage and proposed testing whether they could be consolidated instead of multiplied. One session coached a specialist producing explainer videos for the review process. We hold no completion record for either.
Talent acquisition
In an early session a recruiter built, with us, an agent that turns interview notes plus an AI meeting summary into feedback in the required structure. It was tested in the session on real interview material: a working example, not evidence of team-wide adoption or of time saved.
A later review found the multi-bot sourcing flow too complex for the hiring volume. We delivered three quick wins instead: a CV-generation agent, a form-filling prompt for a recruiting tool and a recorded walkthrough. One of the three was soon superseded by a built-in AI feature from a vendor.
Results
Target of the calendar we designed; assumes HRIS data lands automatically on day one. Designed and coached; no record of implementation.
One participant's account about a year in; previously monthly or quarterly. Built by the client's team, not by us.
The client planned to test the other two items; we hold no adoption record.
What stuck. Reaching data through APIs became a skill that people in the team hold, not a deliverable of ours. The analytics lead reported teaching a talent-acquisition colleague to run an API report of their own, in minutes by that account. That is one reported example, not a measure. The lead later worked through API documentation, permissions and a vendor support request independently. The client also runs its own internal AI programme, so later capability cannot be attributed to us alone.
The strongest reported result is the client's own. The analytics lead rebuilt survey reporting with formulas and linked slide templates, no agent involved. The lead estimated a move from roughly two weeks of several people to three or four days of one person on the first run, with manual steps remaining. Our part was the programme around it. We have not converted the estimate into a percentage or an annual saving.
What stayed partial. The ATS data layer is incomplete and the day-five calendar remains a design. A full run of the feedback analysis and adoption of the recruiting artefacts are unproven. An early status update described HRIS and ATS reporting as implemented; later records narrowed that, and our twelve-month evidence audit flagged overstatements in our own outcomes summary.
Governance and risk
These workstreams were not run as outsourced development with direct access to the client's technology stack. Practitioners built and ran sheets, scripts and agents in the client's own workspace, so each workflow stayed with the person accountable for the process. For a calibration-records idea that never went beyond assessment, we asked for column structure and dummy rows, not real records.
People data is sensitive. The feedback analysis assessed the quality of managers' written comments and their consistency with ratings, and it was built and run by the specialist who owns that process. Model refusals and inconsistent formats were reasons to keep a person in the loop, not defects to hide.
What we learned
- A model used as technical coach lets non-engineers reach data through APIs. In one reported case the skill then passed between colleagues without us.
- Expect mixed maturity from enablement work and label it: working, partial, artefact only, design. The labels are what make the working items credible.
- Agree a baseline and a close-out measure at kick-off. Without them, a participant's estimate stays an estimate and should be published as one.
Client identity, locations and identifying details are withheld under confidentiality. Figures are rounded. Measured figures come from engagement records; client-reported figures are attributed, not audited; modelled figures are projections from the engagement's business case and are labelled as such.