An applied-AI sprint for a People function, set against its plan
A listed payments company piloted role-based AI training in its People function. The client confirmed its first working workflows about five weeks after the first workshop. Part of the plan has no delivery record.
Context
The client is a listed payments company with more than a thousand employees across many markets. The engagement began with its People function: around 30 professionals across recruitment, business partnering, operations, communications, rewards and analytics. It works with every division, which made it a sensible pilot for a wider programme.
Most of the group had no technical background. The company already ran on Google Workspace with Gemini available next to the usual HR systems, so access to AI was not the constraint. The programme was designed under one condition: leave capability installed in the team, on the tools it already used.
The challenge
Reports were assembled by hand, consolidations were copied between spreadsheets, and employee questions were answered one at a time in chat. The open question was not whether AI could help, but where, for whom and in what order.
Generic training does not answer that. A recruiter, a payroll specialist and a divisional head need different things from AI. We had to understand some thirty individual jobs before writing a curriculum, and do it in about two weeks, because our proposal had promised early results.
Our approach
We designed the engagement as a sprint with a fixed order: understand each person's work, personalise, teach by role, apply between sessions. Two paths were planned in parallel: a personal build path, mostly asynchronous and one-to-one, and a collaborative path of live workshops.
- 1Agentic one-to-one assessmentDesigned as a private, hour-long conversation between each participant and an AI interviewer about real workflows, tools and frustrations.
- 2Personal report and deep diveEach assessment was meant to yield a personal report, reviewed one-to-one with a practitioner to agree a first use case.
- 3Use-case prioritisationOpportunities from the conversations were ranked by gain against complexity; high-gain, low-complexity cases shaped the curriculum.
- 4Role-based tracksLive group workshops, planned at two hours, for team leads and for hands-on specialists; a weekly one-to-one track for the divisional leader.
- 5Applied learning weeksWorkshop weeks alternated with applied weeks, when participants worked on their own cases with email and working-session support.
The framework defines four tracks (executive committee, divisional leader, team leads, hands-on specialists) because governing AI, planning a division's use of it, orchestrating work between people and agents, and re-engineering a daily task are different skills. Three ran in this pilot; the executive track was designed and not run.
What was built and designed
Kick-off communications. We drafted three kick-off messages, one per track, for the client's programme coordinator to send. Each introduced a primary contact, booked the assessment and the one-to-one, and linked to a short site explaining that track.
Touchpoints. Each participant's journey was specified as eleven touchpoints. Six belonged to the sprint: assessment, kick-off, one-to-one deep dive, email support, live workshops and a resource repository. Five came afterwards: an impact report, a community hub, monthly briefings, an open monthly workshop and a pulse survey.
Curriculum. The method has four phases: workflow analysis from the assessment conversations, use-case prioritisation, role-specific design, and content updates as the tools change. Our synthesis placed the opportunities in recruitment friction (data entry, interview scheduling, feedback write-ups), analysis of unstructured text, and communications at scale. Sessions worked on participants' own cases; one module covers the AI formula in Google Sheets for bulk data transformation.
Reusable builders. Alongside task-specific agents we handed over agents that build agents: an agent creator, a spreadsheet-formula architect, a script-automation architect with its own knowledge base, and a technical coach for People analytics. The shared folder holds more than ten such artefacts, so a participant can extend a workflow without coming back to us.
Results
Transcript files held; not a clean count of unique completions.
Calendar entries and session notes; a few duplicates exist.
Eight group workshops, seven leader one-to-ones; two leader sessions lack a substantive transcript.
The six sprint touchpoints. The five post-sprint ones have none; one later calendar entry may be a briefing, unverified.
Confirmed in writing for defined scopes. Not a census of production systems.
The sprint ran in the designed order, over roughly twelve weeks from kick-off to the close of the tracks. About five weeks after the first workshop the client confirmed two working workflows in writing: a document-and-email automation for event letters, and a time-off workflow on the HRIS API, built by a People operations specialist with our coaching. A third, a reusable package of Gemini agents, was confirmed shortly after the tracks closed. The scopes were narrow, and these confirmations are not evidence of production use at scale.
We cannot show that the plan was delivered in full. We hold no delivery record for the five post-sprint touchpoints, the closing re-assessment or the planned participant certificate. The design promised one custom agent per participant; many agents and prototypes exist, but no per-person count, and we do not claim one.
When the formal tracks closed, the collaboration continued as applied work with individual teams. We found no formal close-out and no accepted KPI report. No measured before-and-after of skills, hours or error rates exists, so this case reports none. The company-wide second phase remains a proposal.
Governance and risk
Privacy of the assessment. It was presented to participants as a private diagnostic. Its outputs fed their own one-to-one and, in synthesised form, the curriculum.
Client-controlled systems. We did not operate as an outsourced development service with direct access to the client's technology stack. Builds lived in the client's own Workspace environment. Where a workflow touched HR data through an API, the specialist who owned that process built it, with our coaching. Approvals stayed with the people already accountable for them.
Contracting. The engagement went through the client's supplier onboarding. Our file holds the confidentiality and data-processing agreements and the due-diligence paperwork, including the client's anti-corruption policy.
Evidence and attribution. Calendar entries prove scheduling, not attendance, so counts are matched to session notes and rounded. We separate what we built, what participants built with coaching, and what the team built later alone. The company also ran an internal AI initiative, so later capability cannot be credited to this sprint alone.
What we learned
- Interview before you teach. A private agentic assessment per person tied the curriculum to real workflows and gave the one-to-ones a concrete starting point.
- Split by role, not by tool. Leaders, team leads and specialists need different sessions, cadence and examples; one mixed classroom serves none of them well.
- Put the closing measurement inside the sprint. Ours sat after it, we hold no record of it, and so there is no measured before-and-after to report.
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.