The problem
Dentsu Creative is in the middle of a real strategic shift: moving from a model built on single-discipline specialists toward an AI-accelerated hybrid talent model. Making that shift responsibly meant leadership needed a clear, data-driven picture of where the organization actually stood across a defined set of capability dimensions, not a hunch.
Without a structured collection method, talent data tends to be inconsistent, surface-level, and shaped by whichever manager happens to be filling out the form. A static survey could have gathered the same fields, but it wouldn’t have done anything to ensure quality, consistency, or thoughtful reflection from the managers completing it. The real question was whether that bar could be raised without a development team, a budget ask, or months of IT delivery.
What I built
I designed and built a conversational assessment agent inside Microsoft Copilot Studio that guides managers through a structured, 10-question talent assessment of their direct reports. It collects consistent, scored responses across eight capability dimensions and writes the output directly to a SharePoint-hosted Excel workbook via Power Automate, feeding downstream 9-box placement and talent-grouping decisions without a separate manual data entry step.
The eight dimensions aren’t invented. They’re anchored in published external research on future workforce capability, split evenly across current performance (craft excellence, reliability and ownership, collaboration and influence, leadership behaviors) and future potential (strategic capability, growth mindset, hybrid/AI capability, openness to change). Grounding the framework outside of internal HR opinion mattered for credibility with skeptical stakeholders and senior leadership.
The agent uses structured adaptive cards instead of open text to present 1–5 scored questions, which stops managers from skipping the harder dimensions and cuts down on the variance that comes from freeform input. A confirm-before-commit flow lets managers review every response before it’s written to Excel, with a built-in edit loop so they can correct something without restarting the whole assessment.
A couple of design decisions mattered more than they look on the surface. Early prompts leaned into playful, evaluative-sounding language that read as judgment rather than development, so I rewrote the framing around a compass metaphor, directional and developmental, to lower manager defensiveness and signal that the tool exists to help people grow rather than rank who’s expendable. I also ran two phases of testing before rollout: internal HR testing against hypothetical scenarios first, then a supervised pilot with real managers and real employees, with that feedback shaping tone, length, and question clarity before the agent went into broader use.
What happened
The agent is live: managers complete the assessment conversationally, and results land directly in the Excel workbook that feeds 9-box placement, with no one manually transcribing scores. I haven’t tracked a formal completion rate or time-saved figure against the old process. The real proof point is that a strategic initiative that would normally need a developer, a budget line, and months of IT backlog got shipped as something people actually use, inside the existing Microsoft 365 stack, with no engineering support.
What I’d do differently
If I built this again, I’d lean toward a structured form instead of a conversational agent. The conversation format is good at getting managers to slow down and actually think about each dimension, but an agent is worse than a form at holding a long list of distinct inputs consistently over time and writing them cleanly into Excel. A form enforces that structure by default; an agent has to be engineered to fake it.