An agent that writes the first draft

The situation

The analytics team I lead supports brand, sales, and marketing leaders — and more than 2,000 field users — across a multi-brand pharma portfolio. Every week, that support produced a stream of recurring insight documents: narrative summaries with KPI tables, cut by brand and geography, assembled by hand from dashboards into documents.

Analysts were spending large parts of their week transcribing numbers they had already computed into documents nobody enjoyed writing. The work was slow, error-prone, and — worst of all — it consumed exactly the people whose judgment the organization actually needed.

The problem underneath

Document automation fails in enterprises for two reasons: the output can't be trusted (models improvise numbers) or the tool can't be approved (new vendors, new data paths, new risk reviews). Any solution had to clear both bars in a regulated industry.

The insight I started from: the hard part isn't generating text — it's guaranteeing structure and data fidelity. So I designed the system around those guarantees.

Decisions that shaped the work

Build inside the Microsoft estate. Copilot Studio and Power Automate run within the organization's existing compliance perimeter — no new vendor, no new data egress, dramatically faster IT approval. The trade-off is less model flexibility than a custom LLM stack; I took that trade, because the bottleneck was trust, not model quality.

Structure the document; don't freestyle it. The output templates carry 58 content controls — named slots the workflow must fill. The agent drafts narrative into a defined structure instead of inventing one.

The agent never invents a number. Every KPI table is bound to the governed semantic layer and injected by the orchestration flow, not generated by the model. Language is generated; numbers are retrieved. That single design rule is what made adoption possible in a compliance-first culture.

Humans approve, not assemble. Analysts review and sign off. The system changed their role from typist to editor — which is the correct direction for AI in analytics work.

How it works

A requester describes what they need conversationally. The Copilot Studio agent interprets the ask, Power Automate orchestrates retrieval against the semantic layer, and the workflow assembles a structured document — dynamic KPI tables, program context, and drafted narrative — routed to an analyst for review. Minutes of review replaced hours of assembly.

What moved

Across the portfolio's recurring insight documents, roughly 90% of manual drafting time disappeared — about $180K a year of analyst time, measured against the hours the documents took to assemble before, redirected from transcription to interpretation. Cycle time for recurring insight documents dropped from days to same-day. And because the numbers come from the governed layer, consistency across documents stopped being a review problem at all.

Alongside it, I deployed Power BI Copilot over the same governed semantic layer, letting field stakeholders ask territory questions in natural language — cutting rep-to-insight time by about 60% with no technical training.

Lessons

The semantic layer was 70% of the work. The agent is only as trustworthy as the definitions underneath it. Organizations that skip this step get demos; organizations that do it get adoption.

Guardrails are the feature. Content controls and retrieval-bound numbers are what let a regulated business say yes. In enterprise AI, constrained is what buyers can approve.

Automate toil, return judgment. The measure of success wasn't documents per hour. It was what analysts did with the week they got back.

Sources: resume master jun 2026 · build documentation