Embedded conversational analytics for the Controllers' Workbench.
A 12-week build to deliver auditable, conversational answers across the 14 KPIs and 37 tables already powering the Workbench.
Contents.
Country controllers need answers in plain language — not another set of cards to navigate.
An embedded, auditable conversational layer that turns the Controllers' Workbench into a decision tool, not just a dashboard.
Three reasons we're the right partner for this build.
Built for controllership analytics
An active engagement with Unilever Finance is already in flight — with deep familiarity for the CWB data model and KPI catalog.
Multi-agent architecture
Built on MaxAI and an enterprise-grade research framework aligned to Anthropic's standards. Every answer ships with full data provenance — citation, query, and visualization.
Production-grade observability
Telemetry plugs into your existing dashboard. Every prompt and model is benchmarked against your test set before promotion.
What controllers — and the business — get back.
Reclaim controller time.
No more hunting through 14 cards and 37 tables to answer one-line questions. Exception triage starts as a conversation, leaving more room for judgment calls.
Audit-ready by default.
Every answer carries a citation back to source row, KPI, and refresh timestamp. Review trails are built into the workflow, not reconstructed after the fact.
Faster month-end posture.
VAR drivers, BSA breaches, and aged vendor advances surface as questions controllers ask — not cards they have to remember to check. Reactive review becomes proactive triage.
Five workstreams,
twelve weeks.
Five workstreams, sequenced for parallel progress.
Five workstreams across foundation, build, and operate phases.
Discovery & Semantic Layer
- Align on KPI logic, anomaly definitions, and Phase 2 scope
- Map fact and dimension tables across the 37-table model
- Document SharePoint inputs feeding LLR, SoD, and Corporate Audit
Custom Skill Development
- Three accounting / FP&A skills tuned to controller workflows
- VAR driver analysis · Balance-sheet review · Exception triage
Multi-Agent Build
- Conversational orchestration across structured (SQL) and unstructured (SharePoint) sources
- Citation-grounded responses with deterministic data lineage
CWB Embed & Identity
- UI integration with the Controllers' Workbench (or initial standalone option)
- Per-user tokens, SSO, Active Directory groups, row-level security
Observability & UAT
- Telemetry feed to Unilever's observability dashboard
- Prompt and model benchmarking · UAT · structured hypercare handoff
What we bring — and what we plug into.
Multi-agent orchestration with a citation engine.
- Pre-built accelerators for KPI semantic mapping, query routing, and document grounding
- Identity & security — per-user tokens, SSO, AD groups, RLS
- Observability + AI benchmarking surfacing utilization, tool-call success, and model fit
Your stack stays your stack.
- Azure SQL Server, Data Factory, Data Lake, and Databricks
- SharePoint inputs for LLR, SoD, and Corporate Audit
- Controllers' Workbench UI as the embed surface
- Existing agent observability dashboard for telemetry consolidation
12-week phased delivery.
Target start: late May 2026. Final dates confirmed at kickoff.
The Workbench, with a conversational layer.
What you have at the end of 12 weeks.
Embedded conversational layer
Live in the Workbench across all 14 Phase-1 KPIs, ready for Phase-2 insights.
Three custom controller skills
VAR drivers, balance-sheet review, and cross-card exception triage.
Identity & row-level security
Per-user tokens, SSO, AD-group provisioning, RLS — built in, not bolted on.
Semantic & lineage layer
KPI definitions and lineage documented across the 37-table model and SharePoint inputs.
Observability feed
Telemetry, prompt and model benchmarks streaming into your existing agent dashboard.
Hypercare runbook
Documented support cadence, escalation paths, and a structured handoff at week 12.
Investment summary.
All figures USD · annual support shown separately.
One-time, 12-week delivery.
Discovery, semantic layer, three custom skills, multi-agent build, CWB embed & identity, observability and UAT.
Annual platform support.
Monitoring, prompt & model benchmarking, observability upkeep. Begins at go-live and renews annually.