Enterprise AI Dashboard for KPI Forecasting
The client was a mid-market enterprise SaaS vendor whose executive team relied on a manually-maintained weekly KPI deck across sales, churn, and demand. The deck took roughly thirty hours per week to produce across finance, RevOps, and customer success, and by the time it landed in the executive meeting it represented a backward-looking view that was already a week stale. Strategic decisions were being made on lagging indicators because no forward view existed at executive cadence.
Trading, FinTech & Analytics
Enterprise SaaS
12 weeks from kickoff to executive cutover
5 specialists
The full story
The practical problem was that the data the team needed was spread across Salesforce, Stripe, HubSpot, the product analytics tool, and the warehouse, with no consistent reconciliation across them. Each team owned a slice of the truth, and the weekly deck spent most of its production time arguing about whose number was correct rather than reasoning about what the numbers meant. Forecasting did not exist as a function — the closest substitute was finance’s quarterly forecast, which used annual seasonality and missed within-quarter signals entirely.
We built an executive KPI dashboard that pulled directly from the source systems into a unified warehouse, reconciled disagreements via canonical join keys, generated forecasts with confidence bands on the metrics the team cared about, and exposed driver decomposition so the executive could see why a forecast was moving. Anomaly detection ran continuously and surfaced events that required attention before the weekly meeting.
What shipped was a live executive dashboard with twelve-month forecasts on sales, churn, and demand, driver decomposition per forecast, and an alert layer that flagged anomalies in real time. The weekly deck production effort collapsed substantially, the executive meeting shifted from arguing about numbers to acting on them, and the forecast accuracy made forward-looking decisions defensible at board level.
Executives ran the business on a backward-looking deck that took thirty hours per week to produce and was stale on arrival.
Weekly KPI deck production consumed thirty hours across finance, RevOps, and CS, with most time spent reconciling source disagreements.
No forward-looking forecast existed at executive cadence — the closest was a quarterly finance forecast using annual seasonality only.
Source systems disagreed on basic numbers because of inconsistent join keys and timing differences across Salesforce, Stripe, HubSpot, and product analytics.
Anomaly detection was manual and lagged real events by a week, so the executive team learned about issues only at the next deck cycle.
Forecast confidence bands and driver decomposition did not exist anywhere, so strategic decisions lacked defensibility at board level.
How we structured the engagement
Reconciled source data first, then layered forecasting on a single source of truth so the executives stopped debating which number was right.
- 01Phase 01Weeks 1-2
Discovery
Mapped the executive deck contents to source systems and identified every reconciliation seam between Salesforce, Stripe, HubSpot, product analytics, and the warehouse. Worked with finance and RevOps on canonical join keys and timing conventions. Output: a unified KPI schema and a per-source reconciliation specification.
- 02Phase 02Weeks 3-4
Architecture
Designed a warehouse-centric data pipeline with daily syncs from source systems, canonical join keys enforced at ingestion, and modeled metrics defined once. Picked Prophet for the forecasting layer because it handled seasonality, holidays, and changepoints well on this data shape. Built driver decomposition on top using SHAP-style attribution.
- 03Phase 03Weeks 5-10
Build
Shipped the source syncs and the reconciliation layer first because every downstream metric depended on it. Built the forecasting layer next with twelve-month horizons and confidence bands per metric. Implemented driver decomposition and the anomaly detector. Built the executive dashboard last because it was a render of work done elsewhere.
- 04Phase 04Weeks 11-12
Launch
Rolled out to the executive team for two weeks of parallel review against the manual deck, with finance and RevOps validating every forecast and decomposition. Tuned the anomaly detector against the first month of live alerts to suppress noise. Promoted to primary surface after the executive team chose the dashboard over the deck in the third weekly meeting.
What we built, component by component
- 01
Source syncs
Daily pulls from Salesforce, Stripe, HubSpot, product analytics, and the warehouse with structured logging and reconciliation.
- 02
Reconciliation layer
Enforces canonical join keys and timing conventions across sources so every downstream metric resolves to a single truth.
- 03
Metrics model
Defines each KPI once over reconciled data with provenance per number, used by both the dashboard and the forecast layer.
- 04
Forecasting layer
Prophet-based forecasts with twelve-month horizons, confidence bands, and seasonality and changepoint handling per metric.
- 05
Driver decomposition
SHAP-style attribution that explains why a forecast is moving in terms of underlying drivers the executives recognize.
- 06
Anomaly detector
Continuous detector on reconciled metrics that surfaces real events to the executive surface in real time, not weekly.
Source systems sync daily into the warehouse, the reconciliation layer enforces canonical keys, and the metrics model produces a single truth per KPI. The forecasting layer projects twelve months forward with confidence bands, driver decomposition explains forecast movement, and the anomaly detector watches reconciled metrics continuously, surfacing events directly to the executive dashboard between weekly cycles.
The trade-offs we made and why
Solved reconciliation before any forecasting
Forecasting on disagreeing source data produces disagreeing forecasts that no one trusts. Spending the first two weeks on canonical keys and reconciliation paid back every week thereafter, because every metric, every forecast, and every alert sat on a single source of truth.
Used Prophet over a deep model
The KPIs had strong seasonality and identifiable changepoints, which Prophet handled cleanly with interpretable output. A deep model would have required more data, more training overhead, and produced less interpretable forecasts that the executive team would have trusted less.
Built driver decomposition into every forecast
A forecast without explanation is a number to argue about. Driver decomposition gave the executives a way to act on forecasts in terms of underlying business drivers they recognized, which converted forecasts from prediction into decision support.
Made the dashboard the primary surface, not the deck
Keeping the deck alive in parallel would have split executive attention and kept the thirty-hour production cycle running. Cutting the deck after the executive team chose the dashboard forced full adoption and unlocked the time savings the project was sold on.
What changed for the client
forecast horizon
Forward-looking horizon on sales, churn, and demand forecasts with confidence bands per metric and per scenario.
weekly deck time
Reduction in executive deck production time across finance, RevOps, and customer success after dashboard cutover.
decision accuracy
Self-reported lift in confidence and accuracy on strategic decisions made off the dashboard versus the prior backward-looking deck.
anomaly alerts
Replacement for the prior weekly discovery cycle, with anomalies surfaced as they occur on reconciled metrics.
The tools behind the system
Built with a deliberate stack chosen for production reliability and operational velocity.
Lessons learned from the build
Reconciliation was the unsung hero of the project. Without it, every other component would have produced numbers people argued about. Spending the first two weeks on canonical keys before writing any forecast code was the highest-leverage time investment we made.
Driver decomposition turned forecasts into actions. A confidence-banded prediction is interesting; a confidence-banded prediction with attributable drivers is actionable. We would treat decomposition as a first-class feature on any future executive forecasting work, not as a polish item.
Killing the parallel deck was a change-management decision more than a technical one. The deck could have lived forever as a comfort artifact, and cutting it forced the executive team to commit to the dashboard. We would plan that decision into the rollout schedule, not leave it to chance.
Similar delivery work usually starts in these service areas
If you are exploring a similar product, workflow, or implementation challenge, these are the service tracks that usually fit best.
Where this project sits in the bigger market picture
Patterns for AI features, internal tooling, and product delivery in SaaS businesses.
Build a result-driven AI product with a team that has shipped before
If you are exploring a similar product, workflow, or AI use case, we can help scope the right architecture, delivery model, and first milestone.
Related case studies worth reviewing next
Have an AI idea, messy workflow, or product vision? Let's make it buildable.
Bring the problem. We'll help shape the product, define the architecture, and show the fastest path to a serious first version.
A practical first roadmap in the discovery call
Architecture, timeline, and delivery options in plain English
Security, scalability, and reliability discussed upfront
Model registry
softus-rag-v4.2
187ms
Latency
128k
Context
$0.004
Cost / req
Evaluation suite
Deploy pipeline
prod / canary 25% — healthy