What is the difference between a BI dashboard and a predictive analytics dashboard?
A classic BI dashboard shows what has happened — sales history, customer volumes, stock movements. A predictive analytics dashboard looks ahead and forecasts what is coming: demand for next month, churn probability per customer, anomalies in the financial close. Under the hood sits an ML pipeline that continuously retrains on new data, plus a feature store that keeps input variables consistent across models. You set up a BI tool on your data and it works; a predictive platform requires a design phase in which you decide which decisions need to be supported and which data is relevant for them.
Will you replace our current Power BI or Tableau environment?
Not for the standard BI use case — Power BI and Tableau are excellent for analysts and day-to-day reporting. What we do is add a predictive layer alongside your existing BI environment. Often Power BI keeps running for standard reporting and our predictive dashboard is added for forecasting and anomaly use cases. For some clients the predictive platform eventually replaces the classic BI layer as well, but that is not a requirement.
How accurate is a demand forecast or churn prediction, really?
That depends heavily on data quality and the use case. For seasonal retail categories with several years of history, forecast errors of around ten per cent (WMAPE at SKU-category level) are a realistic target; for highly volatile or new products, higher. For churn prediction we look at AUC and precision/recall — good models typically reach an AUC of 0.80–0.90, but the real business value lies in how much of your retention budget lands on the right customers. We always measure against a baseline and only move forward if the model beats the baseline by a measurable margin.
How far along is a natural-language layer with LLMs, really?
Further than most people think, provided you build in the right guardrails. An LLM such as Claude or GPT-4 can reliably turn a question like "what is the expected cash position at the end of Q3 if the top 10 deals go ahead?" into SQL plus a visualisation, provided you supply a good semantic layer that explicitly defines column names, business definitions and joins. We build text-to-SQL with a validator that checks the query against a whitelist of tables, columns and aggregations, plus row-level security that prevents the LLM from retrieving out-of-scope data. For AI Act compliance we log which questions were asked and which SQL followed from them — that audit trail is mandatory for high-risk AI systems.
What data do you need to get started?
As a rule of thumb: at least two years of history for the variable you want to predict, plus as many contextual features as possible. For demand forecasting: sales per day or week per SKU per location, plus pricing, promotions, stock levels, weather and seasonality. For churn prediction: contract data, usage data, support tickets and NPS. In our experience, data quality matters more than volume. During discovery we run a data quality audit and tell you honestly whether both volume and quality are sufficient. Sometimes the conclusion is that you need to clean up your data first.
How do you handle the EU AI Act and GDPR?
First, classification: what risk level does your use case fall under in the EU AI Act? Minimal-risk systems (demand forecasting, anomaly detection on operational data) carry limited obligations. High-risk systems (recruitment, credit scoring, health advice) require strict documentation, an audit trail, human oversight and sometimes a conformity assessment. For GDPR: a DPIA for every model that touches personal data, PII redaction in the feature layer wherever possible, and EU residency for both data and models. For clients in regulated sectors, we work with European cloud tenants as standard.
Can this be integrated with our existing data stack?
Almost always, and that is usually the point. We build connectors to warehouses (Snowflake, BigQuery, Synapse, Redshift, ClickHouse, Databricks), operational databases (Postgres, MySQL, SQL Server), CRM systems (HubSpot, Salesforce), ERP systems and specific APIs. For ETL we use dbt, Airflow or Fivetran, depending on your existing set-up. If you don't have a warehouse yet, we can build one too, usually as a combined
data warehouse project, so the predictive platform rests on a reliable data foundation.
What is the difference from a KPI dashboard?
A
custom KPI dashboard shows how your business is currently performing against agreed goals: KPIs, target versus actual, trends per department. A predictive dashboard forecasts where those KPIs are heading and shows the drivers behind that forecast. In practice the two overlap: many clients start with a KPI dashboard and, after a few months, want a predictive layer on top, for example an end-of-year forecast alongside the actuals, or a drill-down that predicts which department is likely to miss its target. We also regularly build both in a single project when the audience and data source largely coincide.
How often do models need to be retrained?
It depends on data velocity and the use case. For demand forecasting in retail we often work with weekly retraining; for churn prediction in subscription businesses, monthly is usually enough; for anomaly detection on fast streams, daily or even continuous retraining may be needed. There is no manual work involved: the orchestration layer runs on a schedule and MLflow logs every run reproducibly. Drift detection raises an alert when a model starts to deteriorate. Set-and-forget is not an option: data changes, behaviour shifts, and without periodic retraining every model eventually degrades.
What if the model makes mistakes?
Expect it to happen, and build for it: a confidence margin for each prediction, a SHAP explanation for each prediction, human-in-the-loop for high-impact actions, an audit trail of who decided what based on which prediction, and a rollback procedure to an earlier model version. Models without an error procedure are not production-ready. With ML this is more urgent than with traditional software, because the nature of the errors is less predictable.
Do you work with our data team?
Almost always. For some clients we deliver the entire platform, including maintenance; for others we work alongside an in-house data team that takes over the models. We provide knowledge transfer in the final sprints, a runbook for incidents, and clear responsibilities for monitoring, retraining and further development. You always own the code, so there is no vendor lock-in.
What determines the cost of a predictive build project?
Three factors weigh the heaviest. One: the number of models and their complexity — a single demand forecast is a fundamentally different project from a multi-model platform with churn, forecasting and anomaly detection. Two: data maturity — is the warehouse already in place, or does a data engineering track need to run alongside? Three: the compliance and explainability layer — high-risk AI Act use cases require extra documentation and human oversight. We work with a fixed sprint budget so you can adjust the scope every sprint.
How long until we can go live?
For a first working model on your existing data stack, the foundation is in place after a few sprints. A full multi-model platform with a feature store, MLOps and multiple dashboard views takes several sprints on top of that. We work in two-week sprints with a working build every sprint, so you see early on where it's heading and can adjust along the way. A phased approach is also possible: first one model in production, then expanding to further use cases on the same infrastructure.