Blog

Three signs your Excel forecast has hit its limit and you need to build a new model

The limit of the spreadsheet shows up in the infrastructure around the forecast: one person holding it together, series that don't scale without rewriting, and manual decisions accumulating without traceability.

7 min · Data Science

Why so many AI agent projects don't survive the move from demo to production

Current tools allow one person to build an autonomous agent in an afternoon without much prior knowledge. Taking it to production requires architectural decisions and engineering discipline that the demo never considers.

7 min · AI & Automation

Why do production models degrade?

The assumptions behind training age and the model's signal becomes noisy long before the prediction becomes wrong. Detecting that erosion and addressing the problem requires planning and a proactive detection system.

9 min · Data Science

A model that performs well is not the same as a solution that works.

Training a model is a technical problem. Building a business solution with it is a different problem, and the metric to optimise is chosen mostly by the second, not the first.

6 min · Data Science

Before training a churn model, there is a business question that determines the metric to optimise

The model delivers a probability. The metric that decides how to use it depends on the relative cost between letting a customer leave and spending a campaign on a false alarm — and that cost does not live in the technical pipeline.

8 min · Data Science

Dashboard or model: the temporal question that decides where to invest

BI and Data Science answer questions that point in opposite temporal directions. Confusing them is the most common reason data projects solve well what nobody asked.

6 min · Data Science

Machine learning is AI, but the market stopped calling it that

Classical machine learning and generative AI solve different problems. Choosing one or the other depends on the type of signal the system has to process — though sometimes the choice is made based on whichever word is more fashionable.

6 min · Data Science

Treating recurring and one-off customers as the same population can ruin your model

Mixing two populations with different dynamics does not produce a more general model. It produces a model that optimises for one and is blind to the other.

6 min · Data Science

From lab to production: three conditions agreed before training.

80% of models that never reach production do not fail for a technical reason. There are three conditions associated with success, and all three are agreed before the first model is trained.

5 min · Data Science

What happens when the prediction arrives but it is already too late to act?

The prediction horizon has two limits: the floor is set by the customer's process, the ceiling by the stability of the model's signal. The useful window is the intersection of the two.

6 min · Data Science

When the multi-agent architecture complicates more than it helps

A multi-agent system has a specific cost. That cost is only justified when the work has multiple phases with contradictory criteria, or requires unifying different criteria to produce the final output. Without that tension, a single agent can produce better results with less complexity.

6 min · AI & Automation

Automating with AI does not remove the human, it moves them.

The human does not disappear when automation shifts from rules to AI. Their task changes — from resolving exceptions to verifying what the system deemed resolved. Ignoring that shift leaves the automated process more expensive to operate than the manual one.

5 min · AI & Automation