What is explainable AI?
Some models are easy to read: a linear model shows a coefficient for every input. Many useful models aren't, including gradient-boosted trees, neural networks, and language models. Explainable AI is the toolkit for opening them up, so a person can see which inputs pushed a result up or down.
There are two levels. Global explanations describe what drives the model overall, such as which channels matter most to the forecast. Local explanations describe one result, such as why this week's forecast dropped.
What are the main explainable AI methods?
| Method | What it shows | Good for |
|---|---|---|
| Feature importance | Which inputs the model relies on most | A quick overall ranking |
| SHAP values | Each input's contribution to one prediction, based on Shapley values from game theory | Explaining a single result, and summing to an overall view |
| Partial dependence | How the output changes as one input changes, holding the rest steady | Seeing a response curve, such as spend vs. sales |
| LIME | A simple model that mimics the complex one near a single prediction | Quick local explanations |
| Model cards | A written summary of what a model is for, how it was tested, and its limits | Governance and handoffs |
What is an example of explainable AI?
Why the forecast dropped
A sales forecast for next month falls 8%. The team needs to know why before the planning meeting.
A SHAP breakdown shows most of the drop comes from a seasonal pattern and a planned pause in TV, with a small effect from a price increase. The forecast isn't broken. It's reflecting decisions already made, and the team can explain it in one slide.
Interpretable models vs. explained models
| Interpretable by design | Complex model plus explanations | |
|---|---|---|
| Examples | Linear regression, small decision trees | Gradient boosting, neural networks |
| Accuracy | Often lower on complex patterns | Often higher |
| Explanation | Built in and exact | Approximate, from a separate method |
| Best when | Decisions need a simple, auditable rule | Accuracy matters and explanations can be checked |
Why does explainable AI matter?
A model's recommendation only moves budget if someone can defend it. Explanations are how a forecast or a ranking gets past finance, legal, and the person whose channel is being cut.
- Catch mistakes. An explanation that leans on a strange input, like a tracking glitch, exposes a problem before it costs money.
- Build trust. People act on numbers they can follow.
- Know the limit. Explanations show what drives the model. Whether that driver causes sales in the real world still needs a test.
How does Alli do this?
Alli by PMG's Scenario Planner gives modeled estimates with the assumptions in view, so teams can explain a forecast to finance. Every scenario splits into moves to make and what they're worth, with channels to boost sorted by the revenue they're expected to add and channels to trim showing what each cut costs.
Across Alli, answers come with their workings. Ask Alli shows the data, logic, and source behind each answer, and Alli Creative Insights shows which creative attributes go with stronger performance, attribute by attribute.
Explore Alli Scenario PlannerExplainable AI FAQ
What are SHAP values?
SHAP values split a model's prediction into the contribution of each input, using Shapley values from game theory. Positive values pushed the prediction up, and negative values pushed it down.
What is the difference between explainability and interpretability?
They're often used interchangeably. Interpretability usually means a model is simple enough to read directly. Explainability usually means using methods to explain a complex model.
Does explainable AI show causation?
No. It shows what drives the model's output. A strong driver in a model may reflect correlation in the data, so causal claims still need tests.
What is a model card?
A short document that describes a model's purpose, training data, performance, and known limits, so people using it understand where it can and can't be trusted.
Last updated Back to the glossary