How do AI forecasts actually work?
Many companies now use AI to analyse customer service and spot trends. But when an AI suddenly gives you a very clear answer – “Change this and your FCR will increase by 12%” – it can trigger equal parts excitement and scepticism. How can it be that sure?
The honest answer is: it isn’t. The AI isn’t “sure” – it’s calculating. AI is as “clever” as a hammer. It only does exactly what it has been instructed to do, based on exactly the data it has been given.
Used properly, that makes it a very powerful tool for anyone who wants to make genuinely data-driven decisions.
In this article we explain how TellMeNow works with different types of forecasting models. We show why human instructions, data quality and traceability back to underlying insights are absolutely critical if you want forecasts that are both accurate and trustworthy.
Why forecasts can feel like magic
AI forecasts are not based on intuition. They’re based on patterns in data.
When the right model gets access to the right data, it can produce surprisingly precise insights. But everything starts with one uncomfortable truth:
- AI has no understanding. Only instructions and data. The better the input, the better the output.
- Every insight must be traceable back to an identified factor or cluster in the data – for example via PCA (Principal Component Analysis) or user clusters that affect contact volume, AHT or FCR.
Model choice is driven by your data – and your instructions
In TellMeNow, the choice of model is based on five factors:
- Data type (time series, categorical, numerical, etc.)
- Data quality and volume
- Forecast horizon (how far ahead you want to look)
- Need for interpretability (how much you need to explain why the model says what it does)
- Type of decision the forecast will support
The model does not “know” any of this on its own. We have to define what we are trying to achieve.
That’s why, at TellMeNow, we put a lot of effort into:
- Translating business questions into technically precise instructions
- Linking every forecast to a data-driven insight that can be traced and validated
The main model types we use
1. Time-series analysis
When customer contacts repeat regularly over time (e.g. per month or week), we use models such as:
- SARIMA / SARIMAX – good when you have both trend and seasonality
- Exponential Smoothing (ETS) – simple and robust
- Facebook Prophet – flexible and well-suited to business data
These are the default starting point in TellMeNow (strategy: "fallback_to_simple_regression": true).
They require at least 30 data points and only work properly when the data is sufficiently regular.
2. Machine learning
When there are complex relationships or many interacting factors, we move to machine-learning models such as:
- XGBoost / LightGBM – efficient when you have many variables
- Neural networks (LSTM) – capture longer-term patterns and sequences
Machine-learning models are only used after simpler models have been tested.
They must always be validated using:
- Residual analysis
- Significance testing
…so we can see whether they are actually adding predictive power or just overfitting noise.
3. Bayesian models
To express probabilities in uncertain scenarios, we use Bayesian approaches:
-
BSTS (Bayesian Structural Time Series) – scenario modelling with uncertainty intervals
All Bayesian models we use:
- Report 95% confidence intervals, and
- Are chosen when the business question is about likelihood rather than a single point estimate
(e.g. “How likely is it that volume will exceed X?” rather than “What exact value will it be?”).
4. Hybrid models
Sometimes we combine different model types:
- SARIMA + XGBoost – captures both underlying trends and “unexplained” variation
- Ensemble models – multiple models are combined to increase accuracy
TellMeNow weights these components based on backtesting against historical data – not by simply adding effects on top of each other.
5. Rule-based logic models
In some situations, domain knowledge and human behaviour are crucial. Then we use explicit rules, for example:
“If it’s Monday after a bank holiday → expected volume +15%”
These rules:
- May only be used if they are grounded in observed data
- Must be confirmed via simulations with before/after values
No folklore, no gut feeling dressed up as “AI”.
How forecasting works in TellMeNow – step by step
Step 1: Classify the question
We start by clarifying:
- Are we predicting a value over time (e.g. contact volume) or a state (e.g. churn risk)?
- Do we need a percentage, a numeric value, or a probability?
Step 2: Build a baseline
We begin with transparent models such as SARIMA or Prophet to identify:
- Trend
- Seasonality
- Basic patterns in the data
This gives us a baseline we can explain and compare against.
Step 3: Test advanced models
If several factors clearly interact, we test:
-
Ensemble models, or
-
XGBoost / similar
…but only if:
- The data volume is sufficient, and
- Results are validated with:
-
- Residual analysis (e.g. Durbin–Watson)
- Prediction intervals
- p-values < 0.05
If the advanced model doesn’t clearly beat the baseline in backtesting, we don’t use it.
Step 4: Add uncertainty intervals
Bayesian methods are then used where appropriate to show:
- Not just what might happen
- But how certain we are that it will
This is crucial if you’re taking financial risk on the back of a forecast.
Step 5: Backtesting and validation
Every model is tested against historical data using:
- 95% confidence intervals
- Residual analysis
Want a short explainer on residual analysis? Click here 🙂
- Significance levels
- Before/after values for each recommended action
Example: Forecasting contact volume after a campaign
- Data: 24 months of history, 4 predictors
-
- Seasonality
- Campaign type
- Day of week
- Contact channel
- Model: Prophet + LightGBM (hybrid)
- Simulated effect:
- Reduction in forecast error by 27%
- Specifically in the component for repeat contacts
Important:
- This is an example based on simulated components – not a single “total improvement” figure.
- TellMeNow never presents aggregated “magic numbers”.
Instead, you get separate, clearly defined insight components.
Why transparency is the key to trust
We do not believe in AI as an oracle.
We believe in AI as decision support.
By:
- Showing which models are used and why
- Validating results openly
- Tracing every forecast back to PCA/identified factors or clusters
…we build a platform that is both accurate and credible.
And above all: we ensure that humans – not machines – have the final say.
Frequently Asked Questions (FAQ)
How much data do I need for a forecast?
It depends on the model:
- Some models manage with around 30 data points
- Others need several years of history
TellMeNow selects the simplest model that can answer your question reliably with the data you have.
Can I see why the AI suggests a particular decision?
Yes.
- Every insight is linked to specific factors (e.g. PCA components or clusters)
- These are visualised in a dashboard, so you can see which drivers affect which KPI.
Can I influence how the forecast is used?
Absolutely.
- You decide which KPIs the forecasts should optimise (e.g. FCR, AHT, churn, volume)
- You see before/after values for each insight or recommendation separately
- You can then choose which levers to pull, and when
Conclusion: AI is powerful – in the right hands
AI can never think for itself. It has no context, understanding or responsibility.
What it can do is:
- Execute exactly what you ask it to do
- Do it quickly
- Do it consistently
- Do it in a data-driven way
So it’s not the AI that is “smart” – it’s the people who use it well.
With TellMeNow, you don’t just get access to advanced models. You get a structure, a process and a system where every forecast is:
- Traceable back to PCA or clusters
- Validated with backtesting and residual analysis
- Reported with uncertainty intervals
- Presented without misleading roll-ups – but with concrete, actionable insight elements
We give you accurate insights.
The decisions? Those are yours.
