This section provides a capability-first view of the modern AI ecosystem. It is primarily intended for use during the capability-matching phase of use-case viability assessments; a key step in forming an AI strategy for your organisation.

The core capabilities of AI-driven analytics include:

Each of these capabilities can be matched to specific business needs and deployed using purpose-built AI agents within TrueState.

​
Event Prediction

Example: Conversion prediction

A conversion prediction model might use historical data on sales, customer demographics, and market trends to predict if a future lead will convert..

Predictive AI systems use historical data to predict future events. These systems typically rely on machine learning models to generate probabilistic estimates of outcomes such as churn, conversion, or user drop-off.

When to use

  • Will this customer churn?
  • Which leads in our pipeline will convert?

Data requirements

Historical dataset containing:

  • The target event or a proxy (e.g. sales or churn at a customer level)
  • Features known to influence the event (e.g. customer demographics, behaviour, market trends)

Implementation notes

  • Train models on historical data to learn feature-outcome relationships
  • Rows should represent consistent observation windows (e.g. one customer per month)
  • Use a tabular classifier model; TrueState’s default XGBoost classifier is a reliable choice
  • Review the predictive analytics guide for more information.

Backup approach

If the data isn’t ready, LLMs can apply a rules-based estimation approach to mimic expert judgment. This is useful early on, but lacks statistical rigor. If you must use this approach, ensure you collect data to satisfy the data requirements to facilitate a future transition to event prediction..

​
Classification

Example: Email categorisation

Automatically categorising incoming emails as ‘Urgent’, ‘Important’, or ‘Regular’ based on content, sender, and metadata.

Classification systems assign inputs to predefined categories. These systems work with structured data, text, and images across a wide range of scenarios. The difference between event prediction and classification more broadly is that event prediction requires a specific unit of analysis (e.g. “person + month”) to predict an outcome, while classification is a broader use of the same algorithms to classify a wider range of inputs.

When to use

  • Tagging support tickets by priority
  • Flagging at-risk transactions
  • Identifying document type or status

Data requirements

Labeled dataset with:

  • Examples of input items
  • Corresponding class labels
  • Features that differentiate classes

Implementation notes

  • Supports binary and multi-class classification
  • TrueState’s default classifier is optimised for structured data while the text classifier is intended for text classification.
  • Review both the predictive analytics and text analytics guides for more information.

Backup approach

Zero-shot classification using LLMs is helpful for prototyping or when labeled data is limited. Performance may vary; treat as a temporary solution.

​
Granular Forecasting

Example: Inventory demand

Predicting daily demand for specific products across different store locations, factoring in seasonality and local events.

Granular forecasting estimates future numeric values—such as sales volume or traffic—at regular intervals across defined groups or entities.

How it differs from Event Prediction

Event prediction focuses on if an event will occur (e.g. churn), whereas granular forecasting focuses on how much of something will occur (e.g. units sold).

When to use

  • Demand forecasting by SKU or location
  • Financial projections at weekly/monthly levels
  • Traffic or engagement prediction by segment

Data requirements

Time series data including:

  • Historical target values (e.g. units sold, visits)
  • Temporal patterns (seasonality, trends)
  • Known external factors (e.g. holidays, campaigns)
  • 2–3 full seasonal cycles preferred

Implementation notes

  • Include time-based features and external context
  • Watch for anomalies and events
  • Use rolling window validation for accuracy
  • Forecast horizon should match business needs

Backup approach

If limited data exists, simple statistical methods (e.g. moving averages) can establish a baseline.

​
Segmentation

Example: Customer segmentation

Grouping customers into distinct clusters based on purchasing behaviour, engagement patterns, and demographics.

Segmentation groups similar entities (e.g. users, products, sessions) based on behavioural or contextual similarities without using labels.

When to use

  • Identifying customer personas
  • Grouping SKUs by sales patterns
  • Discovering emerging usage cohorts

Data requirements

Dataset with:

  • Features that reflect meaningful variation
  • Normalised input (standardised scales)
  • Minimum sample size: ~100+ rows

Implementation notes

  • Choose cluster count based on business insight or silhouette scoring
  • Use dimensionality reduction for large feature sets
  • Collaborate with stakeholders to label segments post hoc
  • Retrain as user behaviour or business context evolves

Backup approach

Simple rule-based grouping based on expert-defined thresholds can provide quick value while you gather richer data.

Example: Document search

Retrieving relevant documents based on natural language queries, with contextual understanding and synonym matching.

AI-powered search interprets user intent and semantic context to return relevant results across structured and unstructured content.

When to use

  • Searching internal documentation or knowledge bases
  • Retrieving relevant support cases, contracts, or articles
  • Enabling natural language access to business content

Data requirements

Corpus with:

  • Searchable content (e.g. raw text, metadata)
  • Optionally: relevance scores or clickstream logs
  • Clean document structure with IDs and timestamps

Implementation notes

  • Configured using the apply embeddings pipeline node
  • Hybrid search (keyword + vector) improves performance when specific keywords are important (e.g. referencing product names)
  • Refresh index regularly to maintain coverage
  • Review the prompting guide for guidance on how to structure search queries and prepare text data for search.

Backup approach

Traditional search with keyword matching and metadata filters can be used initially.

​
Writing

Example: Report generation

Generating structured reports from performance data, including commentary and recommendations.

AI writing systems generate fluent, human-like text, often combining structured data with domain knowledge to explain, summarise, or narrate insights.

When to use

  • Executive summaries from dashboards
  • Weekly team updates from usage metrics
  • Drafting internal comms or client reports

Data requirements

For domain-specific writing:

  • Examples of ideal outputs
  • Business-relevant terminology and tone-of-voice
  • Structured source data (e.g. tables, charts)

Implementation notes

  • Use output templates to control structure and length
  • Layer in fact-checking and constraint validation
  • Define style rules for clarity and consistency
  • Review samples regularly for tone and factual quality

Backup approach

Template-based systems with rule-based logic can be used for consistent, repeatable outputs before layering in generative language.