Kinds of algorithms (in plain language)
Algorithms in advanced analytics fall into a few families. You don't need to implement them; it helps to know what each family is for.
Predicting a number (regression)
What it does: Estimates a numeric value (e.g. next month's revenue, a customer's lifetime value, how long until a machine fails).
Idea: Use past data where you know the outcome, find patterns (e.g. "when X and Y were high, the outcome was Z"), then apply those patterns to new cases where you don't yet know the outcome.
Typical use: Demand forecasting, pricing, risk scores, resource planning.
TrueState’s platform builds and iterates regression-style forecasting and scoring workflows without ML engineering on your side for each new question.
Putting things into categories (classification)
What it does: Assigns a label or category (e.g. "will churn" vs "will stay," "high risk" vs "low risk," "spam" vs "not spam").
Idea: Learn from historical examples where the label is known, then apply that learning to new cases to assign the most likely label.
Typical use: Credit approval, fraud detection, customer segmentation (e.g. likely buyer vs not), prioritisation (e.g. which lead to follow first).
Classification (churn, risk, prioritisation) is where agentic automation removes the biggest queue — TrueState trains and refreshes models as data changes.
Finding groups in the data (clustering)
What it does: Groups similar items together without you having to define the groups in advance. The algorithm discovers structure in the data.
Idea: Items that are "close" in terms of their attributes end up in the same cluster; items that are different end up in different clusters.
Typical use: Customer segments you didn't pre-define, anomaly detection (the odd one out), organising large sets of entities (products, documents, locations).
Clustering and anomaly work still needs careful interpretation — but discovery and re-run cycles are far faster when the platform handles the heavy lifting.
Working with sequences and time (time series)
What it does: Models how something changes over time — trend, seasonality, level — and uses that to forecast the next values or detect when something breaks from the pattern.
Idea: Past behaviour over time is used to project forward or to flag "this week looks different from normal."
Typical use: Sales forecasting, capacity planning, early warning (e.g. drop in conversions, spike in complaints).
Time series forecasting and monitoring are a natural fit for repeated, scheduled agent runs on clean warehouse data.
Working with text and language (NLP)
What it does: Extracts meaning, topics, or structure from text (emails, documents, chat, reviews) so you can search, summarise, or classify it at scale.
Idea: Represent text in a way a machine can work with (e.g. counts, embeddings), then apply rules or models to classify, extract entities, or summarise.
Typical use: Sentiment analysis, document routing, chatbots, summarising feedback or reports.
NLP layers (summarisation, routing, tagging) plug into the same stack once text is available to the warehouse or adjacent systems.
Optimising decisions (optimisation)
What it does: Finds the best combination of choices (e.g. prices, allocations, schedules) given your goal and constraints (e.g. capacity, rules, risk limits).
Idea: Define an objective (e.g. maximise profit, minimise cost) and the rules that must hold; the algorithm searches for the best feasible solution.
Typical use: Pricing, resource allocation, scheduling, marketing spend, inventory.
Optimisation scenarios benefit when models and constraints are versioned and re-solved as inputs change — work that otherwise sits with a small specialist group.
In practice, a single use case often combines several of these (e.g. a classification model whose inputs were built using feature engineering and then tuned with optimisation). The important thing is to know what kind of question you're answering — a number, a category, a group, a forecast, text, or a decision — and that maps to which family of algorithms is relevant.