Use Case Catalog
Twelve sub-categories. Over sixty concrete use cases. One predictive database underneath.
Browse by capability. Each use case shows the outcome, the demo that showcases it best, the Aito operator(s) it uses, and the vertical(s) where it commonly lives. The catalog is the index β for code examples and full query patterns, follow the "Read more" link into the docs. For the live application, follow the demo tag.
What is a predictive application? Read the architectural argument β
Building an agent? See the same operators in a live agent's toolbox, benchmarked β
Analyze
Find what's already in the data.
Pattern discovery, calibration, forecasting, segmentation. The decisions a user makes after seeing the data β and the patterns the data already contains.
Pattern Discovery
Co-occurrence, lift, persona-affinity, statistically discoverable AND predictive patterns. Find what's already in the data.
Pattern Explorer
Co-purchase lift patterns, each shown only when its 95 % interval clears 1. In the generated PetNord data, dog dry food and dental treats go in the same basket 3.17Γ more often than chance.
Rule Mining
AND-rules for GL code and approver, mined live with _relate from the fields known when an invoice arrives (vendor, category, vendor country, amount band), as rule candidates. A person promotes a candidate with sign-off and an audit-trail entry; the ERP demo can also dismiss one.
Supplier Intelligence
Spend leaderboard plus delivery-risk discovery via _relate: which suppliers are over-represented among late deliveries, with the late rate and lift behind each, mined from the purchase history.
360 KPI Dashboard
Pick a segment and every KPI arrives diagnosed: the rate, its root causes via _relate scoped to the segment, and the lever via _recommend, with the rate among customers who already have it.
Market-Basket Rules
Directional A β B basket rules with confidence, lift, and support, mined live by one _relate per anchor category. No precomputed rule table, no Apriori batch job.
Override Mining
Every human correction is logged, and _relate over the override log surfaces the emerging patterns: which values people keep correcting to, and what drives the correction. Each pattern is a rule candidate.
Confounder-Checked Root Cause
The distributor channel looks guilty for churn until climate is held fixed; conditioned, it drops out of the top drivers. A null-control card that finds no lever ships beside it, on purpose.
Exhaustive Driver Map
Every field Γ value Γ slice cell swept with relate() and ranked by how far its lift moves under conditioning: 827 cells from 18 statements. Interactions surface without anyone going looking for them.
Prediction Over a Knowledge Graph
Accounts, contacts and deals are linked records, and one query reads the links both ways and predicts across them: an account's industry judged only by who works there, or a deal's odds from a fact that lives on its account, with $why naming the evidence.
Quality & Calibration
Score your own predictions. Surface where the system can and cannot help.
Honest Evaluation
System scores its own predictions on held-out data. Return Risk reports +0.0 pp gain over baseline and renders as a red row β the system says when it cannot help.
Quality Dashboard
Per-field accuracy and per-confidence-bucket calibration on held-out samples. Green/red diff vs. always-predict-majority baseline.
Calibration Check
Accuracy on held-out cases next to the model's own calibrated confidence, per field, with the individual cases behind each number. The snapshot a calibration monitor tracks over time.
Cold-Start Check
A live _evaluate on the history a tenant has shows how accurate cost center, account code and approver predictions are for that tenant today, field by field, against the base rate.
Rule Precision Replay
Every promoted rule replayed against the tenant's invoices: how many invoices it matches, how often it disagrees with the posted value, and its precision against ground truth.
Forecasting
Predicted quantities at query time. No separate forecasting model to maintain.
Demand Forecast
ERP: _estimate of units_sold per product from the season and last year's sales for the same month, measured on six held-out months against same-month-last-year and the trailing three-month average. E-commerce: _estimate of next month's units per SKU on Supply, seasonal category peaks via _relate on Patterns, and held-out accuracy via _evaluate on Model checks. No separate forecasting model.
Project Success Probability
For every active project, _predict of six outcomes (makes money, team stays happy, customer stays happy, lands on time, the thing works, opens a door) and a combined reading from outcome_score, plus the properties of the work that move each outcome across completed projects from _relate, so the portfolio review starts from the projects most at risk.
Opportunity Scoring
A deal sheet from your own history: win odds with $why, effort in person-days, and reference projects, the numbers an LLM or a gut call would otherwise invent.
Utilization Forecast
Per-consultant load from active assignments, next to each person's historical norm and the share on projects at risk, plus _predict of the role and allocation_pct a person typically takes on a given project type. Capacity planning without a separate forecasting model.
Price Estimation
Fair-price band per SKU from its realised prices with outliers flagged, and _estimate of units, revenue and profit at a candidate list price, answered only within the prices the product has sold at.
Stockout & Overstock Risk
Stock and lead times checked against next month's predicted demand for every SKU: a reorder queue ranked by revenue at risk, and the overstock that ties up capital.
Revenue Outlook
Scheduled revenue per month next to predicted revenue, from _predict of on-time and on-budget delivery for every active project, with $why behind each.
Pipeline Close Likelihood
Every open deal carries a probability of closing, learned from the deals that closed before and shown beside the rep's own typed-in probability, so the deals where the two disagree stand out.
Segmentation
Discover groupings from behavior, not from hard-coded rules.
Customer Segmentation
Segments cut by the outcome you care about, not by similarity: rank the fields that split customers furthest apart on churn or conversion, step into a segment, and get its calibrated rate, what divides it further, and the lever that moves it.
Click-to-Narrow Slice Explorer
Every value is a link: each click adds a condition to the WHERE and re-asks the question. predict() gives the calibrated churn rate for the slice, recommend() the levers, and GROUP BY counts show the raw rates beside them.
Funnel Leak Diagnosis
The funnel names its biggest drop, and for any slice (source, campaign, device, country, landing page) a calibrated conversion rate arrives with the conditions that pull it up or down and the lever that moves it.
Assist
Suggest the next best action β with calibrated confidence.
Search, recommendations, smart forms, conversational. The human stays in control; the system boosts speed and accuracy.
Search & Discovery
Personalized ranking. Match similar items. Find what the visitor is actually looking for.
Smart Search
Persona-conditioned re-ranking. Same query returns different rankings for different shopper segments β derived from the data, not from a curated rule.
Cross-Lingual Semantic Search
A question in French over English notes finds the right documents although it shares no word with them: the text match returns nothing, nearest neighbours over the embeddings fill in, and an exact English query keeps its exact matches.
Recommendations
Cross-sell, similar items, bought-together, for-you. All from the same _recommend operator.
Bought Together
Cross-sell ranked by co-purchase lift: category pairs on Patterns, and per product "often in the same basket" in the Shop. Mined live from order data β not from a curated rule.
For You
Personalized tile rankings per shopper. Same query, different per customer: _recommend with the household's pets, the month, the cart and the shopper's own history as context.
Win-back Ranking
At-risk customers ranked by the spend at stake, each with the offer and the action that bring the most back after their cost. Drives outreach prioritization with an expected return per customer.
Next Best Outreach
_recommend the channel and angle that most likely books the meeting for this buyer, with the predicted meeting rate attached. Optimisation, not description.
Cart Completion
Given what is already in the basket, the products this shopper is most likely to add: the Shop's "Picked for you" re-ranks with _recommend, with the cart's categories as context, as items go in.
B2B Cross-Sell
Pick a product and get what is bought with it: _relate over past order baskets ranks the products that co-occur with it by lift, with the counts behind each pair.
Smart Forms
Multi-field prediction wired into input UI. One field in, the rest predict.
Smart Form Fill
Type any field, the rest predict. Vendor β fills GL, approver, cost centre, VAT %. The form does the thinking, the user confirms.
Smart Entry
Pick a supplier and four fields predict at once β cost center, account code, project, approver. Tab to accept, Esc to reject.
Cart Autofill
One click prefills the basket with the customer's predicted weekly purchases. Used in grocery; same pattern works for B2B reorder flows.
Smart Autocomplete
Predict text completions from user history and corpus patterns. Lower-cost than an LLM and conditioned on actual usage.
Project Plan Drafting
Start a project and the plan drafts itself step by step: the next phase, its tasks, who should do them, and the materials and suppliers, each a _predict with $why that you accept or swap.
Engagement Planner
For a new engagement, a proposed team, the predicted team size, and the odds of delivery success and of winning the deal, from _predict over past assignments, projects, and quotes, plus a price check against history.
Conversational
Natural language interfaces and agent stacks backed by predictions, not by an LLM-only stack.
Agent Toolbox
Give an LLM agent Aito ops as tools: it predicts win odds, estimates effort, queries references, recommends the outreach that books the meeting, then drafts an action that waits for a person to approve it. Grounded numbers an LLM can't invent.
Agent Tool Shortlisting
Shortlist the tools an agent should consider with one calibrated _predict over its own call history, and hand the LLM ~17Γ fewer prompt tokens for its pick.
AI Assistant
Conversational shopping or support interface backed by predicted product/answer rankings. The LLM frames; the predictive database decides.
Question Answering
Predict canonical answers from a prompt corpus. Question-shaped input, predicted-answer output, calibrated confidence to decide auto-vs-route-to-human.
Help Drawer
In-app help articles ranked by _recommend against click history (CTR-ranked, like product recommendations). Users see what other users on the same page clicked.
Company AI Copilot
A chat agent over one linked customer view: it reads KPIs, diagnoses them with $why and _relate, recommends the lever, and drafts a play that waits for a human to approve it.
Automate
Put predictions into the workflow β auto, review, manual tiers.
Categorization, routing, anomaly detection, multi-field prediction, optimization. Calibrated confidence decides what auto-processes and what routes to human review.
Categorization & Routing
The decisions where the user picks from a menu and the right pick is statistically determined by surrounding fields.
PO Queue Routing
Every purchase order arrives with predicted account code, cost center, and approver, sorted into Rule, Aito and Review rows; the coding on Rule and Aito rows can be accepted in bulk.
GL Coding
Predicted GL code and approver per invoice with $why factor decomposition. The core of predictive AP.
Ticket Classification
Auto-route tickets via _predict on description. Confidence-tiered β high auto-routes, mid suggests, low routes to manual.
Confidence-Gated Handoff
Calibrated confidence as governance: high $p auto-resolves, a borderline read goes to the LLM with Aito's shortlist, a low one goes to a human with the tentative answer attached, and sensitive intents are gated regardless. The number decides who acts.
Tag Prediction
Auto-classify content (products, tickets, documents) into a multi-label taxonomy. Saves manual labeling effort with calibrated confidence per tag.
Approver Routing
Predict the right approver from invoice or PO content. Same pattern as GL coding, applied to the person dimension instead of the account dimension.
Payment Matching
Each bank transaction matched to the open invoice it pays. A payment that quotes an open invoice's reference is matched by plain lookup; the rest go to _predict of invoice_id across the link from bank transactions to invoices, with $why showing which name, description, and amount tokens carried the match.
Invoice Line Matching
Supplier invoice lines matched to the catalog SKU they refer to: _predict of sku across the link to products returns a ranked shortlist with reasons, scored live against held-out truth.
Ticket Resolution: Predict vs LLM
The same support ticket resolved by two engines side by side: two _predict calls for the intent and its parameter, with $p and $why, against a live LLM call, with latency and tokens measured on every run.
Anomaly Detection
Inverse prediction β low confidence on a normally-predictable field is the anomaly.
Inverse Prediction Anomaly
If the actual GL or routing is one the prediction gives little probability, the row is flagged for review. The system surfaces what it didn't expect.
Amount-Spike Detection
An invoice or PO amount far above the supplier's history, scored by how few of the supplier's past POs were that large. Surfaced before posting.
Unknown-Vendor Detection
A PO from a vendor with no purchase history on file is flagged for review: a _search for the supplier finds no prior PO.
Multi-field Prediction
One query, several predicted fields. The same pattern that powers smart forms, applied to bulk data fill.
Catalog Gap Fill
Predict category, HS code, account code and tax class, and estimate unit price and weight, across workflow-blocking products. Smart-forms pattern applied to data instead of input UI.
Product Filling
Predicted product attributes β pet type, category, weight, dietary tag, tax class β from a new SKU's name and brand. Multi-field, multi-output.
Feedback Triage
Three predictions per customer review in one round-trip: issue category, sentiment, and suggested assignee, from the review's text and rating.
Optimization & Triggers
Predicted thresholds that drive auto-action. Predicted ranking that drives outreach order.
Markdown Decisions
_estimate across discount levels picks the shallowest level that clears the overstock within a quarter, or says that no markdown will. In the demo's data it proposes none.
Win-back Triggers
Fire a win-back only where the predicted return exceeds the outreach cost. Combines _recommend, _predict and _estimate.
Churn-ranked Outreach
Predict churn probability, route the top-N to retention. The retention team works the highest-leverage queue first.
Address
Episto Oy
Putouskuja 6 a 2
01600 Vantaa
Finland
VAT ID FI34337429
