The Seven Patterns of AI
Key Takeaways
- The Seven Patterns of AI are Hyperpersonalization; Conversational and Human Interaction; Recognition; Pattern and Anomaly Detection; Predictive Analytics & Decision Support; Goal-Driven Systems; and Autonomous Systems.
- Pattern identification is Phase 1 Business Understanding work: it converts an unscopeable request such as "use AI to grow revenue" into a project with a defined data need, risk profile and success metric.
- Every pattern has a different dominant risk — safety for Autonomous Systems, false-positive investigation load for Pattern and Anomaly Detection, confidently wrong answers for Conversational and Human Interaction, and gaming of the stated objective for Goal-Driven Systems.
- Success is measured differently by pattern: uplift against a control group for Hyperpersonalization, intervention rate for Autonomous Systems, precision — true positives divided by everything the model flagged — at a workable alert volume for Pattern and Anomaly Detection, and straight-through processing rate for Recognition.
- Real solutions combine patterns — claims triage typically pairs Recognition, Pattern and Anomaly Detection, Predictive Analytics, and Conversational and Human Interaction — and each combined pattern should be sequenced as its own CPMAI iteration.
Why a Pattern Comes Before a Technology
An executive asks you to "use AI to grow revenue." That sentence is not a project. It could be personalized offers, a churn model, a sales assistant, dynamic pricing, or revenue-leakage detection — five efforts with different data, different teams, different risks and different definitions of done. The Seven Patterns of AI, the taxonomy PMI carries forward from Cognilytica and lists in the PMI-CPMAI reference materials, exists to close that gap. Naming the pattern converts an ambition into something you can scope, staff, budget and gate.
Pattern identification is work you do in Phase 1, Business Understanding, of the CPMAI methodology — before a vendor demo, before an algorithm is chosen, before anyone quotes a delivery date. The pattern is the first thing that tells you what data you will have to go find in Phase 2 Data Understanding, what your risk register will be dominated by, and what number you will eventually put in front of the sponsor to prove the thing worked.
The Seven Patterns
1. Hyperpersonalization
Business question: how do we treat each customer as an individual, at scale? Examples: next-best-action prompts in a bank, individualized product recommendations, personalized learning paths. Success is measured as uplift against a holdout control group — conversion or retention for treated customers minus untreated — never as raw model accuracy. For the project manager, consent, identity resolution across systems, and the line between helpful personalization and unfair differential treatment become Phase 1 risk items, not late legal surprises.
2. Conversational and Human Interaction
Business question: how do people ask for things in their own words? Examples: tier-one support deflection, voice self-service, a clinical documentation assistant. Success is measured as containment or deflection rate held against satisfaction and clean escalation — an assistant that traps frustrated customers is a failure even at 90% containment. The dominant risk is a confident wrong answer delivered in the organization's voice, so grounding content and escalation rules are scope items, not polish.
3. Recognition
Business question: what is this thing? Examples: invoice and form OCR, defect detection on a production line, medical image triage, identity document checks. Success is measured as straight-through processing rate at a production confidence threshold, with low-confidence items routed to a human. Labeling is the cost driver — somebody must annotate thousands of real examples under real capture conditions. The dominant risk is performance collapsing on inputs unlike the training set, including uneven accuracy across demographic groups.
4. Pattern and Anomaly Detection
Business question: what is unusual here, and what belongs together? Examples: payment fraud, network intrusion, manufacturing quality outliers, claims abuse. Success is measured as precision at the alert volume your investigators can actually work. Precision is a ratio, not a productivity figure: it is the share of everything the model flagged that turns out to be a genuine case — true positives divided by all positives the model raised. Pair it with the number of alerts the team can clear per day, because a model with respectable precision that emits ten times the workable volume still leaves most of its hits uninvestigated. Because true anomalies are rare, the risk profile is dominated by false positives: an alert stream nobody can clear is worse than no alerts. Confirmed labels are also scarce, because you only ever recorded the fraud you caught.
5. Predictive Analytics & Decision Support
Business question: what is likely to happen, and what should we do about it? Examples: demand forecasting, predictive maintenance on pumps, churn scoring, credit risk. Success is measured by the decision outcome — unplanned downtime avoided, stockouts reduced — not by forecast error alone. The implication is that you need long history with the outcome actually recorded, and you must budget for relationships that shift over time, which puts monitoring and retraining inside the original scope rather than in a later phase.
6. Goal-Driven Systems
Business question: what is the best action or sequence of actions to reach a defined objective? Examples: dynamic pricing, crew and resource scheduling, route and bid optimization. Success is measured as objective attained inside a stated constraint set, validated in simulation and in shadow mode before the system is allowed to act. The dominant risk is achieving the stated goal at the cost of an unstated one — margin rises and churn rises with it — so writing the constraints is as much of the job as writing the objective.
7. Autonomous Systems
Business question: what can run with minimal or no human intervention? Examples: warehouse guided vehicles, inspection drones, document processing that files without review. Success is measured as intervention rate — how often a human has to take over — alongside a safety or error-incident rate. Failure consequences are physical, financial and often irreversible, so the risk profile is dominated by safety, liability and edge-case coverage. How much human-in-the-loop control you keep is a scope decision made in Phase 1, not an implementation detail.
Pattern Comparison
| Pattern | Business question | Example | Dominant data need | Dominant risk |
|---|---|---|---|---|
| Hyperpersonalization | How do we treat each person individually at scale? | Next-best-action offers | Per-individual behavioral history, resolved identity | Privacy and consent; unfair differential treatment |
| Conversational and Human Interaction | How do people ask for things in their own words? | Tier-one support deflection assistant | Transcripts, intent taxonomy, curated knowledge source | Confident wrong answers; failed escalation |
| Recognition | What is this thing? | Invoice OCR, defect detection | Large labeled sets covering real capture conditions | Labeling cost; failure on unfamiliar inputs |
| Pattern and Anomaly Detection | What is unusual here? | Payment fraud detection | High-volume transactions with scarce confirmed cases | False-positive load; severe class imbalance |
| Predictive Analytics & Decision Support | What will happen, and what do we do about it? | Predictive maintenance on pumps | Long history with the outcome actually recorded | Drift over time; alerts nobody acts on |
| Goal-Driven Systems | What is the optimal action? | Dynamic pricing, crew scheduling | A defined objective, explicit constraints, a safe place to explore | Optimizing the stated goal and breaking an unstated one |
| Autonomous Systems | What can operate without a human? | Warehouse guided vehicles | Sensor and telemetry data across the full operating envelope | Safety and liability; unseen edge cases |
Patterns Combine — So Sequence Them
Real solutions rarely sit in one box. An insurance claims triage solution is typically Recognition (read the submitted documents), plus Pattern and Anomaly Detection (flag suspicious claims), plus Predictive Analytics & Decision Support (estimate severity and route the claim), plus Conversational and Human Interaction (a status assistant for the claimant). Each of those is a separate data pipeline, a separate evaluation approach and a separate entry in the risk register. The right project-management move is to sequence them as distinct CPMAI iterations rather than one build: start with the pattern whose data already exists and whose business value is easiest to measure, ship it, then take the next one.
On the exam, expect a scenario in which a stakeholder describes a business pain in ordinary language and asks what you should do first. The answer is almost never "evaluate vendors" or "produce a schedule." It is to establish the business problem and name the pattern, because everything downstream — the data inventory, the risk assessment, the success criteria, the scope statement — is derived from it.
A VP of Marketing tells you to "use AI to increase revenue this year" and has already booked a recommendation-engine vendor demo for next week. What should the project manager do first?
A bank's new fraud-detection model (a Pattern and Anomaly Detection solution) flags about 4,000 transactions a day in pilot; the investigation unit can clear roughly 300. The sponsor is pleased because the model catches nearly every known fraud case. Which measure should the project manager put in front of the sponsor?
An insurer asks for a claims solution that reads submitted PDFs, flags suspicious claims, estimates claim severity, and gives claimants a status chatbot — all in one release, in two quarters. What is the best project-management response?