If AI makes you uneasy as a business owner, you're not the only one. Most of that unease comes from talking about AI as if it were one thing. The seven patterns of AI split it into seven jobs you can picture in your own business: recognize, converse, personalize, predict, find anomalies, pursue a goal, or act on its own. Once you know which job a project is doing, you can ask the questions you already ask about any investment: what does it need, what can go wrong, and how will we know it works?
Each pattern below comes with an example from a small or mid-size business and the data it usually needs.
- Recognition. The system identifies what something is: reading an invoice, sorting an email, matching a company name to the right record. It needs examples of the thing, correctly labelled.
- Conversation and human interaction. The system talks with people in everyday language: answering staff questions from your policies, drafting replies. It needs the documents it should answer from, and a clear line on what it must not answer.
- Hyperpersonalization. The system treats each customer as an individual: a product suggestion based on what this client actually buys. It needs a history of each customer's behaviour, and their consent to use it.
- Predictive analytics and decision support. The system estimates what is likely to happen so a person can decide: which invoices will be paid late, how much stock to order. It needs enough history to learn from.
- Patterns and anomalies. The system finds what doesn't fit: a duplicate payment, an unusual expense, a statement line that doesn't reconcile. It needs a good picture of what normal looks like.
- Goal-driven systems. The system tries options to reach a goal you set: the best delivery route, a schedule that respects everyone's constraints. It needs a clear goal and the rules it must respect.
- Autonomous systems. The system carries out a task with little or no human help: updating records, filing documents, running a routine check every night. It needs well-defined steps and a safe way to undo a mistake.
Most real projects combine two or three patterns. Name each one, because each brings its own data needs and its own risks.
Because the pattern decides what data you need and what can go wrong, and both stay true whichever vendor you pick.
A common failure starts with a misread pattern. A company asks for a helpful customer chatbot, which is a Conversation job. Along the way it gets built with everything the company knows about each customer, which is a Hyperpersonalization job. The bot starts mentioning things the customer never told it. It works as designed and it feels creepy, because nobody asked which job the customer had signed up for.
Naming the pattern first would have caught it. Conversation needs your documents. Hyperpersonalization needs customer data and consent. In Quebec, Law 25 governs how a business uses personal information: one more reason to name the pattern before you touch customer data.
In our work, most projects land in three patterns: Recognition, Conversation and Autonomous. You don't need all seven. If a proposal for a first project uses all seven, ask what each one is for.
A few examples from client work, described without names:
- Enriching a CRM. Matching incoming companies to the right record is Recognition. Updating the records without anyone retyping them is Autonomous.
- Reconciling statements. Finding the lines that don't match is Patterns and anomalies. Preparing the reconciliation for a person to approve is Autonomous.
- AI assistants for one team. Writing skills that let a team ask questions about its own procedures is Conversation.
Each pattern has known data needs and known ways to fail, so you can plan for its risks the way you plan for any other project. Three habits handle most of them:
- A person reviews every output at first. The system drafts, someone approves. Autonomy comes later, and only for the steps that have earned it.
- Start small. One task, one team, and the smallest access to your systems that gets the job done.
- Decide up front how you will measure it. If you can't say what "working" means, the project isn't ready yet.
Before spending anything, answer ten questions on one page. A question you can't answer yet is the next conversation to have.
- What problem are we solving?
- Who will use it?
- Which pattern, or patterns, is it?
- What data does it need, and do we have it?
- Where does a person review the output?
- What could go wrong, and for whom?
- What does a small pilot look like?
- How will we measure success?
- Who owns it on our side?
- How often will we review it?
The same page works as a brief for any vendor. If you want to see how we turn it into a working system, here is how we build AI agents for mid-size companies.
The seven patterns are part of PMI's Cognitive Project Management for AI (CPMAI) method. Source: PMI, The Seven Patterns of AI.
