With the process mapped and the bottleneck identified, the next question is: which AI pattern fits here? This module is the reference guide—six universal patterns, how they appear across different areas, and where they don’t fit.
🗂️ AI use case patterns
Every generative AI use case in a business context fits one of these six patterns. Recognizing the pattern is the first step toward recommending the right solution.
Categorize text or data into predefined classes.
Examples: ticket triage, email categorization, document classification by type.
Extract structured information from unstructured text.
Examples: extract CNPJ/amount from invoices, addresses from contracts, dates from emails.
Create new content from data or instructions.
Examples: draft a sales proposal, generate a product description, write a reply email.
Condense extensive content into an actionable summary.
Examples: summarize a recorded meeting, synthesize a contract for approval, ticket summary.
Infer a future state or probability from data.
Examples: lead scoring, churn prediction, probability of default.
Dialogue interface for queries or customer service.
Examples: internal FAQ chatbot, onboarding assistant, first-level help desk.
💼 Opportunities by Functional Area
Each department has recurring patterns. Arriving at the meeting already knowing what to look for in each area speeds up the diagnosis and demonstrates expertise to the client.
Sales / CRM
Support / Customer Success
Finance / Legal
HR / People
🧱 Match the AI pattern to the process step
The right fit is from the inside out: you start with the process step and ask which pattern will solve it. Never the other way around — starting with the tool and looking for a problem for it.
✓ Correct approach (inside-out)
- Identifies the step: “manual contract analysis”
- Describes the work: "text reading, clause extraction"
- Asks for the pattern: "is this extraction/summarization"
- Assesses fit: "do we have the contracts? in what format?"
- Recommends the right solution
✗ Wrong approach (outside-in)
- Saw a chatbot that worked for another client
- They want to implement a chatbot here too
- Looks for an area that "could use a chatbot"
- Tries to fit where it doesn’t
- Delivers a solution the client doesn’t use
💡 Fit criteria for the match
- Available data: does the necessary data exist and is it accessible?
- Acceptable variation: can the output have a margin of error, or does it need to be 100% correct?
- Consequence of error: does an error cause rework or create real risk?
- Sufficient volume: does the process happen frequently enough to justify the investment?
🪜 Map to the pyramid level
For each opportunity identified, the consultant places it on the solutions pyramid. The rule: use the lowest level that solves the problem. Moving unnecessarily up the pyramid increases cost and risk without improving results.
Deterministic — pure rule
When: the output is always determined by the same rules. Examples: automatic calculations, routing by type, threshold alerts.
Tools: n8n, Zapier, scripts, RPA.
AI workflow — standard + variation
When: the work follows a recognizable pattern but varies. Examples: classifying a ticket, extracting data from an invoice, summarizing a document.
Tools: LLM in a workflow, Make + GPT, n8n + Anthropic.
Agent — context + chained decisions
When: the task requires multiple linked decisions and accumulated context. Use with caution — agents are expensive and have more points of failure.
Tools: Anthropic Claude, agentic frameworks.
🚩 Red flags: enticing cases that don't deliver
Some use cases look perfect on paper, but fail predictably when the preconditions aren’t met. Knowing the red flags helps avoid projects that will fail before they begin.
⚠ Chatbot without a knowledge base
A customer service chatbot only works if there is structured, up-to-date content. Without it, it hallucinates or says “I don’t know” to everything. Preparing the knowledge base usually costs more than the chatbot.
⚠ Generating financial reports without clean data
Generative AI amplifies the quality of input data. Dirty data produces incorrect reports with false confidence. “Garbage in, garbage out”—but in polished language.
⚠ Automating a broken process
Automating a process with structural problems only produces errors faster. Fix the process before automating — or you’ll spend money maintaining a machine that produces errors.
⚠ Agent in a risk context with no human in the loop
Autonomous agents making decisions with financial, legal, or health consequences without human oversight. Most of these applications aren't mature enough yet.
📋 Building the Client’s Opportunity Catalog
The deliverable for this phase is a structured table of all the opportunities identified. Still no prioritization — the scope should be deliberately broad before you start cutting.
Catalog structure
| Field | What to document | Example |
|---|---|---|
| Stage | Stage name in the VSM | Support ticket triage |
| AI standard | Classify/Extract/Generate/Summarize/Predict/Chat | Classify |
| Pyramid level | Deterministic / AI Workflow / Agent | AI workflow |
| Preconditions | What needs to be in place for it to work | History of labeled tickets (>500) |
| Dependencies | What needs to be done first | Category cleanup and standardization |
💡 Why list everything before prioritizing
Keeping the catalog broad ensures that non-obvious opportunities aren’t discarded prematurely. Prioritization (module 2.4) will filter them — but you can’t prioritize what isn’t on the list.
🎒 Module summary
Next module:
2.4 — Initial prioritization: Value × Feasibility matrix