A team that builds a Workflow Copilot when they need Scoring plus Routing has wasted six months and a budget. The copilot gets deployed, the team uses it a few times, and then it quietly gets ignored. The underlying problem, too many inbound leads getting routed to the wrong rep too slowly, hasn't been solved. The team eventually goes back to procurement with a new vendor shortlist, having learned nothing.

A team that builds a RAG Assistant when they need Generative Research gets an accurate answer retrieval system that's great at finding things you've already documented, but can't synthesize information from multiple external sources. Different tool. Different capability formula. Sounds similar in a demo.

Pattern selection happens before vendor selection. It takes 20 minutes if you use a structured approach. And it's the decision that determines whether the AI project succeeds, more than any choice you make after it. Gartner reports that organizations with successful AI initiatives invest up to four times more in data and analytics foundations, which tracks directly with pattern selection: knowing which pattern you need tells you exactly which data foundations to prioritize.

Here's the framework.

Step 1: Identify your input type

The first thing that narrows the pattern list is the data going into the system. The ACE Framework's foundation layer identifies seven data types: text, structured, image, audio, video, code, and time-series. For pattern selection, five matter most.

Text inputs are the most common: emails, documents, support tickets, contracts, knowledge base articles, chat messages, form submissions. Text inputs are compatible with almost every pattern, but they particularly narrow toward RAG Assistant (if you're answering questions from text), Generative Research (if you're synthesizing across text sources), Document Review (if you're reviewing a specific text document for compliance or risk), and Workflow Copilot (if a human is working with text and needs drafting assistance).

Structured data inputs are records with defined fields: CRM entries, transaction logs, usage telemetry, survey scores, account attributes. Structured data narrows toward Scoring plus Routing (you're scoring records and routing them) and Anomaly Agent (you're watching a stream for statistical outliers).

Image and scan inputs narrow almost immediately to Vision Extract. If your primary input is a photo, a PDF scan, or a document image, and your goal is to extract field values and put them into a system, Vision Extract is almost certainly the right pattern.

Audio and video inputs narrow toward Meeting Intelligence. Call recordings, meeting recordings, customer interviews, training videos as primary inputs point to Ingest + transcription as the first step, and the Meeting Intelligence pattern built on top of it.

Behavioral/event-stream inputs (user clickstreams, product usage events, browsing behavior, purchase history) point toward Personalization Engine or Anomaly Agent, depending on whether you're personalizing output or monitoring for outliers.

Write down your input type. It's the first filter.

Key Facts: AI Pattern Selection and Project Success

Organizations that formally select an AI pattern before vendor evaluation reduce project failure rates by 55% compared to teams that start with vendor demos (Gartner AI Project Success Study, 2025)

62% of failed AI projects cite "solved the wrong problem" as the primary root cause, which is a pattern mismatch by definition (Deloitte AI Failure Analysis, 2025)

Teams that use a structured pattern-selection framework complete AI procurement in 6 weeks on average, versus 22 weeks for teams using ad-hoc feature comparison (Forrester, 2025)

The Pattern-Problem Fit Matrix

Every AI pattern corresponds to a specific input type and a specific output type. Mismatching the two, such as applying a Workflow Copilot to a problem that actually needs Scoring plus Routing, produces a system that works correctly but solves the wrong problem. The Pattern-Problem Fit Matrix forces teams to name their input and desired output explicitly before evaluating any vendor. Input type narrows to 2 to 3 candidate patterns. Output type narrows to 1. Data readiness and risk tolerance confirm the choice or shift it. This four-step sequence cannot be reversed without wasting procurement cycles.

Step 2: Identify your desired output

Now specify what the system should produce. This is the second filter, and together with input type, it usually narrows the field to one or two patterns.

An answer to a question, grounded in a specific knowledge base: RAG Assistant. The user asks, the system retrieves relevant content and synthesizes an answer with citations. The output is a direct response to a specific query.

A researched report or brief, synthesized from multiple sources: Generative Research. The output isn't an answer to a single question. It's a synthesized document that digests many sources into a coherent narrative. Different from RAG: RAG retrieves from a bounded corpus; Generative Research pulls from broader, multi-source inputs.

A scored record with a routing decision: Scoring plus Routing. The output is not text. It's a score (numerical probability or tier label) that drives a routing action: assign to this rep, escalate to this queue, approve, review, or decline.

Structured fields extracted from an unstructured source: Vision Extract. Input is an image or scan. Output is a database record: vendor name, amount, date, line items. The goal is transformation from unstructured to structured.

A transcript plus enriched notes pushed to a downstream system: Meeting Intelligence. The primary output is structured knowledge from an audio or video source: transcript, topics discussed, action items, CRM updates.

An alert or block based on a statistical anomaly: Anomaly Agent. The output is a flag, an alert, or a triggered action based on something that deviates from the established baseline.

Flags, redlines, or a risk summary from a document: Document Review. The output is an annotated version of the input document (or a summary of its issues) structured around a compliance template or standard.

A draft or suggestion for a human to review and act on: Workflow Copilot. The human stays in the loop. The AI assists. Output is a draft artifact that a person edits and approves before it goes anywhere.

Personalized content or product recommendations delivered per user: Personalization Engine. Output is a different experience for each user, delivered at scale.

A completed multi-step goal that required tool-use and decisions: Autonomous Agent. The output is the result of a goal-directed process, not a single artifact. The system itself decided the path.

The pattern selection matrix

Map your inputs against your desired outputs to narrow the field.

Primary Input	Desired Output	Recommended Pattern
Text question	Answer grounded in internal knowledge base	RAG Assistant
Multi-source text corpus	Synthesized research report or brief	Generative Research
Structured CRM or transaction records	Score + routing decision	Scoring plus Routing
Image, PDF scan, or document photo	Structured database fields	Vision Extract
Audio or video recording	Transcript, summary, CRM notes	Meeting Intelligence
Live data stream (transactions, metrics, events)	Alert or block when something looks wrong	Anomaly Agent
Specific document (contract, policy, report)	Risk flags, missing clauses, compliance check	Document Review
User's current task context (text, email, code)	Draft or suggestion for human review	Workflow Copilot
User behavioral history (clicks, purchases, usage)	Personalized content or recommendations	Personalization Engine
Multi-step goal with tool access	Completed goal across multiple systems	Autonomous Agent

If your input type and output type are both on the same row, you've got your pattern. If your use case spans multiple rows (e.g., you want to ingest a document AND score it AND draft a response), you're looking at a combination of patterns, which is how agents at Level 3 are built. See common AI pattern combinations by department for how teams typically stack these.

Step 3: Check your data readiness

The right pattern is useless without the data to feed it. Each pattern has specific data prerequisites, and teams regularly choose the right pattern but fail to check whether their data can support it.

RAG Assistant needs an indexed, current knowledge base. If your policy documents are scattered across a SharePoint folder, a wiki, and 12 email threads, you don't have a knowledge base. You have a data cleanup project that comes first. McKinsey's research consistently finds data availability and quality among the top AI implementation challenges, regardless of organizational AI maturity.

Scoring plus Routing needs labeled historical data. A lead scoring model needs historical deals with outcomes (won/lost) tied to the attributes you're trying to score on. If your CRM history is incomplete, inconsistent, or short (fewer than 12 months), the scoring model won't have enough signal to be meaningful.

Anomaly Agent needs a baseline. You can't flag anomalies if you haven't established what "normal" looks like. For transaction anomaly detection, you need enough transaction history to define normal patterns. For churn risk, you need enough historical churn data to know what pre-churn behavior looks like. No baseline means the system flags everything and nothing.

Vision Extract needs document consistency. If your invoices come from dozens of different vendors in different formats, the model needs to handle that variability. Test with a sample of real documents before committing to a vendor.

Personalization Engine needs behavioral history per user. Cold-start users (new users with no history) are the Achilles heel of every personalization system. You need enough behavioral data per user to make meaningful predictions. Check your average session depth and user retention before investing in personalization AI.

For each of your top pattern candidates, ask: do we have this data, in this form, at this quality level? If the answer is no, that's your first project, not the AI tool. Data readiness: the prerequisite most AI projects skip is worth reading before you go further.

Step 4: Assess your risk tolerance

Not all patterns carry equal risk. Some patterns read and suggest. Others execute and change state. The distinction matters because it determines what happens when the system is wrong.

Low-risk patterns (read-only or human-gated): RAG Assistant and Generative Research produce text output that a human reads and acts on. If the answer is wrong, the human catches it before anything changes in the world. Workflow Copilot is also relatively low-risk: it drafts, but a human approves before anything is sent or committed.

Medium-risk patterns (Execute with defined, reversible actions): Vision Extract pushes records to a system of record; the wrong field value causes a data error, not an irreversible outcome. Meeting Intelligence pushes CRM notes and may send emails; mistakes are embarrassing but correctable. Scoring plus Routing routes inbound items; a misrouted ticket or lead is annoying but fixable.

Higher-risk patterns (Execute with alerts and blocks): Anomaly Agent blocks transactions or escalates incidents. If it fires incorrectly, it blocks legitimate customers or creates false alarms that erode trust. Document Review flags risks; missed flags create legal exposure. These patterns need confidence thresholds, fallback rules, and human review queues for edge cases.

Highest-risk pattern (autonomous execution in a loop): Autonomous Agent combines all five capabilities in a loop with minimal human checkpoints. Every Execute step inside that loop takes an action in the world. Mistakes compound across steps. For Autonomous Agents, start with a human-review gate on every Execute action and remove it only when you have enough production data to trust the failure rate. See the governance requirements article in Learn More for specifics.

If your organization is early in AI adoption, start with the lower-risk patterns. The value from RAG Assistants and Workflow Copilots is real and the blast radius of a mistake is contained. Graduate toward higher-risk patterns as your team builds operational confidence and your data quality improves. McKinsey's agentic AI governance research underscores this: 80% of organizations have encountered risky behavior from AI agents, and most incidents trace back to premature deployment of high-autonomy patterns before the organization had the operational maturity to manage them.

Teams that apply this matrix before contacting vendors cut their evaluation to one or two finalist tools per pattern, instead of eight to twelve tools across a sprawling shortlist. A competitive RFP for a single pattern takes 3 to 4 weeks. A feature-comparison-driven evaluation without pattern filtering typically runs 4 to 6 months.

Common pattern mismatches and their symptoms

Mismatch: Workflow Copilot when Scoring plus Routing was needed. Symptom: "We built a chatbot for our sales team but nobody uses it." The underlying problem was that reps were wasting time on low-quality leads. A copilot that helps them draft emails faster doesn't fix that. Scoring plus Routing, which automatically prioritizes their queue based on conversion probability, would have. The mismatch is input/output: the copilot takes text requests and produces text drafts; scoring takes structured CRM records and produces a prioritized queue.

Mismatch: RAG Assistant when Generative Research was needed. Symptom: "Our AI keeps saying it doesn't have that information." A RAG Assistant retrieves from a bounded, indexed knowledge base. If the question requires synthesizing information from multiple external sources (competitor analysis, market trends, regulatory changes), RAG can't do it. It can only find things you've already documented and indexed. Generative Research uses a broader ingestion approach and is designed for synthesis across diverse sources.

Mismatch: Anomaly Agent when Scoring plus Routing was needed. Symptom: "Our fraud detection fires on everything." An anomaly agent is for detecting statistical outliers in real-time streams. But if the fraud patterns are well-known and rule-based (transactions over a threshold from new accounts in high-risk geographies), you need a Scoring plus Routing pattern with a trained classifier, not an anomaly detector that's learning baselines from scratch.

Mismatch: Autonomous Agent when Workflow Copilot was needed. Symptom: "Our AI keeps doing things we didn't ask it to do." Autonomous Agents are designed to pursue goals with minimal supervision. If your team wanted AI assistance (suggestions, drafts) but retained control over every action, a Workflow Copilot was the right choice. The autonomy of the Autonomous Agent pattern is a feature for some use cases and a governance problem for others.

Mismatch: Document Review when RAG Assistant was needed. Symptom: "We set up document review AI but our team uses it like a search engine." Document Review analyzes a specific document for compliance, risk, or missing elements against a standard. It's not a question-answering system. If your team wants to ask questions about a body of documents ("what does our vendor agreement say about liability?"), that's RAG.

Choosing between two plausible patterns

The two most commonly confused patterns are RAG Assistant and Generative Research. Both involve text input and synthesized text output. The distinction:

RAG Assistant: Bounded knowledge base. The question has a specific answer that exists somewhere in your internal documents. The retrieval is precise: find the most relevant passages, generate an answer citing them. Best for internal policy Q&A, product documentation lookup, historical ticket resolution lookup. The accuracy depends on how well your knowledge base is indexed and maintained.

Generative Research: Multi-source synthesis. The "question" is more like a research request: "synthesize a competitive analysis of our main competitors based on recent developments." There's no single correct answer sitting in a document. The system needs to pull from multiple sources (news, public filings, web content, internal research) and synthesize across them. Best for market intelligence, account research, trend analysis, due diligence.

If you're unsure which fits: ask whether the answer already exists in a specific document you own. If yes, RAG. If the answer needs to be constructed from multiple sources you don't control, Generative Research.

Rework Analysis: The most expensive pattern mismatch we see is Workflow Copilot deployed where Scoring plus Routing was needed. Both feel like "AI for sales," and copilots demo beautifully. But a copilot that helps reps write better emails doesn't fix a queue clogged with low-quality leads. Pattern selection requires the team to state the problem in terms of inputs and outputs before evaluating any vendor. Organizations that invest 20 minutes in the Pattern-Problem Fit Matrix before issuing a shortlist avoid the most common 6-month waste: building the right AI for the wrong problem.

Frequently Asked Questions

How do you choose the right AI pattern for a business problem?

Use the four-step Pattern-Problem Fit framework: identify your primary input type (text, structured data, image, audio, behavioral stream), specify your desired output type (answer, score, extracted fields, draft, alert, recommendation), assess your data readiness for the candidate pattern, and evaluate your risk tolerance for autonomous execution. Input and output type together narrow the field to one or two patterns in most cases.

What happens when you choose the wrong AI pattern?

A pattern mismatch produces a system that works correctly but solves the wrong problem. Deloitte's 2025 AI Failure Analysis found 62% of failed AI projects cite "solved the wrong problem" as the primary root cause. The typical cost is 6 to 12 months of engineering effort, a failed deployment, and a second procurement cycle to select the right pattern retroactively.

What is the most common AI pattern mismatch?

The most common mismatch is deploying a Workflow Copilot when Scoring plus Routing was needed, particularly in sales functions. Copilots help users draft better outputs, while Scoring plus Routing automatically prioritizes and assigns inbound records. Both are described as "AI for sales" in vendor marketing, but they solve different problems. Identifying this mismatch before procurement requires specifying the desired output explicitly.

What data do you need for each AI pattern?

Data requirements vary by pattern. RAG Assistant needs a current, indexed knowledge base. Scoring plus Routing needs labeled historical records with outcome data (won/lost, converted/churned). Anomaly Agent needs enough transaction or event history to define a statistical baseline. Vision Extract needs document samples that reflect real variability in formats. Personalization Engine needs per-user behavioral history. Missing data is not a reason to delay; it identifies the prerequisite project that must come first.

How does the Pattern-Problem Fit Matrix work?

The matrix maps primary input types against desired output types to recommend a specific pattern. For example, text question input with knowledge-base-grounded answer output points to RAG Assistant. Structured CRM records input with score-plus-routing-decision output points to Scoring plus Routing. When input and output land on the same matrix row, the pattern is identified. When they span multiple rows, the team is building an agent (multiple patterns stacked).

Should AI beginners start with low-risk or high-risk patterns?

Start with low-risk patterns, specifically RAG Assistant and Workflow Copilot. Both produce text outputs that a human reviews before any action is taken, which contains the blast radius of mistakes. McKinsey's agentic AI governance research found that 80% of organizations have encountered risky AI agent behavior, with most incidents traced to premature deployment of high-autonomy patterns before the organization had operational maturity to manage them.

Learn more

Victor Hoang

Co-Founder & CMO, Rework

AI Patterns

Choosing the Right AI Pattern for Your Problem