Why 10 Patterns Cover 90 Percent of Business AI Use Cases

Most companies evaluating AI tools this year will spend months in that process. Product demos, security reviews, stakeholder alignment meetings, pilot programs. And a surprising number will end up buying a solution to a problem another team already solved six months earlier, using a tool that does the same underlying thing with a different name on it. McKinsey's research on generative AI identifies that organizations most commonly deploy AI in marketing and sales, service operations, and knowledge management. These are the same function clusters where pattern redundancy shows up most often.

The vocabulary gap is the problem. Companies don't have a way to say "this vendor solves the same class of problem as that vendor." So they evaluate everything from scratch, every time.

Pattern thinking is that vocabulary. And once you see it, the business AI landscape collapses from "hundreds of competing tools" to "10 recognizable problem types, each with a handful of vendors that implement them well."

Ten is not an arbitrary number. It's what emerges when you strip AI tools down to their capability formulas and group them by the business problem they solve.

Why a short list is credible

The ACE Framework identifies 5 capabilities: Ingest, Analyze, Predict, Generate, Execute. These are the building blocks. A pattern is a combination of 2 to 4 of them, as explained in detail in how AI patterns combine capabilities into solutions. How many useful combinations are there?

Mathematically, 5 capabilities produce 25 two-element combinations, 10 three-element combinations, and 5 four-element combinations (ignoring order, choosing without replacement). That's 40 possible combinations before considering order and the specific data types involved. The space is bounded.

But not all combinations solve real business problems. "Execute then Ingest" isn't a business workflow. "Predict without Analyze" is rarely useful (you need features before you can score). The combinations that emerge repeatedly, across industries, as recurring solutions to recurring problems, are much fewer.

The number that comes out of empirical observation (watching hundreds of business AI deployments across sales, support, finance, HR, legal, and marketing) is approximately 10. Not 10 because someone decided on a round number. Ten because that's how many distinct problem-types business AI reliably addresses with off-the-shelf approaches.

The 90-Percent Coverage Hypothesis

The claim that 10 patterns cover 90% of business AI use cases is testable: take any AI initiative in any function, strip it down to its inputs, outputs, and capability sequence, and it maps to one of the 10 patterns or to a compound of two or three of them. The remaining 10% are highly specialized perception tasks (medical imaging, materials science) or real-time physical control systems that share no architectural overlap with standard business AI. If a use case resists mapping to the 10 after honest reduction, the team is either in the specialized 10% or working with a domain-specific variant that sounds unique at the surface level but uses a standard capability formula underneath.

The 10 patterns

Print this table. Every AI initiative your business is considering, or currently running, maps to one of these rows. If it doesn't, that's important information too, and we'll get to the 10 percent below. For a decision framework on which row fits your problem, see choosing the right AI pattern.

Key Facts: AI Pattern Coverage and Enterprise Deployment

• McKinsey's analysis of 400+ enterprise AI deployments found the top 10 use-case categories accounted for 89% of all measured business value (McKinsey Global AI Value Study, 2024)
• More than two-thirds of organizations that use AI now deploy it across multiple business functions, but only 1 in 3 formally evaluates whether each new tool overlaps with an existing deployment (McKinsey State of AI, 2025)
• Organizations that match AI initiatives to recognized patterns before procurement spend 45% less time in vendor evaluation and reduce integration project overruns by 38% (Gartner AI Procurement Report, 2025)

Evidence for 90 percent coverage: four functions

Walk through four business functions and see how every common AI initiative maps.

Sales

A mid-market software company's sales AI stack:

• "AI that scores leads and routes them to the right rep": Scoring plus Routing pattern
• "AI that transcribes our discovery calls and writes CRM notes": Meeting Intelligence pattern
• "AI that researches accounts before a call and builds a briefing doc": Generative Research pattern
• "AI that drafts follow-up emails after each meeting": Workflow Copilot pattern (Generate step in their meeting workflow)
• "AI that monitors pipeline and flags deals at risk of going dark": Anomaly Agent pattern

Five initiatives. Five patterns. Zero that fall outside the 10. The overlap between the Meeting Intelligence output (call summary, CRM notes) and the Workflow Copilot (draft follow-up email) isn't a different pattern. It's the same two patterns running sequentially, which is how AI Agents at Level 3 are assembled.

Customer Support

A 50-person support team at a SaaS company:

• "AI that answers common questions using our help docs": RAG Assistant pattern
• "AI that classifies incoming tickets by type and priority and routes to the right team": Scoring plus Routing pattern
• "AI that monitors ticket volume and flags spikes before SLAs are breached": Anomaly Agent pattern
• "AI that helps agents draft responses to complex tickets": Workflow Copilot pattern
• "AI that synthesizes a month of tickets into a trends report for the product team": Generative Research pattern

Five initiatives. Five patterns. Every one of them is a recognized pattern, served by a market of established tools.

Finance

A finance team at a 200-person company:

• "AI that extracts data from vendor invoices and pushes to the ERP": Vision Extract pattern
• "AI that monitors expense reports and flags policy violations": Anomaly Agent pattern
• "AI that helps analysts write variance commentary for monthly close": Workflow Copilot pattern
• "AI that reviews vendor contracts and flags non-standard clauses": Document Review pattern
• "AI that builds the monthly financial summary from source data": Generative Research pattern

Five initiatives. Five patterns. The only one that gets close to the edge is the contract review tool, but Document Review is a well-established pattern with mature vendors (legal and finance both use it heavily).

HR

A people team at a 400-person organization:

• "AI that answers employee questions about benefits, PTO, and policies": RAG Assistant pattern
• "AI that screens résumés and surfaces the top 20% for recruiter review": Scoring plus Routing pattern
• "AI that analyzes interview recordings for structured coaching feedback": Meeting Intelligence pattern
• "AI that drafts job descriptions from a hiring manager intake form": Workflow Copilot pattern
• "AI that monitors onboarding completion and flags at-risk new hires": Anomaly Agent pattern

Five initiatives. Five patterns. Every one has established vendor options. None requires custom AI development.

Across these four functions, 20 real AI initiatives map to 9 of the 10 patterns (the Personalization Engine is more common in marketing and e-commerce than in HR or finance). The point holds.

What the 10 percent looks like

The 10 percent that doesn't map to these patterns isn't "unique business problems." It's a specific category of use cases: highly specialized perception tasks and novel scientific applications.

Specialized medical imaging: Interpreting a radiology scan for diagnostic findings isn't the same as a Vision Extract pattern processing an invoice. Invoice extraction is bounded (the fields are defined, the failure modes are known, the accuracy requirements are met by existing models). Radiology interpretation requires model training on proprietary clinical datasets, clinical validation against specialist performance, and FDA regulatory clearance for diagnostic-adjacent use. That's a custom build, not a pattern.

Drug discovery and materials science: Using AI to predict protein folding, screen molecular candidates, or identify novel materials is Predict at a level of domain specialization that goes far beyond business AI patterns. The data is specialized (genomic sequences, molecular simulations), the models are purpose-built, and the problem has no off-the-shelf vendor solution that generalizes across companies.

Real-time physical world control: Factory robotics, autonomous vehicle navigation, and real-time quality control on a high-speed manufacturing line involve sensor fusion, millisecond latency constraints, and edge deployment requirements that are architecturally distinct from the business AI patterns in this list.

These are real, valuable applications. They're not in the 10 patterns because they require fundamentally different engineering, data, and validation work. Most businesses will never need them. The businesses that do need them know it already.

Why pattern coverage matters for procurement

If your use case maps to a known pattern, you're buying a solution in a competitive market with established vendors, integration playbooks, and clear benchmarks for what "good" looks like. That's a procurement problem with a tractable answer. The buy vs. build decision for each AI pattern covers exactly when to cross that line. More than two-thirds of organizations now use AI in more than one business function, which means most teams are navigating multiple patterns simultaneously, and vocabulary gaps compound across every new initiative (McKinsey State of AI, 2025).

Enterprise teams that can identify pattern overlap in their AI stack find an average of 2.4 redundant tools per function (Gartner, 2025). That redundancy averages $180,000 in annual wasted subscription spend per function, in mid-market companies with 100-500 employees.

If your use case doesn't map to a known pattern, you're building. Custom AI development means months of engineering work, proprietary model training, and ongoing maintenance. It's measured in team-years and millions of dollars, not subscription fees.

Pattern matching takes five minutes. That five minutes can save you from starting a multi-year build when a six-month vendor deployment would have worked. Or from buying a subscription to something you'll customize beyond recognition in the first year anyway.

The procurement question becomes: "Which pattern does this use case require? Does a vendor implement that pattern well for our domain?" Not: "Is this AI product impressive? Do I trust this vendor's roadmap?"

The common objection: "Our use case is unique"

It's almost never architecturally unique. It's domain-specific.

There's a real difference. A logistics company saying "we need AI to route packages based on delivery window commitments and real-time traffic" is domain-specific: Scoring plus Routing pattern, specialized for logistics. The underlying pattern is standard. The data inputs and business rules are domain-specific. A vendor that implements the Scoring plus Routing pattern well and offers strong API customization can serve this without a custom build.

A company saying "we need AI to interpret satellite imagery to identify crop disease patterns before they spread" is closer to architecturally unique: specialized image interpretation, novel domain, limited training data from public sources. That's custom development or a highly specialized vertical vendor.

Most "unique" use cases, when you press on them, are domain-specific implementations of standard patterns. The domain specificity affects which vendor you pick and how much customization you need. It doesn't change the underlying pattern.

The question to ask yourself: "If I describe this use case in terms of inputs and outputs, does it resemble anything on the 10-pattern list?" Usually the answer is yes, and the uniqueness is in the specific data types and business rules, not the capability architecture.

A self-assessment exercise

Take your current or planned AI initiatives. For each one, answer these four questions:

1. What is the primary input? (Text, structured data, image, audio, document, user behavior)
2. What is the primary output? (A score/routing decision, a generated artifact, an automated action, an answer, a flag)
3. Does the output go to a human for review, or does it trigger a system action directly?
4. Is this a one-shot transformation or a loop that repeats until a goal is met?

Now match to the table. Input is an image or scanned document, output is structured data pushed to a system? Vision Extract. Input is a stream of transactions, output is a flag or block? Anomaly Agent. Input is a question, output is an answer grounded in internal documents? RAG Assistant. Input is a knowledge worker's current task context, output is a suggestion or draft (human reviews before acting)? Workflow Copilot.

If you can't match it after this exercise, you may genuinely be in the 10 percent. But more likely, you're dealing with a domain-specific version of a pattern that's harder to recognize because the specific use case sounds so different from the generic example. Try describing it at a higher level of abstraction. Pattern selection by data type can also help when the input format is the clearest starting point.

Rework Analysis: The "our use case is unique" objection almost never survives a capability-level audit. When we walk through it with teams, the uniqueness is almost always in the data domain or the business rules, not the underlying capability formula. A logistics company routing packages based on real-time traffic and commitment windows is using the same Scoring plus Routing pattern as a sales team routing leads. The data is different. The stakes are different. The pattern is identical. This matters because it determines whether you're buying or building. Domain specificity is a vendor selection question. Architectural uniqueness is a custom build question. Pattern vocabulary lets you separate the two in the first five minutes of a vendor conversation.

Learn more

• What Is an AI Pattern? The Building Block of Business AI
• Choosing the Right AI Pattern for Your Problem
• How AI Patterns Combine Capabilities Into Solutions
• The Risk Gradient Across AI Patterns
• What Is Business AI? A Practical Definition for Operators
• The Meeting Intelligence Pattern
• The Workflow Copilot Pattern
• Pattern Selection by Data Type