A Map of Where AI Creates Value in Your Business

TL;DR:

• Most AI strategies start with a use case list. It feels rigorous, but it scatters spending across pilots that never build on each other.
• Plan by mechanism instead: what the AI actually does inside a process, rather than the label on the use case. There are thirteen mechanisms, eight for how you operate and five for what you sell.
• The mechanism you pick decides the infrastructure you need, so choose your top three to five and build for those. If you get it wrong, you pay for plumbing that never reaches the value you were after.

The typical AI strategy starts with a use case list. Someone surveys the departments, collects ideas, scores them on a matrix (impact vs. feasibility, or some variant), and picks the top five to pilot. This process feels structured. It produces a spreadsheet. It gets presented to the Board.

It also produces scattered investments that do not compound.

The problem is that use cases are the wrong unit of analysis for strategic planning. “Automate customer service responses” is a use case. But it can create value through at least four different mechanisms: substituting for human cognitive labor (1,000 fewer tickets handled by agents), improving customer experience (response time drops from hours to seconds), providing surge capacity (handling 4x volume during a product launch without hiring), and accumulating proprietary training data (every interaction improves the model). Each mechanism has different infrastructure requirements, different ROI profiles, and different implications for what you need to build next.

When you plan by use case, you build for one of those mechanisms and accidentally leave the other three on the table. When you plan by mechanism, you build infrastructure that captures all four.

The organizing principle

The cleanest way to map AI value is by mechanism of value creation: what the AI is fundamentally doing in the business process that a non-AI alternative could not do as well, as cheaply, or at all. This avoids the double-counting problem that plagues financial-outcome frameworks (revenue vs. cost vs. risk) and the replication problem that plagues function-based frameworks (marketing vs. ops vs. finance).

The map has two branches that are mutually exclusive. Internal value is how AI changes how the company operates. External value is how AI changes what the company sells or how customers experience it. A given deployment can produce both, but the mechanisms are distinct.

The 13 mechanism categories and 43 leaf-level value sources are in the explorer below. Each leaf is annotated with the public examples that exemplify it, the infrastructure it requires, and where the framework strains. The explorer lets you filter by company archetype to surface the mechanisms your business model structurally over-indexes on, and the Surprise me button picks a random mechanism to seed brainstorming.

Internal value: how AI changes how you operate

Eight mechanisms cover the space.

Cognitive labor substitution

AI performs work that previously required a human knowledge worker, end-to-end or in a bounded task. Most companies think of this mechanism first, and it has the most visible ROI.

• Routine task automation. Klarna’s customer service AI handled two-thirds of chats in its first month, equivalent to roughly 700 full-time agents, with resolution time dropping from 11 minutes to under 2 (figures as of February 2024; Klarna later reintroduced human agents for complex cases).
• Non-routine task automation. An agentic system takes a sales discovery call transcript, drafts a tailored proposal, pulls relevant case studies, and routes for review. That work used to require four to six hours from a senior consultant.
• Specialized expertise arbitrage. A mid-market law firm uses AI for first-pass contract review, reducing reliance on senior associates billing $400 per hour.

The infrastructure requirement scales with the task. Routine automation needs structured inputs and clear success criteria. Non-routine automation needs unstructured-data access and retrieval infrastructure. Expertise arbitrage often needs domain-specific fine-tuning or heavily curated retrieval-augmented generation (RAG).

Cognitive labor augmentation

AI does not replace the human; it makes the human more productive or more capable per unit of time.

• Throughput amplification. Same person, more output per hour. The canonical example is GitHub’s 2022 controlled study showing developers complete a benchmarked HTTP-server task 55% faster with Copilot than without.
• Skill ceiling expansion. The human now does work they previously could not do at all. A non-designer founder produces brand-quality marketing assets via image generation, removing the need for an early-stage agency engagement.
• Quality floor elevation. Variance shrinks across a workforce. Brynjolfsson, Li, and Raymond’s 2023 contact-center study found a 34% productivity improvement for novice agents and minimal effect on experienced ones.

The infrastructure requirement here is lighter than substitution. Augmentation typically works through APIs and tool integrations rather than custom pipelines. The architecture investment is minimal until you want to augment with company-specific context, at which point you need the same retrieval infrastructure as non-routine substitution.

Decision quality improvement

AI changes which decisions get made. Faster execution of the same decisions is augmentation; this mechanism is reaching a different and better decision.

• Synthesis at scale. A pharma medical-affairs team uses AI to synthesize 10,000+ clinical papers and identify emerging treatment patterns in days rather than months.
• Counterfactual and scenario reasoning. A private equity diligence team uses AI to surface negative cases and management red flags that pattern-match to prior failed deals.
• Real-time decision support. A B2B sales rep gets a real-time briefing during a customer call that flags recent support tickets, contract renewal dates, and competitor mentions.
• Resource allocation optimization. A retailer dynamically reallocates inventory across stores based on real-time demand signals, at a granularity manual ops cannot maintain.
• Lead and pipeline prediction. Prospects, accounts, transactions, or subscribers get ranked by likelihood to convert, churn, or default. Sales, marketing, or credit attention follows the ranking.

This is the heaviest infrastructure surface of the internal mechanisms. Synthesis needs retrieval over large unstructured corpora. Real-time decision support needs low-latency access to multiple structured and unstructured sources at once. Resource allocation needs a closed-loop system with write-back capability. The context layer investment pays off most directly here.

Process reconfiguration

AI enables operating models that were previously infeasible. The question is not whether AI can do the existing task; it is what becomes possible that was not possible before.

• Workflow compression. A media company eliminates its dedicated localization team because content can be translated, culturally adapted, and quality-checked by an agentic pipeline.
• Hyper-personalization at scale. Every customer or employee gets what was previously feasible only for a high-value subset. Duolingo’s Birdbrain model generates individualized lesson paths for every learner rather than only paid premium users.
• Always-on operations. Functions that previously had business-hours constraints run continuously without proportional cost increase. An ecommerce company’s AI shopping assistant handles pre-purchase questions overnight in non-English languages.

The infrastructure requirement is integration and orchestration. The AI has to reach across the systems a workflow touches and coordinate multi-step actions, often with write access, not just read from a single source. That is a heavier lift than retrieval, because it means connecting to operational systems and handling what happens when a step fails.

Knowledge capture and activation

AI turns latent organizational knowledge into a usable asset. The source material is internal and proprietary, and the value comes from making tacit or buried knowledge accessible.

• Institutional memory unlock. A consulting firm builds a retrieval system over fifteen years of project deliverables, so proposal teams find and adapt prior work in minutes rather than weeks.
• Tacit knowledge externalization. A field service company records, transcribes, and structures senior technicians’ diagnostic approaches into an AI assistant that junior techs consult on the truck.
• Onboarding and ramp acceleration. New employees become productive faster because AI provides the institutional context they previously had to acquire by tenure.

The infrastructure requirement is primarily retrieval: document processing, embedding, indexing, and governed search. Companies whose primary mechanism is knowledge capture should invest in retrieval infrastructure before they invest in semantic layers over structured data.

Risk and compliance reduction

AI reduces the probability or magnitude of adverse events.

• Error and defect reduction. An AI code review system flags security vulnerabilities before merge.
• Regulatory and policy adherence. A bank monitors customer-facing communications for disclosure compliance.
• Threat detection. A payments company detects novel fraud patterns in real time across billions of transactions.
• Incident response and triage. AI investigates production incidents in flight, clustering related alerts, surfacing similar past outages, drafting root-cause hypotheses, and routing to the right responder. Distinct from defect reduction; the system is already broken, and the AI’s job is to compress mean-time-to-resolution.

The infrastructure requirement is monitoring and logging. This mechanism depends on scoring events in real time and on the audit trails that let you prove what happened and reconstruct it afterward, so coverage and latency matter more here than in almost any other mechanism.

Innovation and R&D acceleration

AI compresses the cycle from question to discovery to product.

• Discovery acceleration. AI generates and evaluates candidate solutions at scales researchers cannot match. Drug-discovery platforms compress target-to-lead timelines from years to months.
• Experimentation throughput. A consumer product company runs thousands of AI-generated packaging variants through simulated shopper-response models before committing to physical prototypes.

The infrastructure requirement is specialized and compute-heavy: domain models, simulation environments, and pipelines to generate and evaluate candidates at volume. It is the least reusable of the mechanisms, so the investment only makes sense once you have confirmed that discovery, and not something downstream of it, is your real bottleneck.

Organizational resilience and optionality

AI changes the company’s capacity to absorb shocks and pivot strategically.

• Surge capacity. A tax-preparation firm absorbs 4x volume during peak filing season via AI agents, where previously it recruited and trained seasonal labor with quality variance.
• Strategic speed. New product launches, market entries, and competitive responses happen in weeks instead of quarters.
• Talent attraction. The company’s AI-equipped operating environment becomes a recruiting asset because top performers want to work where their output is amplified.
• Burnout reduction. AI removes the most demoralizing parts of work, with measurable retention impact. A 2024 multi-site quality-improvement study of ambient AI scribes across six US health systems found burnout dropped from 51.9% to 38.8% after 30 days, and 62% of clinicians said they were more likely to extend their clinical career.

There is no single infrastructure requirement here, because resilience draws on whatever you have already built for the other mechanisms. What it does assume is elastic capacity: systems that scale up on demand instead of forcing you to provision for the peak in advance.

External value: how AI changes what you sell

Five mechanisms cover the space.

Product capability expansion

The product does things it could not do before.

• New AI-native product lines. Cursor, the AI-first code editor, crossed $2B in annualized revenue by February 2026 (up from $500M in mid-2025) in a category that was created by AI.
• Feature additions to existing products. Notion AI is bundled into Notion’s Business and Enterprise tiers as the differentiator over Free and Plus; Adobe Firefly is built into Creative Cloud apps as Generative Fill, Generative Expand, and similar capabilities.
• Interface transformation. Replit’s Agent (launched September 2024) shifted the company from a code editor for developers to a natural-language app builder usable by non-developers.

The infrastructure requirement is production-grade serving. The AI is now inside the product, so it needs the reliability, latency, and monitoring you would give any customer-facing system, not the looser standards internal tools tolerate. The durable advantage still comes from proprietary data, not the serving stack, which competitors can build too.

Customer experience differentiation

The product is the same; the experience around it is meaningfully better.

• Service responsiveness. Support tickets resolve in minutes instead of hours.
• Personalization of experience. Each customer gets a tailored experience that drives engagement. Spotify’s AI DJ creates individualized commentary about the tracks it picks; on days users open the feature, they spend roughly 25% of their listening time with it.
• Self-service capability. Customers accomplish more on their own. An insurance company’s AI claims agent walks customers through the entire filing process, replacing multiple phone calls.

The infrastructure requirement is a live, unified view of each customer: the profile, history, and context the experience personalizes against, served fast enough to use in the moment. Personalization is only as good as the customer data behind it, so clean, connected customer records come first.

Market reach expansion

AI lets the company sell to customers it previously could not economically reach.

• Geographic and language expansion. Markets that were uneconomic due to localization costs become accessible. A US-based legal-tech company enters 12 European markets simultaneously because AI handles translation and legal nuance.
• Down-market expansion. Customer segments that were previously unprofitable become profitable. A wealth-management firm offers AI-driven advisory to clients with $50K in assets, where human advisors started at $500K minimums.
• Long-tail coverage. Niche use cases get profitable coverage. An ecommerce platform generates product descriptions, images, and SEO content for millions of long-tail stock-keeping units (SKUs) that previously got minimal attention.

The infrastructure requirement is mostly translation and localization pipelines, plus the automation that makes a low-value segment cheap enough to serve. The harder work is often operational rather than technical: the compliance, support, and payment plumbing each new market or segment demands.

Pricing power capture

AI lets the company charge more or shift to better-aligned pricing models.

• Outcome- and consumption-based pricing. Revenue ties to usage or results rather than seats. A legal AI vendor prices by document processed, not by user.
• Premium tier creation. AI features become the basis for an upsell. Microsoft 365 Copilot launched in November 2023 at $30 per user per month as an add-on to Microsoft 365 enterprise subscriptions.
• Switching cost increase. AI personalization makes the product harder to leave. An assistant trained on a customer’s documents and history would take months to recreate elsewhere.

The infrastructure requirement is metering. To charge by usage or outcome, you first have to measure usage or outcome accurately and bill on it, which many companies cannot do today. The instrumentation to track what each customer actually consumes has to exist before the pricing model can change.

Network and data effects

AI deployment generates assets that compound.

• Proprietary data accumulation. Customer interactions improve the product for everyone. GitHub Copilot’s quality improves as more developers accept, reject, and modify suggestions.
• Workflow lock-in through fine-tuning. Models adapted to a customer’s domain become more valuable with use.
• Ecosystem and platform effects. AI capabilities draw third-party developers whose contributions make the platform more valuable.

The infrastructure requirement is the feedback loop: capturing the signal from real usage, feeding it back into the models, and closing the loop so the product improves as adoption grows. This is the mechanism where the data pipeline itself is the advantage, so it is worth building deliberately rather than assuming the advantage will compound on its own.

Where to start

The framework is a diagnostic tool, not a theory exercise. Three steps:

1. Identify your three to five primary mechanisms. Walk the thirteen categories. For most mid-market companies, the answer clusters around two or three internal mechanisms (typically some mix of cognitive labor augmentation, decision quality, and knowledge capture) and one or two external mechanisms (typically product capability or customer experience). If you mark eight or more, force-rank to the top three before moving on; trying to cover the field dilutes infrastructure investment.
2. Audit alignment. Most companies discover a mismatch. They are investing in cognitive labor substitution (chatbots, automation) when their highest-value mechanism is decision quality or knowledge capture. The infrastructure differs: a chatbot needs a narrow task scope; decision quality needs broad retrieval across structured and unstructured data. Building for the wrong mechanism means the infrastructure you are paying for does not serve the use cases that would create the most value.
3. Let the mechanism drive the architecture. Cognitive labor augmentation needs API access and tool integrations. Knowledge capture needs retrieval infrastructure. Decision quality needs a context layer with entity models, gold models, and a semantic layer. Product capability expansion needs embedded infrastructure with tenant isolation. Operational action-taking needs identity resolution, authorization, and reversibility.

The mechanism map is the prerequisite. Without it, you do not know what to build.

If you have a list of pilots scattered across departments and need to convert them into a mechanism-ranked investment sequence, start a conversation; this is where our AI strategy practice begins. If you already know your primary mechanisms and your current infrastructure does not match what they need (chatbot tooling serving a decision-quality use case, or a warehouse without retrieval serving knowledge capture), our data assessment and data engineering practices are where most engagements begin.

Sources

• Klarna AI assistant handles two-thirds of customer service chats in its first month. Klarna press release, February 2024. Source for the 700-FTE-equivalent figure and the 11-to-2-minute resolution-time drop.
• Klarna reinvests in human customer service. Customer Experience Dive, 2025. Context on Klarna’s later reintroduction of human agents for complex cases.
• Research: quantifying GitHub Copilot’s impact on developer productivity and happiness. GitHub Research, 2022. Source for the 55%-faster benchmark on the HTTP-server controlled experiment.
• Generative AI at Work. Brynjolfsson, Li, and Raymond, NBER Working Paper 31161, 2023 (later in Quarterly Journal of Economics, 2025). Source for the 34% productivity gain among novice contact-center agents.
• Learning how to help you learn: Introducing Birdbrain. Duolingo Blog. Source for Duolingo’s individualized lesson-path personalization system.
• Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout. JAMA Network Open, 2024. Source for clinician burnout dropping from 51.9% to 38.8% after 30 days and the 62% career-extension figure across six US health systems.
• Past, Present, and Future. Cursor (Anysphere) Series D announcement. Source for Cursor’s annualized-revenue trajectory ($500M in mid-2025 to $2B by February 2026).
• Notion Pricing. Source for Notion AI being bundled into the Business and Enterprise tiers (May 2025 pricing change consolidated the standalone AI add-on into paid tiers).
• Adobe Releases New Firefly Generative AI Models and Web App, Integrates Firefly Into Creative Cloud. Adobe Newsroom, 2023. Source for Firefly’s integration into Creative Cloud apps as Generative Fill, Generative Expand, and Generative Recolor.
• Replit Agent. Replit product page. Source for Replit’s natural-language app builder, launched September 2024.
• Spotify Debuts a New AI DJ, Right in Your Pocket. Spotify Newsroom, February 2023. Source for the AI DJ feature and the 25%-of-listening-time engagement figure.
• Announcing Copilot for Microsoft 365 general availability and Microsoft 365 Chat. Microsoft 365 Blog, September 2023. Source for the November 2023 commercial launch at $30 per user per month.

Browse the value map for your business →