Definition
The AI Capability Framework, also called the Master Stack, is a four-tier model for reasoning about artificial intelligence once the term “AI” stops being a useful unit of decision-making. It was developed by Greg Kihlström and is the organizing structure of his book Stop Saying “AI”.
The framework starts from a specific prediction: “AI” as a label is on the same path the word “electric” took. We once said electric light and electric refrigerator; the qualifier fell off the moment electricity became the assumed default. “AI-powered” is that qualifier now, and it will erode the same way. The strategic problem is that a capability nobody names is a capability nobody budgets, governs, assigns, or measures. The umbrella term is comfortable, and that comfort is what blocks management.
The Master Stack replaces the single word with four categorically distinct tiers. Read from the top down, it forms one sentence: why → what AI does → what that requires → within what limits.
- Goals / Strategy — the apex. The directional aim a person or organization brings to the system. Goals are not a layer and not a capability; they are what everything else serves.
- Pillars — what AI does: Augmentation, Insights, Orchestration, Generation. Each is a capability.
- Layers — the conditions every pillar requires but that are not themselves acts: Interface, Memory/Context, Trust/Verification, Identity/Permissions.
- Forces — the external pressures that bend the whole system from outside: compute economics, energy, regulation.
Cutting across the pillars, on its own axis rather than inside the stack, is the mechanism — the how each capability is built (rules → predictive → generative → agentic). Mechanism is orthogonal to the tiers, not a fifth band.
The distinctions between tiers are load-bearing. A pillar is something AI does. A layer is a condition the doing requires — you don’t “do” trust, you establish the conditions for it. A force is contended with, not designed in. Goals aim the work; layers enable it. Keeping these separate is the discipline that lets the framework survive contact with a fast-moving technology, because the categories stay stable even as the tools inside them churn.
How It Relates to Marketing
Marketing is where the “AI” umbrella does the most quiet damage, because marketing technology is where the largest number of AI features hide under a single word. A marketing organization that says “we’re investing in AI” has not said anything a CFO can fund or a CMO can measure. The Master Stack turns that sentence into a set of decisions.
Applied to marketing, the framework does several things at once:
- Turns a budget line into a portfolio. “AI in the martech stack” becomes a set of named capabilities — a Generation pillar producing subject-line variants, an Insights pillar scoring leads, an Orchestration pillar routing journeys — each with its own cost, owner, and risk profile.
- Separates the safe-to-fade from the must-stay-named. Subject-line variants and send-time optimization can disappear into the workflow. A price quote, a credit decision, or a claim denial cannot. The framework draws that line down the middle of the funnel.
- Exposes over-built features. Naming the mechanism under each capability reveals where a heavy, expensive technique is doing a job a lighter one would clear — the recurring “forklift carrying a coffee cup” pattern.
- Gives governance a target. You can only govern what you can name. Once a capability is named, the AI Governance Board, the audit trail, and the human-in-the-loop control have something specific to attach to.
The Four Tiers at a Glance
| Tier | What it answers | Elements | Marketing example |
|---|---|---|---|
| Goals / Strategy | Why, and toward what? | The organization’s aim | Grow retention in a defined segment |
| Pillars | What does AI do? | Augmentation, Insights, Orchestration, Generation | Generation drafts variants; Insights scores intent |
| Layers | What does the doing require? | Interface, Memory/Context, Trust/Verification, Identity/Permissions | Provenance on generated content; permissions on an agent’s spend |
| Forces | What bounds it? | Compute economics, energy, regulation | Token cost per send; consent and disclosure rules |
The mechanism axis (rules → predictive → generative → agentic) runs across the pillars rather than sitting in the stack, because any pillar can be built with any mechanism.
How to Apply the Framework
The Master Stack is a decision structure, not a maturity score. A practical sequence:
- Name the goal first. State the business aim the capability serves, in plain language, before naming any technology. If the goal is vague, everything below it will be too.
- Sort every “AI” item into a pillar. Take each feature currently filed under “AI” and place it in Augmentation, Insights, Orchestration, or Generation. Use the Author Test to keep Augmentation and Generation distinct: remove the AI, and ask whether a skilled human could still produce the output, just slower.
- Name the mechanism under each pillar. Is the capability delivered by a rule, a predictive model, a generative model, or an agentic loop? This is where most cost and most accuracy risk live.
- Check the four layers for each capability. What does the interface look like, what memory or context does it hold, what does trust and verification require here, and on whose behalf and within what bounds does it act?
- Locate the capability against cost-of-wrong. Decide whether being wrong is cheap and loud (safe to fade) or expensive and quiet (must stay named and watched).
- Read the forces. Note the compute, energy, and regulatory pressures acting on the capability, since they move independently of your design choices.
The output is a portfolio in which every capability has a pillar, a mechanism, a trust condition, and a cost-of-wrong — the four-column view the book calls the decomposition table.
Comparison to Similar Frameworks
| Framework | Focus | What it organizes |
|---|---|---|
| AI Capability Framework (Master Stack) | What AI does, requires, and is bounded by | AI capabilities, conditions, forces, and the goal they serve |
| Wardley Mapping | Evolution of components from genesis to commodity | How capabilities drift toward assumed infrastructure |
| Diffusion of Innovations | How new technology spreads through a population | Adoption over time |
| Crossing the Chasm | The gap between early adopters and the mainstream | Go-to-market sequencing |
| Balanced Scorecard | Translating strategy into balanced measures | Objectives across four perspectives |
The Master Stack shares Wardley Mapping’s interest in capabilities becoming commoditized, but it is built for a technology that is probabilistic and fallible rather than reliable, which is why it keeps a Trust layer and a cost-of-wrong axis that a commodity-evolution map does not need. It pairs naturally with a Balanced Scorecard: the Master Stack names the AI capabilities, and the scorecard measures whether they served the goal.
Best Practices
- Name the capability before naming the tool. The pillar comes first; the vendor and the mechanism come after. Reversing the order is how organizations buy the impressive technique for a modest job.
- Hold the two collisions apart. A generative mechanism is not the Generation pillar, and an agentic mechanism is not the Orchestration pillar. The shared words invite a category error that leads straight to over-buying.
- Keep goals out of the stack. Because everything serves goals, folding strategy in as another component lets it absorb the whole model and stop distinguishing anything. Keep it at the apex.
- Let low-stakes capabilities fade on purpose. Managing an ambient, forgettable feature as if it were high-stakes is its own form of waste.
- Don’t let the forces tier sprawl. Compute, energy, and regulation are external pressures. Resist the urge to redraw internal design choices as forces.
Future Trends
- The qualifier keeps dropping. As “AI-powered” erodes from product language, the pressure to name capabilities precisely will grow, not shrink, because invisible capabilities are the ones that go unmanaged.
- Mechanism churn under stable pillars. The pillars are designed to stay stable while the mechanisms serving them keep re-versioning and repricing. Expect the framework’s value to concentrate at the mechanism axis and the Trust layer, where the movement is.
- Governance shifts to the capability level. Regulation and internal governance increasingly attach to specific capabilities and their cost-of-wrong rather than to “AI” as a category, which is the level at which the Master Stack operates.
- Strategy becomes the scarce input. As capability becomes cheap and ambient, advantage migrates to the apex — the choice of what to aim the stack at.
FAQs
1. Who created the AI Capability Framework? Greg Kihlström, author of Stop Saying “AI”, where the Master Stack is the book’s organizing structure.
2. Why is it called the Master Stack? Because it stacks the tiers — goals, pillars, layers, forces — into a single readable structure, with the mechanism axis crossing it. Read top to bottom it forms one sentence: why, what AI does, what that requires, within what limits.
3. What are the four pillars? Augmentation, Insights, Orchestration, and Generation — the four things AI does, each a distinct capability.
4. What’s the difference between a pillar, a layer, and a force? A pillar is a capability AI performs. A layer is a condition that capability requires but that isn’t itself an act, such as trust. A force is an external pressure the system contends with, such as regulation. Keeping the three categories separate is the framework’s core discipline.
5. Where is “AI setting strategy” in the model? It’s the Insights and Augmentation pillars applied to the strategy domain — a use case inside existing pillars, not a new building block. Goals stay at the apex.
6. Why does the framework exist if “AI” is fading? Because the fade is uneven. Capabilities that become invisible stop being managed, and the framework restores the vocabulary needed to budget, govern, assign, and measure them.
Related Terms
- The Fade (Uneven Disappearance)
- Place vs. Property
- The Four Pillars (Augmentation, Insights, Orchestration, Generation)
- The Mechanism Axis
- Cost-of-Wrong
- The Four Layers
- The Three Forces
- The Four Management Verbs (Budget, Govern, Assign, Measure)
- Artificial Intelligence (AI)
- Generative AI
- Agentic AI
- Wardley Mapping
- Balanced Scorecard (BSC)
Sources
- Kihlström, Greg. Stop Saying “AI”. https://amzn.to/4wilWcA
- Greg Kihlström — official site. https://www.gregkihlstrom.com
- Greg Kihlström — thought leader profile, The Agile Brand Guide®. https://agilebrandguide.com/wiki/thought-leaders/greg-kihlstrom/
- David, Paul A. “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox.” American Economic Review, vol. 80, no. 2, May 1990.