The Capability Matrix — What AI Can and Can't Do (and Why It's Not About Capability)

Most organizations delegate work to AI the way they delegate work to a new hire: they look at what the system seems capable of and hand it a task. The Capability Matrix replaces that intuition with structure. It is the Grammar of the Language of Work — a two-dimensional map that crosses nine work verbs against four platform types, producing a clear, auditable answer to the question every executive should be asking before deployment: can this platform legitimately perform this verb?

The answer is not always about technical capability. In several critical cells, the constraint is structural — rooted in accountability, governance, and the physics of how work actually gets done. Ignoring those constraints does not produce innovation. It produces liability.

The four platform types

The Grammar of Work recognizes four platform types that can execute work. Understanding what each platform is matters because it determines what each platform can do — and, more importantly, what it cannot.

Humans remain the only platform that bears accountability. A person can be deposed, subpoenaed, fired, or promoted. No other platform can. This is not a sentimental observation; it is a legal and organizational fact that constrains every cell in the matrix.

Prediction Machines — large language models, recommendation engines, image generators, and other statistical inference systems — operate by identifying patterns in training data and projecting those patterns forward. They are extraordinarily powerful at interpolation. They are structurally incapable of bearing responsibility for what they produce.

Logic Machines — rules engines, RPA bots, deterministic software, and traditional automation — execute predefined instructions with perfect fidelity. They do exactly what they are told, which is both their strength and their fatal limitation. They cannot handle ambiguity, and they do not degrade gracefully when inputs deviate from expectations.

Matter Machines — robots, autonomous vehicles, drones, industrial actuators — operate in the physical world. They sense, move, and transform physical materials. Their constraints are dictated by physics: latency, precision, environmental variability, and the irreversibility of physical action.

The matrix

The Capability Matrix maps nine work verbs against these four platforms. Each cell contains one of four values: YES (the platform can legitimately perform the verb), NO (it structurally cannot), PARTIAL (it can contribute but cannot own the outcome), or LIMITED (it can perform the verb only under narrow conditions).

Read the matrix column by column and the profile of each platform becomes clear. Humans can do everything but struggle to sustain monitoring. Prediction Machines cover a wide range but cannot decide or formulate on their own. Logic Machines are reliable within their lane but blind outside it. Matter Machines dominate actuation and monitoring but falter wherever context or judgment is required.

Read it row by row and you see something different: which verbs are contested, which are shared, and which belong to a single platform. That row-level view is where the real strategic insight lives.

Four cells that define the boundaries

Most of the matrix is intuitive. Four cells are not, and misunderstanding any one of them is enough to sink a deployment.

DECIDE is human-only

DECIDE appears as YES for humans and NO for every other platform. This is the matrix’s most important cell and its most frequently violated constraint. The restriction is not based on a capability gap that future AI might close. It is based on accountability. Decisions create consequences. Consequences require someone who can be held responsible. A prediction machine can recommend, score, rank, and draft — but the moment an organization treats that output as a decision, it has created an accountability gap with no one standing behind the outcome.

This constraint holds regardless of the decision’s apparent simplicity. An algorithm that automatically rejects loan applications is not “automating a decision.” It is making a decision with no accountable party, which is precisely the structure that produces regulatory exposure and reputational damage at scale.

INTERPRET is impossible for Logic Machines

Logic Machines receive a NO for INTERPRET because interpretation requires handling ambiguity, and ambiguity is the one thing deterministic systems cannot process. When a rules engine encounters an input it was not designed for, it does not degrade gracefully or ask for help. It either crashes, throws an error, or — worse — applies the wrong rule with perfect confidence. This is the Brittleness Trap: the failure mode where rigidity is mistaken for reliability.

Prediction Machines can interpret because they are designed to operate on probability distributions. They handle ambiguity natively, even if imperfectly. Logic Machines cannot, and no amount of rule-writing changes that structural fact.

MONITOR is where humans are LIMITED

Humans receive a LIMITED for MONITOR — the only cell where humans are not fully capable. The reason is physiological: vigilance decrement. Research consistently shows that human monitoring performance degrades significantly after approximately twenty minutes of sustained attention to a low-event-rate task. This is not a training problem. It is a biological constraint.

Organizations that assign humans to monitor dashboards, audit logs, or quality control screens for extended periods are building failure into their process. The work feels like oversight, but after the first twenty minutes it is closer to theater. The Vigilance Fallacy is the named violation: assigning sustained monitoring to humans and assuming the assignment provides actual coverage.

FORMULATE is PARTIAL for Prediction Machines

Prediction Machines receive PARTIAL for FORMULATE because they can draft, generate, and propose — but they cannot bear accountability for novel creation. The same statistical mechanism that allows a large language model to produce fluent, creative text is the mechanism that produces hallucination. Pattern completion and confabulation are not separate features; they are the same feature viewed from different angles.

A prediction machine can formulate a first draft, generate candidate solutions, or sketch options for human evaluation. What it cannot do is own the output. Formulation that matters — strategy documents, legal arguments, medical treatment plans, creative work with reputational stakes — requires a human who reviews, modifies, and takes responsibility for the final artifact.

How the matrix reveals hidden violations

Consider a mid-size insurance company that has deployed an AI system to process claims. The system ingests claim documents (SENSE), extracts key fields and flags anomalies (INTERPRET), checks the extracted data against policy rules (VERIFY), and routes approved claims to payment processing (ROUTE and TRANSFER). On paper, every step maps to a cell where the platform has a YES.

Now look more closely at what actually happens at the anomaly-flagging step. When the prediction machine encounters an ambiguous claim — a medical procedure code that could indicate either a covered treatment or an excluded cosmetic procedure — it does not flag the ambiguity. It resolves it. It picks the more statistically likely interpretation and moves the claim forward. In practice, the system is not just interpreting; it is deciding which reading of an ambiguous input governs a financial outcome.

The Capability Matrix makes this violation visible. DECIDE is human-only. The system has crossed from INTERPRET (where prediction machines have a YES) into DECIDE (where they have a NO). The fix is structural: ambiguous claims must be routed to a human adjudicator who can weigh context, apply judgment, and bear accountability for the resolution. The matrix does not just identify the problem — it specifies exactly where the human must re-enter the process.

This is the matrix’s primary function. It does not tell organizations what AI can do. It tells them where the boundaries are — which work is AI Handoff Work that can be liberated, which is AI Amplified Work where humans and machines collaborate, and which is Human Reserved Work that must stay with a person. Designing systems that respect those boundaries before a failure teaches the lesson at greater cost is the point.

Named violations

When the matrix’s constraints are broken, the resulting failure modes are predictable enough to name. The Grammar of Work identifies four primary violations:

Accountability Gap. Assigning DECIDE to any non-human platform. The output looks like a decision, functions like a decision, and creates consequences like a decision — but no one is accountable for it. This is the most common and most dangerous violation in enterprise AI deployment.

Brittleness Trap. Assigning INTERPRET to Logic Machines. The system performs flawlessly on expected inputs and fails catastrophically on unexpected ones. Organizations mistake long runs of correct output for robustness, when they are actually observing a system that has not yet encountered its breaking point.

Phantom Authority. Treating prediction machine output as authoritative truth. This violation does not map to a single cell; it is an emergent property of over-relying on PARTIAL and YES cells without maintaining verification loops. When an organization accepts AI-generated analysis without human validation, it has granted the system authority it cannot structurally possess.

Vigilance Fallacy. Assigning sustained MONITOR to humans. The process design assumes continuous human attention, but human biology cannot deliver it. The monitoring looks active but is functionally passive after the first twenty minutes.

The matrix as compiler stage

The Capability Matrix is Stage 1 of the Compiler — the Language of Work’s two-stage validation system for work allocation. The matrix answers the threshold question: is this platform structurally permitted to perform this verb? If the answer is NO, no further analysis is needed. The delegation is invalid regardless of how well the platform might perform technically.

Stage 2 — the Physics of Work — addresses the question of whether the platform can sustain the assignment given its architectural constraints, and whether a better-suited platform is being overlooked. But Stage 1 comes first, because a system that clears physics constraints for a verb it cannot structurally perform is still a violation.

The matrix is deliberately simple. Nine rows, four columns, thirty-six cells. That simplicity is the point. It is a tool designed to be used in operating reviews, architecture decisions, and procurement evaluations — contexts where a framework that requires a PhD to apply will not be applied at all. Executives can read it. Engineers can implement against it. Auditors can verify compliance with it. And when a deployment goes wrong, the matrix provides a shared language for diagnosing exactly which constraint was violated and why.

The Language of Work

The Capability Matrix is the Grammar — one component of a larger system. The Language of Work provides a complete architecture for describing and validating work allocation:

• Vocabulary: The Four Platforms define who performs work. The Nine Verbs define what operations work consists of.
• Grammar: The Capability Matrix (this page) defines which platform-verb assignments are structurally valid.
• Physics: The Physics of Work defines which assignments are sustainable given each platform’s architectural constraints.
• Compiler: The Compiler runs Grammar then Physics as a two-stage validation — catching delegation errors before deployment.

Related Concepts

• The AI Readiness Scale — The three categories of work that emerge from Language of Work analysis

Further Reading

• The Four Kinds of Actors in Hybrid AI Architecture — Names the Accountability Gap and maps which actors can perform which operations
• Speed Is a Trap: Why Armadin’s “Autonomous Defense” Is a CISO’s Nightmare — The Accountability Gap and Probability Problem in action when AI decision-making lacks human oversight

Is your AI deployment violating the Capability Matrix without anyone noticing? Seampoint’s Discovery engagement audits your work allocation against the Grammar and Physics of Work — identifying violations before they become failures.