the diagnostic · outcomes economy

The Bottleneck Taxonomy

A complete, mutually exclusive map of where enterprise value gets trapped: three failure types across five groups, fourteen cells, and one causal bottleneck at a time. Diagnose the right one before you spend a dollar removing it.

By Vin Vashishta · V-Squared AI

Run the diagnostic Then remove it: MAIT →

three failures, three tests

Every trapped dollar is one of three failures, and each has a test.

Name the failure precisely and the remedy names itself. The three run in sequence: a model must exist before it can move, and it must move before execution can align to it. So the tests run in order, and diagnosis stops at the first one that fails.

F1 · don't understand

A modeling failure: no accurate model of the group exists anywhere. Test: would two people, working independently from what the organization actually knows, describe this group's decision logic the same way? If not, F1 is present.

F2 · can't work with

A translation failure: the model exists but hasn't crossed from the person who holds it to the group that owns the decision. Test: does that group act on the model without its author in the room? If not, F2 is present.

F3 · can't align

A coordination failure: the model reached the decision-owner, but strategy, build, and execution still point elsewhere. Test: would a fully informed, fully empowered actor build what's being built today? If not, F3 is present.

The diagnostic rule, run the tests in order and stop at the first failure, sounds procedural. It's the taxonomy's most valuable edge.

Most problems present as alignment problems: the roadmap misses, the numbers disappoint, execution looks sloppy. But run the tests and you'll often find the coordination failure is downstream of a translation failure, because the roadmap drifted when customer understanding never reached it, and that translation failure is downstream of a modeling failure, because the understanding that got stuck was itself a guess. Fixing alignment on top of a missing model just aligns everyone to the guess.

The causal bottleneck is the earliest failed test. Everything after it is noise until that gate opens.

The reverse error matters too. If your AI pilots convert to production at a healthy rate, modeling isn't your constraint, and pouring more money into understanding is fixing an open gate while translation or coordination stays shut. The taxonomy isn't just for finding the problem; it prevents you from overpaying to fix the wrong one.

Five groups. Fourteen cells, not fifteen.

The failures recur across five groups, every population whose model governs whether enterprise value gets created. Technologists own the model of how technology is built and monetized. C-level leaders own the funding decision, and each executive runs a different decision model. Business units own the operating reality that corporate strategy makes assumptions about. Customers own the buying decision, the model most companies describe most vaguely. And the marketplace owns the external dynamics that decide which positions get rewarded.

The marketplace is also the taxonomy's one deliberate hole. There is no "can't work with the marketplace" cell, because a market is a system you read and align to, not a counterparty you work with. Fourteen cells, not fifteen, and the missing one is a claim, not an omission. Every bottleneck except one is ultimately a people-or-process problem; the single exception is the one force you calibrate.

the evidence · priced and documented

Each failure has a price tag, and a cautionary tale.

Three decades of research converge on the same point: these failures are almost never technological. They're modeling, translation, and coordination.

modeling failure

MIT's Project NANDA found roughly 95% of enterprise GenAI pilots delivered no measurable P&L impact in 2025, a directional but blockbuster snapshot, and RAND put AI-project failure near 80%, about double conventional IT, with organizational rather than technical causes leading. Blockbuster is the story: its model of the customer was built on the late fee (~$800M in 2000), so it couldn't imagine a customer who hated the fee more than they loved the store. Netflix could.

translation failure

PMI priced it at $75M of every $135M at risk per $1B of project spend, put on the line by poor communication alone. Nokia is the story: its engineers knew Symbian was inferior, but organizational fear kept the truth from traveling upward while ~90% of the company's value evaporated. Boeing's MCAS risk never reached regulators or pilots, costing 346 lives and over $20B.

coordination failure

McKinsey found even high-performing companies leave ~30% of their strategy's value undelivered, tracing the gap to the operating model rather than the strategy. Target Canada is the story: strategy, an untested supply chain with up to 70% inaccurate product data, and execution never pointed the same way, roughly $2B gone, shelves empty above full warehouses.

from symptom to cause

Three steps to the causal bottleneck.

Diagnosis is mechanical once you run it in order.

Locate the group

Identify which of the five groups is producing the most friction: a decision that keeps stalling, a build that keeps missing, a relationship that costs more than it returns.

where is the friction

Run the tests in order

Apply the independence test, then the decision-owner test, then the informed-actor test. Stop at the first failure. That cell is the causal bottleneck.

stop at the first failure

Hand off to MAIT

The causal cell defines what MAIT's Model step must build: a shared model accurate enough to pass the first test, transferable enough to pass the second, operationalized enough to pass the third.

diagnosis → removal

the evidence · earnings calls, july 2026

The July 2026 calls read as a diagnostic sheet.

The taxonomy locates value in one of three failures across five groups. Run the July earnings calls through it and each of the largest technology companies is visibly working a specific cell, which is a useful demonstration that the cells are real rather than tidy.

MicrosoftFY26 Q4 · Jul 29, 2026

F2 translation, addressed structurally rather than rhetorically

Frontier Company embeds 6,000 industry and engineering experts inside customer organizations to co-design and continuously improve AI systems, after a year of piloting across 330 projects and 164 customers. Separately, time from license purchase to high usage fell from months to days.

what it confirms

The understanding existed and was not crossing to the group that owned the decision, which is the definition of the translation cell. Microsoft did not fix it with documentation or enablement content. It fixed it by putting people who hold the model in the room where the decision gets made.

MicrosoftFY26 Q4 · Jul 29, 2026

F1 modeling, applied to the marketplace group

Nadella told analysts that an enterprise cannot depend on any one model and cannot be subject to the refusal of a single model, framing multi-model architecture as business continuity. Zuckerberg made a parallel argument that other companies do not want to rely on a small number of closed labs.

what it confirms

Model availability is a marketplace dynamic, and most enterprises have no written model of it at all. That is the modeling failure against the marketplace group, and it stays invisible until an access decision made by somebody else lands in the middle of a production system.

AlphabetQ2 2026 · Jul 2026

F1 modeling, applied to the customer group

Pichai's answer on defensibility was that the model is an ingredient rather than the solution, and that what customers need is their own data and trajectories kept confidential. Gemini Enterprise now reaches roughly 90% of the Fortune 100.

what it confirms

The vendor is describing the customer's real decision criterion, which is control over the learning loop rather than benchmark position. Any enterprise whose written model of its own buyers still says they choose on capability is carrying a modeling failure in the cell that costs the most.

Sources: Alphabet Q2 2026, Microsoft FY26 Q4, and Meta Q2 2026 earnings calls and releases, plus Amazon Q2 2026 results reported July 30, 2026. Figures are as stated by company executives on those calls. Amazon reported after market close on July 30, so Amazon figures here come from the release and initial call remarks rather than the full transcript.

the wider evidence · each failure type, priced separately

Each failure type has its own separate literature.

The taxonomy is only useful if the three failure types are genuinely different rather than three descriptions of one problem. The best support for that is that each type accumulated its own body of evidence, measured by different people using different methods.

peer-reviewedsurvey researchcompany filingsmarket dataindependent audit

F1 · modelingsurvey research6,000+ executives · 625 leaders

The people funding the technology aren't operating it

A survey of over 6,000 senior executives found nearly 70% using AI less than an hour a week, including 28% who never use it. BCG's May 2026 survey of 351 CEOs and 274 board members found around 75% of directors rating their own AI understanding at or above their peers, an assessment their CEOs didn't share, with the least confident directors most likely to feel the organization was moving too slowly.

Sources: executive AI usage survey reported March 2026; BCG Split Decisions, May 2026.

what it confirms

Both measure the same thing from different angles. No accurate model of the technology exists at the level where funding gets approved. The independence test fails here immediately, and everything downstream inherits the error.

F2 · translationmarket dataMay to July 2026

Five vendors decided the gap has to be staffed

Palantir built the embedded engineering model and reported 85% revenue growth in Q1 2026. Anthropic and OpenAI announced enterprise deployment ventures within hours of each other on May 4. AWS committed $1 billion to a Forward Deployed Engineering organization on June 30, structured around shared business outcomes instead of billable hours. Microsoft announced 6,000 embedded experts in July.

Sources: company announcements and contemporaneous coverage, May to July 2026.

what it confirms

Translation is the only one of the three failures a vendor can sell you a remedy for, which is why it's the one attracting capital. The remedy is people placed inside your decision process, which is exactly what the discriminating test for this cell asks about.

F3 · coordinationsurvey research895 transformations · 1,200+ managers

Value gets created and then fails to hold

BCG's assessment of 895 transformation programs found 44% created some value but missed their targets and produced only limited lasting change, with another 26% producing no sustainable change at all. McKinsey surveys of more than 1,200 managers found 20% saying their organizations excel at decision making, and high quality decisions reported at 70% in organizations with one to three reporting layers against 45% at seven or more.

Sources: BCG digital transformation research; McKinsey Global Surveys on decision making.

what it confirms

Coordination failures show up in outcomes rather than intentions, which is why they get misdiagnosed as weak execution or missing commitment. Keep the reporting layer gradient. Alignment decays with every structural handoff no matter how good the strategy was.

These three bodies of evidence were produced independently, by different researchers, using different methods, at different times. They map onto three separate cells, which is the argument for the taxonomy being a real partition rather than a convenient one.

The taxonomy classifies a state, not a permanent condition

Removed bottlenecks reopen. Customers shift, executives turn over, markets move, so every model decays back toward "don't understand" and every alignment drifts back toward "can't align." That's why removal is a cycle, not an event, and why the diagnosis hands off to a method built to run continuously.

from diagnosis to removal · mait

The taxonomy names the bottleneck. MAIT removes it.

Diagnosis is the input to the cycle. MAIT, meaning Model, Apply, Improve, and Transform, builds the shared model the diagnosis calls for and turns it into a compounding advantage.

Go to MAIT

one causal bottleneck at a time

Every cell, correctly named, is a located, priced, removable reason the next stage of growth hasn't happened yet.

Not a diagnosis of what's wrong with the business, but a map of where the value is waiting.

Run the diagnostic V-Squared advisory