← Home Notes from the Jagged Frontier
01
The Production Readiness Stack
Four layers, all necessary, none optional

Teams build AI systems top-down, starting with the use case and the application. Systems fail bottom-up, starting with infrastructure assumptions nobody stress-tested during the pilot. The Production Readiness Stack names each layer and makes the dependency explicit.

L4
Application & Industry Workflows
Where value is delivered. Where most teams start. Where most pilots live permanently.
L3
Data & Retrieval Layer
Context, lineage, freshness, trust fabric. The connective tissue between AI and the system of record.
L2
Governance & Trust Platform
Identity, authority, audit. Engineered in from the start — not bolted on after the first finding.
L1
Infrastructure
Reliability, latency, data residency. Non-negotiables that determine what can run at all.

The production readiness gap almost always opens between layers: a governance assumption the infrastructure cannot support, a retrieval design that skips identity enforcement, an application scoped without knowing where the execution boundary sits. Build all four before you ship any one.

See the full essay: The Production Readiness Gap →

02
The Execution Boundary
Where model output meets humans, rules, and systems of record

Every enterprise AI workflow has an execution boundary — the point where the model's output hits a human who must approve, a rule that enforces a constraint, or a system that owns identity, permissions, lineage, or SLAs. This boundary is not optional. It determines the architecture.

Event
Model
Rules / Constraints
System / Human
Action
New Event

AI can recommend, classify, draft, route, and summarize. It almost never completes an enterprise workflow end-to-end. Companies that go from pilot to production understand this early — they design around the execution boundary, not around the model.

Healthcare
Clinical review — AI flags, human owns the action. PHI never leaves the boundary.
Edge · Federated
Finance / Trading
Every AI-influenced decision must be explainable, reproducible, and auditable.
Deterministic core · HitL
Manufacturing
Edge latency and physical safety gate every actuator action. PLCs control all timing.
Edge · Safety rules
Enterprise ERP
SOX, identity, transactional consistency, approval workflows override all model output.
Governance fabric · HitL
03
Enterprise AI Operating Model
The system that turns AI capability into production outcomes

Models and tools are not the limiting factor in enterprise AI anymore. The constraint is the operating system built around them: capital allocation, platform infrastructure, production workflows, telemetry, adoption systems, and ecosystem leverage. Four components are required — and all four must be working simultaneously.

01 — Telemetry & Intelligence
Connect usage to workloads
Product usage data linked directly to customer workloads. Visibility into which services drive growth and which customers have high expansion probability. Without telemetry, you are guessing.
02 — Organizational Alignment
Shift incentives to adoption
Incentives aligned to adoption milestones, not bookings. Teams trained to architect workload chains, not just deploy first instances. The compensation model has to point at the same metric as the product roadmap.
03 — Repeatable Implementation
Reference architectures at scale
Standardized solution playbooks and reference architectures that let customers move from proof of concept to production without starting from scratch each time. Repeatability is what makes the economics work.
04 — Operational Cadence
Weekly workload growth reviews
Right metrics identified, executive alignment achieved, and a consistent review cadence focused entirely on workload growth. The cadence creates the discipline. The discipline creates the compounding.

See the full essay: From Consumption to Outcomes →

04
Capacity → Consumption Yield Model
Capital allocation against revenue conversion velocity

Enterprise AI infrastructure planning typically starts from the supply side: how much capacity do we need, where do we deploy it, how fast do we build? The right question is different: where will deployed capacity generate durable, compounding revenue yield within an economically rational time horizon?

Capital Yield Formula
Capital Yield = (Projected 5-Year Regional Recurring Revenue) ÷ (Deployed Infrastructure CapEx)

Five factors determine the yield of any infrastructure deployment — and all five must be assessed before capital is committed.

1
Revenue Conversion Velocity
Speed at which deployed capacity converts into recognized recurring revenue. Tracks the lag between infrastructure availability and production workload adoption.
2
Pipeline Maturity Index
Validated late-stage enterprise demand in the target region. Anchors deployment decisions to revenue-backed demand rather than speculative TAM.
3
Regulatory & Data Sovereignty Trajectory
Compliance lead times and architecture localization cost. Prevents capital from being trapped behind compliance bottlenecks.
4
Latency & Workload Fit
Whether regional demand requires high-performance AI accelerators or standard infrastructure. Protects gross margin by aligning infrastructure cost to workload economics.
5
Competitive Density & Ecosystem Readiness
Integrator presence, field enablement strength, competitor saturation. Determines whether the region can generate yield immediately or requires subsidy.
05
Constraint Augmentation Matrix
Match the architecture to the constraint — industry by industry

The right AI architecture is the one that survives your hardest constraint. Different industries hit different walls first. This matrix maps the primary constraint in each domain to the augmentation paths that actually solve for it.

Vertical Primary Constraint Edge AI Accelerators Neuromorphic Sparsity
Enterprise ERP / Supply ChainReal-time optimization, integration, trustHighHighEmergingMedium
Healthcare — DiagnosticsPrivacy, PHI, latencyHigh · PivotalMediumEmergingMedium
Healthcare — GenomicsData volume, compute complexityLowHighLowHigh
Finance / TradingUltra-low latency, determinism, auditabilityHigh · Co-locationHigh · FPGAsExperimentalMedium
Manufacturing / RoboticsMillisecond timing, safety, edgeHigh · MandatoryHighHigh · EmergingMedium
Recommendation SystemsScale, real-time personalizationMediumHigh · CustomLowHigh · Native
Telecom — 5G / 6GNetwork complexity, real-time QoSHigh · MECMediumHigh · EmergingMedium

See the full research: Beyond the Ceiling — Whitepaper →

06
Agent Portfolio Go/No-Go

Three-step filter before you commit to building agents

Most agent projects fail not because the technology does not work but because the conditions for it to work were never in place. The Go/No-Go filter asks three questions before a single line of code is written: does the workflow have enough business value to justify the build, does the data required to run the agent actually exist and connect, and is the organization ready to trust and act on agent output?

All three gates must pass. A workflow that passes value but fails data quality will produce broken outputs at speed. A workflow that passes both but fails org readiness will sit unused. The filter exists to save months of engineering on the wrong bets.

Gate 1
Business value
Is the workflow worth automating? Does the value justify the build and maintenance cost?
Gate 2
Data quality
Is the data reliable and connected? Garbage in, garbage out — at agent speed.
Gate 3
Org readiness
Will the organization trust and act on the output? An unused agent is a failed agent.

See the full framework: Agent Portfolio Go/No-Go →

07
The GTM Motion Diagnostic

Revenue outcome to workflow cause to operational lever to EBITDA impact

Most GTM problems are not what they look like on the surface. A pipeline conversion problem is often an ICP problem in disguise. A churn problem is often a pricing and packaging problem that was never addressed at the point of sale. The diagnostic exists to find the real source of the leak before you spend time and capital fixing the wrong thing.

The framework examines four areas: ICP and segmentation, pipeline conversion, expansion and NRR, and pricing and packaging. For each: what is the revenue outcome at risk, where is the workflow leaking, what lever fixes it, and what is the EBITDA impact?

ICP + Segmentation
Better targeting raises conversion on the same pipeline. Lower CAC. Better sales productivity without adding headcount.
Pipeline Conversion
Higher conversion increases revenue without adding headcount. Better qualification reduces wasted pursuit cost.
Expansion + NRR
NRR compounds ARR. Expansion is more efficient than new logo growth. Better retention protects gross profit.
Pricing + Packaging
Revenue lift at near-zero incremental cost. Less discount leakage. Monetization discipline directly improves gross margin.

See the full framework: The GTM Motion Diagnostic →  |  Framework at a Glance →

08

Product Health Diagnostic

Where the product system breaks, and which lever to pull first

Product organizations that struggle to grow revenue usually have the same underlying problem: the system is broken in more than one place, and the breaks compound each other. A weak discovery process produces unclear requirements. Unclear requirements slow delivery. Slow delivery delays adoption. Weak adoption produces shallow feedback. Shallow feedback makes the next roadmap worse.

The diagnostic covers six areas. Start where the pain is loudest.

Strategy and Discovery
Are PMs spending time in the customer workflow? Is the roadmap tied to a specific business outcome? Are priorities changing from evidence or from pressure?
Build and Velocity
What share of engineering time goes to new features versus maintenance? Is tech debt tracked formally? Are the same issues repeating across every new installation?
Portfolio and Investment
Which bets should be stopped? Are zombie products alive only because nobody made the call? Is the build versus buy versus partner decision made explicitly?
Validation and Adoption
Are pilots designed to mirror production? Is success defined before the pilot starts? Are performance, stability, and security blocking go-live after a successful pilot?
Feedback and Iteration
How does feedback actually reach the roadmap? Does every piece of feedback get a response? Is win/loss analysis used to influence product decisions?
Telemetry and Intelligence
Do we know which features customers actually use? Can at-risk accounts be identified before they notify us? Is competitive intelligence commercial, not just feature comparisons?

See the full framework: The Product Health Diagnostic →  |  Framework at a Glance →