Our Approach

The eight words of every AI system

AI feels chaotic because vendors rename the same parts. Strategy gets easier when you can see the system underneath.

The Problem

Vocabulary churn makes decisions harder

Too many names for the same thing

Agents, co-pilots, assistants, autonomous systems — the industry invents new vocabulary faster than buyers can learn it. Meanwhile, every AI deployment is built from the same small set of primitives.

Vendors sell the dream, hide the parts

Most vendors operate on only some layers but sell the whole dream. If you can't name the parts, you can't see where the black boxes are — or what will be expensive to change later.

The Insight

The shape of AI systems is stable

Every AI deployment can be decomposed into the same eight words. The hard decisions are what to put into each word, what risk to accept, and what economics to optimize.

The Framework

The eight words defined

Each word represents a decision point. Each decision has cost, risk, and governance implications.

Model

The brain. Trained software that takes input and produces output — answers, scores, predictions, classifications.

Examples: GPT-4, Claude, Gemini, Llama, Mistral

Decision question: Which model's capability and cost profile matches this workflow's requirements?

Gateway

The front door. One controlled entry point for governance, spend tracking, access control, and compliance logging.

Healthcare context: PHI boundaries, audit trails, model routing

Decision question: Where is the control point that routes work to appropriate intelligence?

System Message

Standing instructions. The persona, behavior rules, guardrails, and boundaries encoded into the AI's behavior.

Also called: Playbooks, rules, behavioral policy

Decision question: What behavior is required, what is prohibited, and who governs these rules?

Knowledge

AI-ready encodings of policies, clinical guidelines, payer rules, patient data — your ground truth for retrieval.

Not the same as: Training a model (retrieval is cheaper and more controllable)

Decision question: Which documents and data should the system access, and who governs them?

Tools

Actions the model can take in other systems — schedule, send, update, query, approve.

Healthcare examples: EHR lookup, billing system, scheduling, prior auth

Decision question: What systems can the AI read from or write to, and with what limits?

Orchestration

Structure connecting models, humans, and systems through defined workflows — the choreography of multi-step processes.

Also called: Agents, workflows, multi-turn processes

Decision question: At what points do humans need to review, approve, or take over?

Interface

The surface where people meet the system — chat, voice, forms, dashboards, embedded in existing applications.

Key consideration: Meet users where they work; don't add friction

Decision question: Where do different roles interact with the system, and what do they need?

Training

How models are created and improved — feeding examples until patterns are learned, then evaluating outcomes.

Includes: Fine-tuning, evaluation, monitoring, feedback loops

Decision question: Do you need a custom model, or is retrieval + prompting sufficient?

Impact

How the Eight Words change buying decisions

Vendor evaluation

Identify which layers a vendor controls and hides. Is this a model company, a wrapper, or a system?

Cost control

Route work to the least expensive intelligence that meets the need. Don't use premium models for routine tasks.

Governance

Rules, logs, and measurement live in explicit places. You can locate what you need to audit and govern.

Implementation

Use cases become configurations of a repeatable system. The parts stay stable even as vendors change.

Healthcare Example

An AI-powered patient intake system

How the Eight Words map to a specific use case

Model: Frontier model for ambiguous patient descriptions; smaller model for structured form-filling
Gateway: Routes by request type; logs all PHI access; enforces rate limits and cost caps
System Message: "You are a supportive intake coordinator. Elicit symptoms without diagnosing. Escalate urgent cases immediately."
Knowledge: Provider specializations, insurance coverage rules, location availability — retrieved not memorized
Tools: EHR lookup, scheduling system, insurance verification API — read-only during intake, writes require human approval
Orchestration: Information gathering → provider matching → appointment scheduling → confirmation → handoff to human for complex cases
Interface: Patient-facing chat, staff dashboard for handoffs, embedded in patient portal
Training: Evaluation of match quality, escalation patterns, patient satisfaction; feedback improves prompting and routing

Use the framework on your next AI decision

Bring us the use case or vendor proposal you're evaluating. We'll map it to the Eight Words together.

Start a conversation