On this page
Workflow Classification: Mapping Your AI Touchpoints by Complexity
Step 0 in the Cromus.ai Measurement Framework
The Most Expensive Mistake in AI Deployment
The most common and most expensive AI deployment mistake is using a frontier model for a task that doesn't need one. It's invisible: the workflow runs, the output looks fine, the API bill arrives — and no one questions whether a lighter model would have produced the same result for 10% of the cost.
This isn't a hypothetical. It's the default. Most organizations deploy a single model tier across all workflows because it's simpler than classifying tasks by complexity. The result is systematic over-provisioning — frontier-level reasoning applied to formatting, extraction, and tagging tasks that a lightweight model handles identically.
Workflow Classification makes the question answerable before the bill arrives. It maps every AI touchpoint to the model tier it actually requires — so the right capability handles every task, and nothing more.
20–50x
A frontier model running a formatting task costs approximately 20–50x more than a lightweight model producing the same output. (Illustrative estimate)
What Is Workflow Classification?
Workflow Classification is the process of mapping every AI-assisted task in your organization to the model tier it actually requires — based on reasoning complexity, output stakes, latency tolerance, and cost sensitivity.
Classification comes before measurement. You cannot set a meaningful Baseline Cost per Workflow without first knowing what model tier is appropriate for the task. A workflow running on the wrong tier doesn't have a meaningful baseline — it has a distorted one. Classification establishes the frame; baseline measurement fills it in.
Classification is an ongoing discipline, not a one-time audit. Models change. New tiers arrive — the Mythos/Capybara tier is the most recent example. Tasks evolve as teams expand what they delegate to AI. A classification that was correct six months ago may be generating Croms today.
The Four Model Tiers
Tier 1 — Mythos/Capybara (Breakthrough Reasoning)
Representative model: Claude Mythos (Capybara tier)
- •Advanced cybersecurity vulnerability analysis
- •Complex multi-domain research synthesis
- •Novel software architecture design
- •High-stakes legal or regulatory reasoning
- •Tasks where Opus-level output is measurably insufficient
Deploy when: The task requires reasoning capability beyond what frontier models reliably produce, and the output stakes justify the cost premium. Currently in limited early access.
Read the Mythos Benchmark FrameworkTier 2 — High-Reasoning Frontier (Complex Multi-Step Tasks)
Representative model: Claude Opus
- •Research synthesis across multiple sources
- •Complex code generation and debugging
- •Legal/medical document analysis
- •Multi-step strategic analysis
- •Novel problem-solving requiring judgment
Deploy when: The task requires multi-step reasoning, synthesis across sources, or judgment calls where errors are costly and a lighter model produces measurably worse output.
Tier 3 — Mid-Tier Balanced (Structured Generation)
Representative model: Claude Sonnet
- •Long-form drafting
- •Content summarization
- •Nuanced classification
- •Data interpretation with context
- •Customer-facing generation at moderate quality bar
Deploy when: The task requires coherent, contextually aware output but not deep multi-step reasoning; quality matters but so does cost.
Tier 4 — Commodity Lightweight (Deterministic Tasks)
Representative model: Claude Haiku
- •Text formatting and cleanup
- •Structured data extraction
- •Simple Q&A with factual answers
- •Rule-based tagging and classification
- •Translation of straightforward content
Deploy when: The task is deterministic or near-deterministic, high-volume, latency-sensitive, or cost-critical; a lightweight model produces acceptable output.
The Four Classification Dimensions
Reasoning Complexity
How much multi-step judgment and synthesis does the task require?
Output Stakes
What is the cost of an error in this workflow's output?
Latency Tolerance
How much does response speed matter relative to accuracy?
Cost Sensitivity
Is per-call cost critical at the volume this workflow runs?
The Workflow Classification Matrix
| Reasoning | Stakes | Latency | Cost Sensitivity | Recommended Tier |
|---|---|---|---|---|
| 4 | 4 | 1–2 | 1–2 | Mythos/Capybara |
| 4 | 3–4 | 2–3 | 1–2 | High-Reasoning Frontier |
| 3 | 2–3 | 2–3 | 2–3 | Mid-Tier Balanced |
| 1–2 | 1–2 | 3–4 | 3–4 | Commodity Lightweight |
This matrix is a starting guide. Always validate with your actual baseline cost data. See Baseline Cost per Workflow.
What Misclassification Costs You
Over-Provisioning: Frontier Doing Commodity Work
A formatting workflow running on a frontier model at $0.04/execution when a lightweight model would cost $0.002/execution. At 10,000 executions/month, that's $380/month in pure Cost Waste — before human review or rework is counted.
This generates Croms in the Cost Waste dimension (35% weight).
Under-Provisioning: Lightweight Doing Frontier Work
A research synthesis workflow routed to a lightweight model to save cost. Error rate increases from 5% to 40%. Human review rate jumps. Rework time doubles. The per-execution cost appears lower on the API bill and dramatically higher on the fully-loaded Baseline Cost.
This generates Croms in the Failure Risk dimension (25% weight).
Both directions of misclassification increase TCWO. Classification is how you prevent both.
How to Classify Your Workflows
- 1
Inventory your AI touchpoints
List every place in your organization where an AI model is called — including automated pipelines, user-facing features, and internal tools. Most teams undercount by 30–50% on first pass because they forget internal tools, one-off scripts, and embedded model calls in third-party platforms.
- 2
Score Reasoning Complexity
Rate 1–4: Does the task require multi-step judgment and synthesis (4), structured generation with nuance (3), simple classification with rules (2), or deterministic extraction (1)? Be honest — most tasks score lower than teams expect.
- 3
Score Output Stakes
Rate 1–4: What is the cost of an error? Mission-critical with legal or financial consequences (4), significant rework required (3), easily corrected (2), or inconsequential (1)? This dimension often determines whether over-provisioning is justified.
- 4
Score Latency Tolerance
Rate 1–4: How much does speed matter? Real-time user-facing response required (4), moderate latency acceptable (3), batch processing acceptable (2), or asynchronous overnight acceptable (1)? High latency tolerance opens the door to heavier models.
- 5
Score Cost Sensitivity
Rate 1–4: Is per-call cost critical? High-volume commodity task where cost scales dangerously (4), moderate volume (3), low volume where quality dominates (2), or one-off task (1)? Volume is the multiplier that turns small per-call differences into significant budget impact.
- 6
Map to the four-tier matrix
Combine scores to identify the appropriate model tier. Use the classification matrix above as a starting guide, then validate with actual baseline cost data. The goal is not perfection on the first pass — it's visibility. You can refine as you measure.
How Workflow Classification Connects to the Full Framework
Workflow Classification
Map every AI touchpoint to the model tier it actually requires.
(This page)
Baseline Cost per Workflow
Once you know the right tier, measure what each workflow actually costs today — fully loaded.
Croms
Score each workflow across Cost Waste, Latency Overhead, Failure Risk, and Structural Gaps to identify where waste is occurring.
TCWO
Calculate the financial cost of that waste over time — the number your CFO needs to make AI investment decisions.
You cannot optimize what you haven't mapped. Workflow Classification is where it starts.
Frequently Asked Questions
Map Your Workflows. Match Your Models.
Most AI waste is invisible because nobody asked the first question: does this task actually need this model? Cromus.ai makes classification, measurement, and optimization continuous — not a one-time audit. The first Workflow Intelligence as a Service platform.