Operating System Skill · v1.0

AICThe AI Corporation

Not a normal company that uses AI tools. A lean human core surrounded by agents, workflows, dashboards, reviewers, community, and governed execution. This is the playbook for building one.

AIC = MTP Constitution + Agent Fleet + Workflow Interfaces + Live Dashboards
      + Human Review + Community Loops + Governance + FinOps
Core
Small, smart, human
Edge
Large, agentic, leveraged
Stack
12 operating layers
Scarce resource
Trusted coordination
01 · Thesis

Built for abundance, not scarcity.

The old organization was built for scarcity. The exponential organization was built for abundance. AIC updates the source of that abundance, from cloud and crowds to on-demand intelligence and tool-using agents.

The old abundance

Cloud, mobile, platforms, APIs, gig labor, social networks, crowds, shared assets. The leverage of the last era.

The new abundance

On-demand intelligence, tool-using agents, human specialist clouds, automated research and execution, machine-readable knowledge, real-time dashboards, AI-assisted experimentation.

The scarce resource is no longer information
The scarce resource is trusted coordination.

The shape of the next company.

Small
at the core
Large
at the edge
Smart
in the middle
Fast
in experiments
Careful
in governance
Relentless
on economics
02 · Non-negotiables

The hard rules.

Speed multiplies execution, so the guardrails come first. Break these and the system turns fast danger into a feature.

R1
Do not overbuild
Manual service → AI-assisted → repeatable workflow → internal tool → client dashboard → platform. In that order.
prove the workflow first
R2
Do not sell vague AI
Name the job, the buyer, the inputs, the workflow, the review, the cost, and the measurable outcome.
outcome over technology
R3
Keep human judgment where trust matters
Regulated advice, strategy, public claims, brand and safety-sensitive output, and consequential decisions stay human-reviewed.
trust layer required
R4
Do not invent market proof
Research current claims. Label every assumption as an assumption until verified.
evidence or it is a guess
R5
Do not ignore unit economics
Cost per run, cost per accepted output, cost per client, gross margin, review time, error and rework rate.
cost per accepted output
R6
Do not confuse autonomy with chaos
Bound it with purpose, roles, authority limits, escalation, audit logs, approval gates, and kill switches.
bounded authority
R7
Do not let agents run blind
Every agent gets a mission, allowed tools, forbidden actions, required I/O, review, evidence, memory, cost limits, and failure handling.
no charter, no run
03 · Architecture

The twelve-layer stack.

Define the company as an operating system, not an org chart. Each layer ships a first version, an owner, a tool, and a metric.

01
MTP Constitution
Purpose, rules, principles, decision filters, and evidence standards.
machine-readable soul
02
Sensing
Market, customer, competitor, regulatory, and technology signals.
live, not static
03
Intake
Forms, emails, calls, uploads, APIs, CRM entries, chat, internal requests.
structured inputs
04
Knowledge
Company memory, customer data, sources, style guides, SOPs, past work, facts.
retrievable memory
05
Agent
Specialized AI agents with defined roles, tools, authority, and outputs.
narrow, governed
06
Interface
Workflows, routers, APIs, MCP servers, approvals, automations, handoffs.
the control plane
07
Human Review
Editors, strategists, experts, QA, legal when needed, final approvers.
the trust layer
08
Dashboard
Live metrics for work, quality, cost, speed, revenue, risk, and learning.
the manager
09
Experiment
Hypotheses, tests, variants, success thresholds, learning archives.
cheap tests over opinions
10
Community
Customers, contributors, expert network, advisors, creators, testers, fans.
contribution engine
11
Governance
Risk controls, permissions, audits, privacy, security boundaries, escalation.
trust as a feature
12
FinOps
AI, cloud, and tool spend, cost per output, gross margin, utilization.
margin by offer
04 · The Workforce

The agent fleet.

Agents are not random chatbots. They are role-specific operating units with defined authority, tools, memory, and review rules. Narrow agents beat magical ones.

ST
Lvl 2
Strategy
Turns goals into plans, tradeoffs, briefs, and roadmaps.
RE
Lvl 2
Research
Gathers evidence, sources claims, tracks competitors and signals.
CU
Lvl 1
Customer
Analyzes pain, feedback, personas, journeys, objections, support.
PR
Lvl 1
Product
Defines specs, priorities, MVPs, user stories, acceptance criteria.
WF
Lvl 2
Workflow
Maps operations, handoffs, automations, bottlenecks, and SOPs.
DA
Lvl 1
Data
Handles schemas, data quality, source lineage, reporting inputs.
CR
Lvl 1
Creative
Creates copy, concepts, visual direction, campaigns, brand voice.
EN
Lvl 1
Engineering
Drafts technical plans, scaffolds, API plans, data flows, specs.
QA
Lvl 2
QA
Checks outputs against requirements, facts, brand, and risk rules.
GV
Lvl 2
Governance
Checks policy, permissions, sensitive data, compliance, escalation.
SA
Lvl 1
Sales
Builds outreach, proposals, qualification, CRM updates, follow-up.
FI
Lvl 2
Finance
Tracks cost, revenue, margins, forecasts, pricing, AI spend.
05 · Bounded Agency

Six levels of authority.

No agent should have more authority than the company can monitor. New agents default to Level 1 or 2 until proven safe.

L0   No agencyanswers questions only
L1   Draftingdrafts only; cannot send, publish, spend · DEFAULT
L2   Recommendationrecommends with reasoning; humans decide · DEFAULT
L3   Controlled executionexecutes bounded tasks after approval
L4   Limited autonomylow-risk tasks within predefined limits
L5   Full bounded autonomycontinuous, governed domain, with logs, alerts, kill switches
06 · The Operating System

How work moves.

AIC companies win by converting messy requests into repeatable workflows. Design the common path first, make status visible, and capture learning on every run.

1 Intake
→
2 Clarify
→
3 Source
→
4 Route
→
5 Agent execution
→
6 Human review
→
7 Revise
→
8 Deliver
→
9 Feedback
→
10 Dashboard
→
11 Capture learning
→
12 Improve template
The moat
Models will keep changing. The defensible layer is the workflow, not the model: customer knowledge, proprietary examples, evaluation criteria, expert review, distribution, trust, data feedback loops, and community. Every run should improve the system.
07 · Monetization

Offer architecture.

Turn the operating model into something a buyer can purchase. Start with a service. Software comes only after the workflow repeats.

STEP 01
Diagnostic
A fast, low-friction entry product that reveals the problem and builds trust. An audit or opportunity scan.
creates trust
STEP 02
Implementation
Done-for-you or done-with-you build that fixes the highest-value problem. The first workflow or agent team.
fixes the pain
STEP 03
Managed operating
A recurring service that runs the system: monitoring, optimization, reporting, and experiments.
recurring revenue
STEP 04
Platform
Only after repeatable demand exists. Client portal, workflow library, agent marketplace, self-service.
after repetition
08 · Mission Control

The dashboard is the manager.

Dashboards are not decoration. They are the management system. Start with decisions, not vanity metrics. Every metric needs an owner and an action.

Company
  • Revenue
  • Gross margin
  • Retention
  • Active clients
Workflow
  • Cycle time
  • Error rate
  • Rework rate
  • SLA performance
Agent
  • Acceptance rate
  • Escalations
  • Cost per run
  • Tool failures
Experiment
  • Active tests
  • Win rate
  • Decisions made
  • Learnings
Trust & Governance
  • Approval exceptions
  • Risk flags
  • Policy violations
  • Audit gaps
FinOps
  • AI spend
  • Cost / deliverable
  • Margin by offer
  • Unused capacity
09 · Keep speed from becoming danger

Governance and risk.

Agents multiply execution capacity, so governance is needed from day one. Trustworthy operation is not back-office compliance. It is part of the offer.

Risk surface

Data privacySecurityHallucinationBad adviceBrand damageRegulatory claimsBiased outputIP misuseRogue spendData leakageModel driftPrompt injectionVendor lock-inCost sprawl

Controls

  • Role-based access and tool permissions
  • Sensitive data rules and approval gates
  • Audit logs and escalation paths
  • Incident response and vendor review
  • Data retention rules and model evaluation
  • Customer disclosure rules where needed
⏻KILL SWITCH · every governed agent ships with one.
10 · Go / No-Go

The evaluation scorecard.

Score each idea from 1 to 5 across ten criteria. The total tells you whether to build, test, narrow, or wait.

Criterion
Score 1
Score 3
Score 5
Pain intensity
Mild
Annoying
Urgent & expensive
Buyer clarity
Unknown
Likely buyer
Obvious budget owner
AI advantage
Cosmetic
Helpful
Changes cost or speed
Trust feasibility
Risky
Manageable
Strong review path
Workflow repeatability
Custom each time
Partial
Highly repeatable
Data availability
Weak
Some data
Strong inputs & feedback
Speed to first test
Months
Weeks
Days
Revenue path
Unclear
Plausible
Immediate paid wedge
Moat potential
None
Some learning
Data, trust, community
Cost control
Unknown
Trackable
Strong unit economics
40–50 · Strong AIC candidate
30–39 · Test before building
20–29 · Narrow the wedge
Below 20 · Do not build yet
11 · Execution

The 90-day roadmap.

Clarity, then workflow, then sandbox delivery, then package, then sell, then systemize. Do not scale acquisition until delivery works.

DAYS 1–15
Clarity
  • MTP Constitution
  • Customer & wedge
  • Offer & metrics
  • Governance rules
DAYS 16–30
Workflow
  • First workflow
  • Agent charters
  • Intake & delivery
  • Dashboard v1
DAYS 31–45
Sandbox
  • Recruit partners
  • Deliver real output
  • Track cost & quality
  • Capture examples
DAYS 46–60
Package
  • Productize workflow
  • Set pricing
  • Proof assets
  • Onboarding
DAYS 61–75
Sell
  • Outbound tests
  • Content tests
  • Referral tests
  • First paid clients
DAYS 76–90
Systemize
  • Automate repeats
  • Deepen dashboards
  • Add staff on demand
  • Decide platform path
12 · Avoid

Common failure modes.

The same eight mistakes sink most AIC builds. Each has a one-line fix.

Platform fantasy
Building software before proving anyone wants the workflow.
FixStart as a service.
Agent sprawl
Too many agents with overlapping jobs.
FixMap the workflow first, then assign agents.
No review gate
AI output reaches customers without accountability.
FixDefine review gates before delivery.
No dashboard
Running the company by vibes.
FixTrack cycle time, quality, cost, revenue, learning.
No buyer
Building for a market instead of a person with a budget.
FixDefine the economic buyer and buying trigger.
No unit economics
Ignoring AI, cloud, tool, and review costs.
FixTrack cost per accepted output.
Generic positioning
Saying "AI-powered" instead of naming the result.
FixSell the outcome, not the technology.
Unbounded autonomy
Humans or agents acting without clear authority.
FixUse authority levels and escalation rules.
13 · Minimum Viable Company

The build kit.

For the first version, build only these twelve pieces. If they are not working, do not build more.

01One MTP Constitution
02One customer profile
03One wedge offer
04One intake form
05One workflow spec
06Three agent charters
07One review checklist
08One delivery template
09One dashboard
10One experiment backlog
11One risk register
12One sales page or script
Do not build an AI company around a tool. Build it around a job that matters, a purpose that coordinates people, and a system that improves every time it runs.
The Final AIC Principle