Operations, run as a system.

Over two years I rebuilt a real company’s operations around AI. Not a folder of prompts: a six-level operating system with governed access, autonomous loops, and a ledger that prices its own output.

Every figure below is from production, logged the day it shipped.

ops-ledger · six-week windowverified

window:        2026-06-12 → 2026-07-22
shipped:       91 production deliverables
value_range:   $75.9K–$179.3K (rate-card)
reverts:       0
ai_spend:      $200/mo
status:        verified live, per deliverable

Six weeks, measured in production

$75.9K–$179.3K

Replacement value of the six-week window: each deliverable priced against a freelancer and agency rate card, conservative to market. Methodology at the end of this report.

0 reverts

Shipped means verified live: a cache-busted page load, a passing live test, a confirmed status code. Deployed is not done.

$200/mo

Total AI subscription cost behind the output. The spend is not the story. The system is.

Where the ninety-one landed
Development & infrastructure fixes	37
Internal tools built	19
SEO / AI-visibility work	14
Marketing strategy	6
Finance & bookkeeping ops	5
Site redesign	5
Brand & design	3
Copywriting & other	2
Logged deliverables	91

Six levels, built in order

Each level compounds the ones beneath it. The progression matters: context before skills, skills before access, access before autonomy, and measurement closing the loop over all of it.

Level 01 · Context

The company brain

25 context files · 283 memory files

A versioned repository of institutional knowledge: org structure, entity facts, brand rules, decision history, per-project memory. Every AI session starts already knowing the company, its people, its constraints, and its past corrections, instead of starting from zero.

Level 02 · Skills

The skills library

85 packaged skills

Recurring workflows encoded as named, reviewable, runnable procedures: payroll runs, SEO audits, deploy pipelines, client reporting, financial reconciliation. A skill is a workflow the company owns, not knowledge locked in one person’s head.

Level 03 · Access

The integration layer

1 permissioned API server

A purpose-built server (Model Context Protocol) that gives AI agents governed access to production systems: task management, orders, reporting. OAuth in front, a tiered permission model behind it. Agents work with real company data without ever holding the keys.

Level 04 · Routing

Model routing

3 independent model lanes

The right model for each job rather than loyalty to one vendor: a primary lane for operations, a separate cross-vendor lane for adversarial code review, and a router to specialist models for design and analysis. Independent second opinions are part of the architecture.

Level 05 · Autonomy

The agent fleet

17 scheduled autonomous jobs

Loops that run without being asked: a weekly multi-agent revenue radar whose findings are adversarially verified before a human sees them, an inbound-lead watcher that drafts replies but never sends, a monthly client-pulse review, and daily site-health, backup, and configuration-drift monitors that alert on failure.

Level 06 · Measurement

The closed loop

91-entry ledger · 5 dashboards

Every deliverable logged and priced. Token spend tracked. A session-intelligence dashboard that flags stalled work before it is lost. What ships gets priced, what stalls gets surfaced, and the findings feed the next build cycle. Most AI programs cannot say what they are worth; this one prices itself weekly, in dollars.

The command center

One operator at the top, dashboards in the middle, autonomous agents underneath. Escalation flows up; nothing outbound leaves without a human.

operator (human, final say)
├─ command center · 5 live dashboards
├─ integration layer · OAuth · 3 permission tiers
├─ agent fleet · 17 scheduled jobs
│  ├─ revenue radar · weekly · adversarially verified
│  ├─ lead watcher · 2-hour cadence · draft-only
│  ├─ client pulse · monthly
│  └─ site-health / backups / drift · daily
└─ ledger · 91 entries priced · closes the loop

Findings are verified before they surface. The radar’s opportunities pass an adversarial review stage; weak findings die before they cost attention.
Drafts, never sends. Outbound communication is machine-drafted and human-sent, without exception.
Failure is loud. Monitors alert on failure rather than reporting success. Silence means healthy.

Inward, for the team

The work system talks back. Task queues, order status, and workload reports are queryable in plain language, with a weekly digest delivered to team chat automatically.
Tooling ships weekly. Payroll tooling, internal order paths, and department dashboards, rolled out opt-in with release notes. Adoption is earned, not imposed.
People interact with results, not prompts. Nobody needs to know how the system works to benefit from it. That was a design goal.

Outward, for clients

AI-visibility analytics. Measurement of how AI assistants cite and recommend brands, a service line that exists because the capability was built internally first.
Faster production turnaround. Fixes that once waited behind a backlog now ship the same week, verified live.
Reporting people can read. Narrative, rule-driven analytics reports rebuilt from raw dashboards into documents a non-technical owner actually understands.

Governance came before scale

The fastest way to lose an organization’s trust in AI is to skip the guardrails and apologize later. These were designed first:

Tiered permissions. Three access tiers between agents and production systems, with hard red lines no automation may cross.
Draft-first. Anything outbound is drafted by the system and sent by a human.
Verified-done. A deliverable counts as finished when it is checked live in production, not when a deploy turns green.
Confidentiality boundaries. Explicit rules for what internal context may never surface in shared or client-facing output.

These rules are why the revert count is zero, and why the people affected by the system trust what it produces.

The playbook transfers

The system was built inside a marketing services company, but nothing in the method is industry-specific. It is a sequence any organization can run:

Map the workflows people actually run, not the org chart. The real ones.
Encode them as skills with the organization’s own context, vocabulary, and rules.
Gate access before granting it: permissions, red lines, human sign-off on anything outbound.
Automate the loops that repeat, with verification built into every loop.
Measure in dollars and adoption, not demos. A program that cannot price itself cannot defend itself.

It took two years of building and correcting to arrive at this playbook, under the constraints of a small team and a real P&L. An organization that starts with the playbook, and a person who has already run it, moves considerably faster.

About the operator

Two decades in operations. Twelve years scaling a technology services firm from $500K to $30M in annual revenue and 5 to 200+ people, then President of a marketing services company, where the system in this report was built and is running today. The work above is recent, measured, and demonstrable.