Why a method, not a template
Most AI projects start from a tool or a template. Ours start from the work. A template captures software. It does not capture what your team actually does, who decides, which systems hold the truth, what must never happen without a person approving it, or how anyone will know the result is better. Those facts are different in every business, and they are found by mapping the work.
The Company OS Method is how we take a business function that people, scripts and scattered AI tools run today, and turn it into a digital employee: a clearly defined role, with its own instructions, knowledge, connections, schedule and approval rules, running in a workspace your company owns.
Principles
- Outcome before automation. Every engagement starts from a business outcome with a named owner and a measurable source of truth. If the outcome cannot be measured, fixing measurement comes first.
- Map the work as it really runs. We inventory people, agents, scripts, schedules, data and connections before designing anything. Contradictions between written rules and real behavior are the most useful findings.
- Every step gets one owner type. Deterministic software handles rules, routing and records. An AI agent handles interpretation, research and drafting. A person decides, approves and owns relationships.
- Risk is classified per step. Does the step only read, draft, or act on something real? Can it be undone? Will customers or the public see it? Does it involve money? The approval rule follows from the answers.
- Use the platform as shipped. We use the workspace's own features first and build custom pieces only where nothing native fits, recording each exception.
- One thin slice, end to end. One process, one channel, one approval path, measured, before adding breadth.
- Calibrate before trusting. Shadow runs and draft-only work come first. More autonomy is earned with evidence and can be withdrawn.
- What is yours stays yours. Your names, accounts, voice and thresholds live in your own configuration, and your data stays in your accounts. Reusable logic lives in the role template.
- Designed for handover. Even when we operate it with you, you can run it without us.
The nine stages
Each stage produces a written record and ends at a gate: a short list of questions the record must answer without anyone having to ask.
1. Discover
Inventory the function as it runs today: people, AI tools, scripts, automations, schedules, data stores, connections, metrics and every place where someone approves something. We look in the places people forget, such as scheduled jobs inside other tools, background services, and approvals that nothing actually executes.
You receive: a current-state inventory and a list of findings. It ends when every actor, system and schedule has an owner, and every outcome metric has a known source or is recorded as missing.
2. Map processes
Break the function into processes, each with a trigger, inputs, numbered steps, decision points, outputs, an owner, frequency, time spent and the metric it moves.
You receive: a value-chain map and a process card for each candidate process. It ends when each candidate has a complete card with real numbers or explicit unknowns.
3. Classify steps
For every step, decide who or what should do it (software, agent or person) and classify its risk. Score each process on value, feasibility and risk, then choose the first slice.
You receive: a step classification and a recommended first slice. It ends when every outward-facing, irreversible or money-related action has a human approval step.
4. Design the role
Specify the digital employee before choosing how to build it: purpose, manager, scope, KPIs, capabilities, the knowledge it needs, each connection as a narrow permission, the approval matrix, its schedule, budget and stop switch.
You receive: a role design your manager can read and sign. It ends when the manager accepts the role card and the approval matrix.
5. Fit to the workspace
Map every requirement to a feature of the company workspace, and list the gaps that need building. Our reference workspace is built on Cloudflare OS, an open-source AI workspace deployed in the customer's own cloud account.
You receive: a fit sheet showing what is native, what is configured, and what must be built. It ends when every requirement is covered or is a recorded gap with an owner.
6. Build
Configure the workspace and build only the gaps: instructions, knowledge collections, connections and permissions, small internal apps, schedules, data records and any custom connectors.
You receive: a working first slice in your environment, its configuration and a runbook. It ends when the slice runs end to end in a shadow or non-production setting with records written.
7. Calibrate
Run the slice alongside the current process, then in draft-only mode with human review. Measure accuracy, edit rate, time saved and how failures are handled, against the baseline from stage 1.
You receive: a scorecard comparing the digital employee with the current process. It ends when the agreed thresholds are met and signed off.
8. Operate
Make the role part of how the business runs: a daily approval rhythm, a weekly review, a monthly scorecard, cost tracking, versioned changes and an incident runbook. Autonomy is reviewed with evidence.
You receive: an operating rhythm and a monthly scorecard. It continues with a monthly review of outcomes, cost, incidents and autonomy.
9. Package
Turn what is generic into a reusable role template, keep what is specific to you in your own configuration, and record what we learned in the method itself.
You receive: documentation you can maintain. It ends when the role could be deployed again without redesign.
Why later engagements are faster
The first deployment of a role goes through every stage from scratch. Later deployments of the same role start from its template: process cards, a default role design and approval matrix, connection patterns and knowledge templates. Discovery, mapping, classification and design then become a matter of confirming and tailoring rather than starting from zero. Scope and timeline are still agreed for each engagement.
How it fits an engagement
The method is the inside of our forward-deployed engagements: Find covers stages 1 to 3, Build covers stages 4 to 6, Prove is stage 7, and Hand over covers stages 8 and 9.
We use it on ourselves first
We are applying this method to AIMasterz's own marketing operation before offering the same role to clients. We will publish that case when it has measured results.
Templates
Every stage record has a template you can download and use: see templates.