How it works

Fixed fee. 3x deliverables. 1x gate.

Phase 1 is scoping. The question is: what does Decision OS look like inside your specific business, and is it worth building? You walk out with three artifacts, a quantified automation map, and a clean go/no-go on Phase 2.

No vendor lock-in. No consultant dependency. Three separate SOWs across the engagement — you decide at each gate whether to proceed.

THE THREE DELIVERABLES

1. Business Summary (.md → Notion file) — the training corpus

The story of how the company actually wins. Procurement levers. Premium positioning. Customer rhythms. Unstated foundational assumptions the founder has never written down because they live in their head.

This is founder-altitude work. Not delegable to a Head of AI in their first quarter. It needs the standing to ask questions that produce real ground truth — and the cross-functional context to know which questions matter.

What's inside:

• Story of the product from procurement to delivery
• Company values, guardrails, lessons learned
• Five-question framework: where's my money, where's my product, who's my customer, what have we done, where are we going

2. Roles Org Chart (FigJam) — what every person actually does

Every role. Every responsibility. Every software touched.

Automatable roles get circled, annotated with $ saved per year. Security access tiering per individual: High (brain query), Medium (departmental agentic usage), Low (departmental chat).

This is where the automation savings get quantified — and where the case for Phase 2 either earns itself or doesn't.

3. Ideal Botmap (FigJam) — how information will flow

Visual schematic of each department's LLM stack. Data inputs, trigger cadence, processing layer, outputs, human checkpoints.

Shows how departmental LLMs feed the central brain. Where the claws sit. Where the kill switches sit. What gets handed to the Head of AI to build.

METHOD

Example interview waterfall — top-down through the hierarchy.

• Entry: Finance first. Software footprint. Where the data already lives.
• Then: Operations. The live queries actually being asked.
• Then: Marketing, Customer, Product. Department by department.
• Finally: Founder / CEO. Synthesis. The unstated assumptions.

Each interview pulls roles, workflows, software, and data flows. By week six, the three deliverables are written, the automation map is quantified, and the gate review is scheduled.

The gate

End of week six, three questions get answered:

1. Are the deliverables good?
2. Do the automation savings justify Phase 2 investment?
3. What's the token cost and operational overhead at steady state?

If yes — Phase 2 SOW gets signed.

If no — you walk away with three artifacts you fully own and a clearer picture of your business than you had six weeks ago. Either way, it was worth doing.

Best time to plant the tree was two years ago. Second-best is now.

Built on what Phase 1 scoped — no rediscovery, no false starts.

Three milestones. Each independently approved. Walk away clean at any of them.

2a — Proof of Concept

One departmental LLM. Live. Usually Finance or Supply Chain.

The department with the cleanest data footprint goes first. Finance usually wins — books are already structured, software is already in place. Supply chain is the alternative if procurement complexity is the bigger leverage point.

The POC ships with:

• Department-specific dashboard fed by current, clean data
• Scoped LLM with isolated credentials
• Logged query audit trail
• Kill switch at the LLM layer

You don't pay for 2b until the POC works. That's the gate.

2b — Department Rollout

Dept-by-dept LLM build. Claws, dashboards, integrations.

Each department gets its own LLM with scoped data access. No cross-bleed: Finance cannot read HR. Supply chain cannot read marketing customer data. Each LLM has its own credentials. Every query is logged.

Built throughout this phase:

• Agility scoring — how easily models can be swapped (Anthropic ↔ OpenAI ↔ local)
• Grounding checks — outputs validated against ground truth
• Containment audits — data segmentation verified
• Choke points + kill switches — at every LLM and at the brain layer

2c — Brain Online

Full combination of all dashboards into the cognitive layer.

Departmental dashboards roll up into the CEO dashboard. The CEO dashboard rolls up into the brain.

The brain holds every dataset simultaneously. Surfaces cross-functional correlations no individual in the room could see. Whispers via phone when the room is missing something the data can see.

This is the moment the install pays off.

Most AI rollouts die at the human layer. The brain is only as smart as what flows into it — and what flows into it is decided by a hundred small choices people make every day. Phase 3 is where the system either compounds or rots.

What gets installed

Workflow standardisation

The data that feeds the brain stops being entered retroactively, in different formats, by different people, in different places. It gets logged at the point of event, in a structure the machine can read.

• Project managers log decisions at the same time, in the same format, every cycle
• Meeting notes structured for machine readability — decisions, owners, dates, outcomes
• Customer interactions tagged consistently across channels
• Procurement and supply data entered at point-of-event, not from memory two weeks later

Data hygiene rituals

What doesn't get reviewed degrades. Hygiene needs a weekly cadence to survive.

• Weekly data quality reviews per department
• Automated checks flag missing or malformed entries before they pollute the brain
• "Garbage in, garbage out" made visceral — show teams what bad data does to their own dashboards

Department-lead fluency 1:1s

Each lead trained to query, iterate, and improve their own LLM. Critically — they learn when the LLM is wrong and how to correct it. Blind trust in the output is the failure mode.

Best-practices video library

Onboarding asset for every new hire. "How we use the brain" becomes part of the culture, not just the tooling. Includes hallucination detection training and the case against blind abidance.

How I work with your Head of AI

I don't replace them. I work hip-to-hip.

• Their job: wiring the pipes. Permissions. Connections. Keeping data current. Building infrastructure. Owning the backend.
• My job: founder-altitude work above that. The intent. The decision-making process. The questions worth asking.

They execute against the intent. I make sure the intent is right. By the end of Phase 3, they own the system entirely.

The point isn't for me to know your backend better than anyone. The point is for someone internal to know it better than anyone — so they can fix it, improve it, and run it long after I'm gone.

Success criterion

Your Head of AI ships one new departmental LLM end-to-end without my involvement.

That's the test. Not a milestone deck. Not a satisfaction score. A live system, built independently, working in production.

When that ships, my work is done.

The compounding asset stays internal.