Skip to content
Talk to us →︎
Our AI operating system

How we run on AI.

Written rules at the top. AI doing the work underneath. Every lesson fed back in.

What it isLast updated 20 November 2026

An AI operating system,
written down.

An AI operating system is the set of written rules, procedures and wiring that tells people and artificial intelligence (AI) how an organisation works. Every AI tool starts from the same place, and a person stays accountable for the result. This is the one AI Accelerator runs on. We are an AI capability partner, and we build the same thing with client teams.

It has six layers, from the rules that change least to the work that changes daily. When two rules disagree, the higher layer wins. Below, each layer is set out with why it is there, followed by how any AI plugs in, how security runs through it and the two loops that keep it learning.

The layersTop to bottom

Each layer answers
to the one above.

Six layers, in the order our management system sets them. When two rules disagree, the higher layer wins. Higher layers change rarely and set the limits; lower layers change daily and do the work. A task picks up its rules at the top and is broken into parts as it goes down. Security rules and checks sit on every layer, shown in blue.

01

The Constitution and Objectives

Eight articles and twelve operating principles that bind people and AI alike, with our objectives beside them pointing the way. A person is always accountable, and when two rules collide, the higher layer wins.

8 articles · 12 principles

Why it is there. It settles conflicts before they happen. Everything below answers to it, so people and AI never have to guess which rule wins.

8 articles12 principlesObjectivesA person is accountableHigher layer winsNo silent changes

Security is not negotiableNo secrets in chat or repoA leaked secret is dead

02

Position

What we are, who we serve, how we reach them and how we sound. Every standard below is read against it, so a proposal, a page and an AI draft all start from the same idea of who we are.

4 documents

Why it is there. Without it, every proposal, page and AI draft decides for itself who we are. With it, they all start from the same idea.

What we areWho we serveHow we go to marketHow we sound

03

Standards and records

Written standards say how we do each thing, from proposals to security. One index names a single home for every kind of information, so nothing lives in two places and nobody has to guess.

45 standards · 17 systems of record

Why it is there. Variance is expensive where it matters, so the approach is written once. The index stops information living in two places.

Brand and voiceDocument controlProposal standardResponsible AI policyClient foldersPublishing and claimsDecision logRisk register

Information security policyAccess and identityIncident responseSupplier checksContinuity and backupData protection, UK GDPR

04

SOPs

Standard operating procedures: the whole repeatable job, step by step, owned by a person, with the human judgement kept in. Skills take on the steps an SOP hands to AI; the SOP stays what a person follows.

10 procedures

Why it is there. The judgement in a job has to sit somewhere. An SOP keeps it with a person and says exactly which steps AI can take on.

Client onboardingWorkshop deliveryAI Primer engagementStaff onboardingSocial content productionDemo accounts for AI toolsMedia libraryShowing our intranet to clients

Phishing and suspicious messages

05

Skills

A shared library of skills that carry out the steps an SOP hands to AI, from proposals to decks to research. The same library serves Anthropic Claude Code, Anthropic Claude and OpenAI Codex, and each skill passes team review gates before everyone relies on it.

104 skills · 7 plugins · 3 gates

Why it is there. A skill turns a step into something any AI does the same way every time, and the review gates mean everyone can rely on it.

ProposalMeeting prepPost-call capturePresentationBrand graphicDiscoveryClose sessionWhere does this live

Automated checks on every skillTeam review before sharingUnused skills retired

06

Tools and systems

The systems that hold the work: Linear for tasks, Attio for relationships, Drive for files, GitHub for the rules themselves, Slack and email for conversation. Swap a tool and the layers above stay the same.

One home for each thing

Why it is there. Tools change every few months. Keeping them at the bottom means swapping one never rewrites the rules above it.

Linear: workAttio: relationshipsDrive: filesGitHub: rules and codeSlack: team chatGmail: clientsFigma: design

Password manager vaultTwo-factor everywhereEndpoint protectionCyber Essentials Plus

Base

The physical world

The ground the system stands on: standing meetings, workshops at the AI Adoption Lab in Greenwich, client calls. Each one is captured, so actions land in Linear and decisions land in the log, not in someone’s notebook.

6 standing meetings · the Lab

Why it is there. Most decisions still happen in a room. Capturing them is what stops them living in someone's notebook.

Weekly sales meetingSkills meetingPlatform sprintAll handsWorkshops at the LabClient callsLive from the Lab

Locked screens, locked kitAnyone can report a risk

Around the six layers: one way in for every AI, and the agents that do the work.

In

The way in: CLAUDE.md and AGENTS.md

Not a layer but the wiring. A CLAUDE.md and an AGENTS.md side by side, at every level from the company down to a single project folder, point any AI at the six layers. Each one points back to the Constitution and never restates it, so Anthropic Claude, OpenAI Codex or the next AI reads the same rules.

CLAUDE.md + AGENTS.md · 4 levels

Why it is there. One door for every AI means one set of rules to keep current, however many AI tools you use.

CLAUDE.md + AGENTS.md, personalCLAUDE.md + AGENTS.md, workspaceCLAUDE.md + AGENTS.md, repositoryCLAUDE.md + AGENTS.md, folderUK English, no em dashesCheck Linear firstBranch, then pull request

No client data in gitNever paste a credential

Do

Agents

The workers. Agents pick up a task, read the rules through the way in, use the skills and work in the tools, often splitting the job into parts that run in parallel. Scheduled runs work overnight. A named person signs off the result.

A person accountable for every agent

Why it is there. Agents take the legwork. A named person signing off keeps the accountability where it belongs.

Research subagentsOvernight briefNightly lesson harvestWeekly lesson promotionLibrary rebuild botModel routing by task

Only cleared AI tools touch client dataLimited permissions by defaultAutomated checks before anything is savedA person signs off

The loopsNothing learnt is lost

The system gets better
every week.

Lesson loop

Finished work teaches the system

When a piece of work closes, its lessons are captured. Each week they are promoted into a skill, an SOP, a standard or an instruction file, so the next piece of work starts from a better place.

Meeting loop

Every meeting turns into work

A meeting becomes a transcript, the transcript becomes actions and decisions, the actions become work, and the next meeting reviews it. Nothing depends on memory.

Plug in another AI

A new AI knows how we work from day one

We run on Anthropic Claude Code, Anthropic Claude and OpenAI Codex today. Microsoft Copilot, Google Gemini, OpenAI ChatGPT, xAI Grok, Meta Muse or whatever comes next can join the same way: it is cleared for our data first, then reads one short AGENTS.md file and gets the same Constitution, skills, checks and systems as the rest. No rewrite, no second set of rules.

Where does it live?Nine examples

Follow one thing
through the system.

Pick a question under the drawing and follow it through the system, step by step. All nine are written out here.

Show all nine, step by step

Who signs off AI work?

  1. A task is picked up from Linear, with a named owner
  2. The Constitution sets the limits: a person stays accountable
  3. Position says who it is for and how it should sound
  4. The standards say what good looks like
  5. The SOP sets the steps, and where a person decides
  6. A skill carries out the steps handed to AI
  7. Agents break it into parts and work them in parallel
  8. The results are filed in the right system
  9. A person checks it and signs it off

How does every AI write in our voice?

  1. The way we work is written down
  2. How we sound is part of our Position
  3. Brand and voice standard: the writing rules
  4. Every CLAUDE.md and AGENTS.md points to it
  5. Proposal, deck and graphic skills follow it
  6. Automated checks catch house-style slips before work is saved
  7. Rules in GitHub; brand files in Figma, Drive

How do the rules change safely?

  1. No silent changes to how we work
  2. Owner, version, status and review date
  3. Agents check its status before relying on it
  4. Agents propose; a person approves
  5. Every change is a reviewed pull request

How is a proposal put together?

  1. It starts with a discovery call
  2. The proposal skill drafts from the brief
  3. Position: who we serve and how we sound
  4. The Proposal standard sets the bar
  5. Filed in the client’s Drive folder
  6. The deal and next step tracked in Attio

What happens to a meeting decision?

  1. Agreed in a meeting
  2. The transcript is filed, Drive or GitHub
  3. The decision goes in the decision log
  4. Actions are captured
  5. Each action becomes a Linear issue

What happens to a password?

  1. Secrets never touch a chat or a repo
  2. The Secrets management standard says how
  3. An agent handed one stops and says so
  4. It lives in a password manager vault
  5. If it leaks, it is dead: rotate it

Where does a client’s personal data go?

  1. A client shares personal data in a workshop
  2. It is kept only in our approved company systems
  3. It goes into an AI only with a data agreement and training switched off
  4. A consumer AI tool? Never. It stops at the edge of the system
  5. We are registered with the ICO, and UK GDPR sets what we keep
  6. A slip is reported at once: the 72-hour clock starts

How does a new AI learn our rules?

  1. Cleared at the door: a data agreement, and no training on our data
  2. It reads one short AGENTS.md file
  3. AGENTS.md points it to the Constitution
  4. It learns who we are and how we sound
  5. It finds the standards and where everything lives
  6. It follows the same procedures
  7. It loads the same shared skills
  8. It works in our tools, and a person signs off

How does a new skill get made?

  1. Someone spots a job worth repeating
  2. It is a step an SOP can hand to AI
  3. Its author builds and tries it first
  4. Skills lifecycle: team review gate
  5. Shared through the skill library
  6. Agents run it in Claude Code, Claude and Codex
ISO 9001Quality, layer by layer

Mapped to ISO 9001,
layer by layer.

ISO 9001 is the international standard for quality management. Each layer covers part of it, and the two loops follow its plan, do, check, act cycle. This is a map of what each layer covers, not a certification.

Show the clause map
The Constitution and Objectives5 Leadership · 6.2 Quality objectives
Position4.1 Context of the organisation · 4.2 Interested parties
Standards and records7.5 Documented information · 6.1 Risks and opportunities
SOPs8.1 Operational planning and control · 8.5 Service provision
Skills8.3 Design and development · 8.5.6 Control of changes
Tools and systems7.1 Resources · 8.5.3 Property belonging to customers
CLAUDE.md + AGENTS.md7.3 Awareness
Agents8.4 Externally provided processes
The physical world8.2 Customer requirements · 9.3 Management review
Meeting loop: check9.1 Monitoring and evaluation · 9.3 Management review
Lesson loop: act10 Improvement
QuestionsIn short

Questions about the
operating system.

What is an AI operating system?

The written rules, procedures and wiring that tell people and AI how an organisation works. It sits in layers, from a constitution at the top to the tools at the bottom, and the higher layer wins when two rules disagree.

How does a new AI tool learn how you work?

It is cleared first, with a data processing agreement in place and no training on our data. Then it reads the same instruction files as every other AI, a CLAUDE.md and an AGENTS.md at each level, which point it to the Constitution, the standards, the procedures and the skills.

Who is accountable for work AI does?

A named person, always. Agents can split a task into parts and work on them in parallel, but a person checks the result and signs it off.

How is client data kept out of the wrong AI tools?

Personal and client data stays in approved company systems. It only goes into an AI with a data agreement in place and training switched off, never into a consumer AI tool, and any slip is reported straight away.

How does the system stay current?

Through two loops. Lessons from finished work are promoted each week into a skill, a procedure, a standard or an instruction file, and every meeting becomes actions and decisions that land in the right system.

Can you build one for my organisation?

Yes. As an AI capability partner, we build the same shape with your team: your rules written once, one way in for whichever AI you use, procedures that say where a person decides, and loops that keep it current.

Want your organisation
to run like this?

As your capability partner, we build the same system with your team: the rules, the skills, the agents and the habits that keep them working. Start with a conversation, or see the results first.