AI does the work. We do the relationships.

We focus our time and energy on solving real health problems and building long-lasting, trusting relationships with people. The rest is work for agents. And this is how we work with them.

Our operating system starts with three value drivers.

KMG Global opens regulated markets for evidence-based health brands. Every piece of work has to serve at least one of the three.

Every run pays into a value driver
  1. Value drivers

    What creates value

    • GrowthMarkets × partners × revenue per partner
    • EfficiencyCost to serve, speed, return on digital spend
    • Trust & ResilienceCompliance, trust, continuity
  2. Functions

    What we do for a partner

    • WinA qualified partner
    • GrowRepeat demand
    • DeliverProduct availability
    • AssureDefensible releases
  3. Workstreams

    Recurring work with one owner. Two of the sixteen open here.

    • Partner QBR
    • Sell-in / sell-out
    • Performance marketing
    • Pricing and access
    • Demand forecast
    • Production order
    • Logistics and customs
    • Order-to-cash
    • Registrations, labelling
    • Claims and releases
    • Complaints / vigilance
    • Batch release
  4. Runs

    One logged execution

    NDA and LOIProcess BD-05Human in the loop

    One NDA for one partner, with its own log

    Trigger
    A deal reaches the stage "NDA needed".
    Nikita
    Fills the unchanged template, files the PDF and books the meeting.
    A person
    Approves every change and signs every agreement.
    The partner gets
    A prepared NDA, ready for a human signature.

    See this run step by step

    Meeting prepProcess BD-07Fully autonomous

    One brief for one meeting, with its own log

    Trigger
    A meeting is on the calendar.
    Nikita
    Writes one page: history, open points, market facts and suggested asks.
    A person
    Reads the brief. Nobody has to approve it, because it stays in the house.
    The partner gets
    A counterpart who remembers everything.

The work decides how far an agent goes.

There are four modes, from an agent working alone to people only. The risk of the work sets the mode.

  1. Fully autonomousThe agent runs the job from start to finish. A person audits it from the log.ExampleA one-page brief before every meeting
  2. Human in the loopThe agent prepares everything. A named person decides, approves or signs.ExampleAn NDA, signed by a person
  3. Sparring partnerA person does the work. The agent helps on request and challenges the thinking.ExamplePreparing a negotiation
  4. People onlyNo agent takes part.ExampleConversations, negotiations, signatures

Ink: the agent's share of the work. Orange: the person's share.

Every run has the same three parts.

  1. TriggerAn event or a state rule starts the run. First it is checked: is the trigger valid?
  2. WorkReading, doing and checking, in as many rounds as it takes. Who does what is fixed before it starts.
  3. OutputThe result lands in a master and in the log. Often it is the trigger of the next run.
R
does the work
V
verifies
G
holds the gate
A
answers for it, always a person
A real example: one NDA, in six steps

RunOne NDA for one partnerProcess BD-05AThe market owner, with Legal

  1. TriggerTriggerA deal reaches the stage "NDA needed"
  2. WorkNikita readsRMasters, templates and SOPs
  3. Nikita doesRThe filled template, with its evidence IDs
  4. Milo checksVPass, flag or block
  5. A person decidesGApproves every change and signs every agreement
  6. OutputNikita writesRFiles the signed PDF and books the meeting. Only Nikita can send.

Half and half: the trigger. It comes from a person or from the system. Set in type: an agent does it. Drawn in marker: a person does it. Double line: the output.

People, agents and tools share one body of knowledge.

The core team sets the goals. Agents and tools do the work between them. What all three know is written down in one place.

Core teamSets the goals, decides, signs and answers

Google WorkspaceWhere the team works day to day

AI

NikitaMilo

Two agents, each on its own machine, on our own hardware

Models

Several, behind one router. Any of them can be swapped.

Backup

On our own storage

Knowledge

What we know, written down

  • Masters
  • Templates
  • SOPs
  • Each agent's identity

Tools and SOPs

Notion

The single source of truth. Every master has one owner and one way to write to it.

SOPs

How each process runs, step by step

Run log

Every step of every run, with its gate and verdict

Models improve every few months, so every part is built to be swapped: model, tools and data connections are separate.

Nikita and Milo are our two agents.

Nikita does the work and Milo checks it. Each has a written identity: who it is, what it is for and what it must never do.

Does the work

Nikita

  • Always on, on its own machine
  • Claude and Gemini do the work. A small local model routes the tasks.
  • Reads the masters and writes the results. Only Nikita can send.
  • Has its own login, a named manager and one switch that cuts all access

Checks the work Still in build

Milo

  • Runs on a different model provider
  • Reads only and has no way to send
  • Can block. Can never approve.
  • A pass from Milo never replaces a person's decision

Both work inside fixed limits.

What our agents do

  • Read the masters, templates and SOPs
  • Prepare the work and attach its evidence
  • Log every step of every run
  • Work under their own login, with the least access the job needs
  • Lose all access when one switch is flipped

What they never do

  • Sign a contract or a release
  • Pay, or start a server
  • Approve a claim
  • Take instructions from an incoming email
  • Change their own goals or the way they are checked
Alain Cappelle

Our way of working only works because agents do the work, and because we keep a firm grip on them.

A conversation takes as long as it takes. So does a negotiation. Nothing else should.

The more agents do, the more it matters which rules they work by. These are ours. We change them when we learn, and we log it when we do.

Alain CappelleManaging Director, KMG Global

House rules for people and agents.

Ten rules. Five are written out with an example and a way to check them. Five are still open.

Rule 1: AI is never responsible.

Every run has one person who answers for it.

The market owner answers. Where no owner is named, a managing director does.

Answering covers the process and its runs. It does not mean performing every task. Goals stay ours, and people own the relationships.

CheckEvery process names its owner.

Example: four workstreams and who answers

  • WinNDA and LOIMarket owner, with Legal
  • GrowSell-in / sell-outCommercial lead
  • DeliverDemand forecastSupply lead
  • AssureClaims and releasesRegulatory lead or signatory

Rule 2: Humans sign off high-risk results.

Money, signatures and submissions stay with people.

Work that is legal, financial, regulatory or irreversible stops at a person every time. An agent may prepare it. No number of clean runs changes that.

CheckEvery process carries its risk tier.

Example: one NDA

NDAProcess BD-05

  • TemplateUnchanged, filled in by Nikita
  • CheckMilo, once built
  • FilingNikita files the PDF and books the meeting
  • ChangesA person approves every change
  • SignatureA person signs every agreement
  • Money
  • Signatures
  • Submissions

Rule 3: Independence through modularity.

We go the extra mile and build our own system: a harness on our own hardware, in which every module can be swapped.

If you build your entire system on one company, such as Google or Anthropic, you depend fully on every change it makes to content moderation, quality and pricing.

For infrastructure this critical, we cannot take the risk of depending on one supplier.

CheckAny single module can be swapped.

Example: the models in our harness

HarnessOurs, on our own hardware

  • Model for the workClaude and Gemini
  • Model for routingA small local model
  • Model for checkingA different provider, still in build

A dashed outline means the part can be swapped. Tools and data connections are separate modules in the same way.

Rule 4: Knowledge lives in the system.

What we know goes into processes and tools. It does not stay in heads.

An agent can only work with what is written down. The same goes for a new colleague.

So every process has its SOP, every fact has its master and every run leaves its log.

CheckEvery process has a written SOP.

Example: where knowledge goes

What we knowWhere it lives

  • How we do itThe SOP of the process
  • What is trueThe master, with one owner
  • What goes outThe template, versioned
  • What happenedThe run log, never edited

Rule 5: One owner. One truth.

One person is accountable. One place holds the truth.

Every fact has one master, with one owner and one way to write to it. Everyone else reads.

A copy is not a second truth. Where two places disagree, the master is right.

CheckEvery master names its owner.

Example: one master

ContractsMaster

  • HoldsEvery agreement, its status and the signed PDF
  • WritesOne way in, through the process that owns it
  • ReadsEveryone else
  • OwnerLegal, or a managing director

Changelog

We change these rules when we learn. Every change is listed here.

18 Nov 2026
Version 0.2
Rules 4 and 5 written out. Rules 6 to 10 open.
16 Nov 2026
Version 0.1
Rules 1 to 3 written out. Rules 4 to 10 open.