← Gimbo's Universe/ AI Builder/ Org Evolution
Supply Chain · How the Org Evolves

One company,
two futures.

Hand a distribution company to AI and there are two distinct steps — don't conflate them. Mode A swaps each department's people for agents: automated, yes, but still the old company with a new engine — the three flows still run as a relay. Mode B is AI-native: no departments; one core optimiser treats the three flows as a single problem and solves them jointly, with agents existing only at the edges, on the interfaces. Two animated flow diagrams — watch how information moves, and where the humans stand.

Information flow Goods flow Cash flow Where humans sit
Mode A · Agent-Operated Company

Old company, new engine: the relay pipeline

The org chart stays untouched — each department's people are simply replaced by agents. The three flows still pass the baton segment by segment: one step finishes, then hands off to the next. Coordination relies on an orchestrator plus human approval gates.

Human · General Manager approves POs & payments · takes escalations Orchestrator task routing · still split by department Orders Agent order entry · acks Planning Agent MRP · raises POs Logistics Agent in-transit · customs Warehouse Agent GRN · FEFO Finance Agent 3-way match · pay INFO · order → MRP → PO → ASN → GRN → invoice (relay) PO value gate (human) GOODS · supplier → warehouse → customer (waits on info) Supplier Warehouse human · moves goods Customer CASH · customer pays on terms ← supplier paid after dual sign-off dual-sign payment gate (human)
Same object of management: still managing processes — agents are faster employees slotted into the same departmental boxes
Flows run as a relay: goods wait on information, cash waits on goods — delay compounds stage by stage
Humans are the gates: PO value gate, dual-sign payments, exception escalation — people sit mid-stream in every flow
Mode B · AI-Native Company

A new species: one core, agents at the edge

No departments. At the centre sits one coupled optimiser (a world model + an objective function); edge agents are arranged by interface, not by function. The three flows no longer relay — they are solved jointly against a single objective: cash constrains goods, goods constrain promises, all resolved in one pass.

Human · Owner (central banker) sets objective weights · risk appetite · constraints rare, high-leverage demand loop supply loop capital loop Core Optimiser world model + objective fn three flows solved jointly Demand Interface orders · promises · pricing Supply Interface suppliers · POs · confirms Capital Interface terms · FX · cash cycle Logistics Interface forwarders · customs · ETA Human · Physical Layer moving · receiving · installs · in-person relationships agents scale with interfaces — not with functions
Management moves up a level: managing the objective function — the Owner doesn't approve documents, they tune weights, the way a central bank sets rates
Flows solved jointly: one incoming order resolves its promise date, its procurement and its cash impact in a single pass — no relay latency
Humans at the two ends only: defining what "good" means at the top, touching the physical world at the bottom — no human-shaped gates in between
Side by Side

Where the evolution actually happens.

Mode AAgent-operated company
Mode BAI-native company
Object of management
Processes — agents slotted into departmental boxes
The objective function — departments dissolve
Shape of the three flows
Relay pipeline: information → goods → cash; delay compounds stage by stage
A force field: solved jointly against one objective, coupled in real time
Coordination
Orchestrator scheduling + human gates + the residue of meetings
Explicitly engineered into the core optimiser — no longer free, but no longer meetings either
Where humans sit
Mid-stream: approving POs, dual-signing payments, taking escalations
At the two ends only: defining "good" (weights, constraints) at the top; the physical world at the bottom
How agent count grows
With functions — one per department, the org chart copied over
With interfaces — a new agent only when there's a new external interface
Inventory & cash
Three-policy monthly MRP — the concession to human bandwidth persists
Continuous per-SKU tuning; inventory driven toward zero, the cash conversion cycle can turn negative — growth flips from consuming cash to producing it
Distance from today
Buildable today — my Action Center build and the shadow-mode work pilot sit in this box
What's missing is engineered coordination and earned trust — this is the end state, not the starting point
Part 3 · How Far?

Two very different distances.

So how far is a real company from each mode? The honest answer splits in two: the distance to Mode A is measured in engineering years; the distance to Mode B is measured in company generations. The most honest measuring stick I own is the gap between my own two artifacts — the v1 simulator runs ~85% touchless in its clean synthetic world, and the first real supplier email broke four of its assumptions at once (the honesty layer in Metis 2.0). Every demo lives at 85%; every production system stumbles on its first real email. The industry's true position is that gap — not the demo number.

To Mode A
3–7 years

An engineering problem — already underway

For an ordinary mid-size company, getting the majority of routine back-office touches (order entry, chasing, reconciliation, AP) handled by agents is buildable — seat-agent products are shipping today. What actually sets the pace:

Ontologies are explored, not installed. Every seat's real rules — each supplier's date formats, reference quirks, exception patterns — only surface by running the work and cataloguing where it breaks. There is no shortcut around this fieldwork.
Trust expands on the organisation's clock, not the model's. Every action type moved from "human approves" to "auto" needs an audit trail behind it. That cadence is set by the company — and it is the real rate limiter.
To Mode B
Three walls

None of them technical

The objective function is politics. Service vs working capital vs margin is "computed" today by departments negotiating. Writing it down as one function means writing the power structure down in plain text — nobody signs that.
Segregation of duties is law — and one core violates it by design. Three-way match exists half to catch errors, half to stop fraud: control frameworks require ordering, receiving and paying to be separate roles. Mode B has to wait for the control frameworks themselves to be rewritten for machine actors.
The ontology is only discoverable by running Mode A for years. Mode A isn't the budget version of Mode B — it is Mode B's only construction path.

So incumbents don't reach the single core. The realistic ceiling is one shared world state plus gated cross-department writes — a strong data core with delegated authority. Worth climbing to (most of the relay's losses die there), but it's a plateau, not Mode B.

How Mode B Arrives
By replacement

Species don't migrate — they get replaced

Mode B natives are already being born — tiny teams running one shared state and a fleet of agents from day one. They don't dissolve departments; they never had any.
The order of arrival is set by edge thickness. Businesses whose counterparties are APIs (software, media, services) go first — their edges are born ontologised. Physical goods, multi-currency, sea freight and regulated QA go last: expect 10–15 years before natives contest those niches.
Thick edges are the incumbent's buffer. The strategic question for an existing company is not "how do we become Mode B" — it's "can we climb to the plateau, and turn what it saves into a moat, before a native reaches our niche."

These aren't two options — they're one evolutionary path.

Mode A is the necessary step: trust is earned one action type at a time, and the human-shaped gates are both its signature and its value — the supply chain Action Center I'm building, and the shadow-mode pilot I'm starting at work, sit in this box. Mode B is the extrapolated end state: once coordination is engineered and every SKU can be tuned continuously, the department as a shape loses its reason to exist. The distance from A to B isn't model capability — it's two things: engineering the coordination layer, and earning organisational trust.

Full reasoning: Running a Company on AI Agents · Two Simulations, One Live Wire (retro)