How a company gets its own HQ
Some companies have no system of record. The truth lives on paper in a cabinet, in one workbook on one person's PC, and in an ERP that two people open. For a real-economy business we build the missing layer: an owned workspace where people see what matters, do the next action, and leave a reliable record. The agents live inside it, behind the same login.
A look under the hood: what the HQ is, how permissions work, why the existing systems stay untouched, and where the AI fits without ever becoming the source of truth.
The short version
The HQ is one web app, which installs as an app on phones and desktops, with one login per person secured by a passkey. Each person sees only the modules they were granted: work orders, the document archive, working files, finance views, field books, the fleet, parcels, weather. Every read is scoped to the person, their department and the objects they are allowed to see; every change is written together with its audit row in one transaction. The company's existing systems, the ERP, the state e-invoicing portal, the archive, the file shares, keep their authority. The HQ reads them and never writes into them unless a separately approved path exists.
That is the whole shape. The rest of this page is how the two ways in share one permission model, what the gate on consequential actions looks like, and why we build it this way instead of bolting a chatbot onto the ERP.
Two interfaces, one permission model. A person and their agent enter through the same scoped door (green); the systems the company already runs are read, never written.
One login, and everything follows from who you are
Private mode requires a real identity. The browser only ever holds an opaque session; the identity tokens stay encrypted on the server. From that identity the HQ derives everything: which modules exist for you, which actions you may take, which department's records you see, which archive tags, vehicles and parcels are yours. Hidden objects do not return "forbidden", they return "not found".
This matters more in a mid-sized real-economy company than in a software team, because the people are not equal in what they should see. The owners' view is not the managers' view; a field worker's view is their own orders and nothing else; the bookkeeper sees invoices the agronomist never will. Module capability, tenant, department and object scope intersect on every single request, and the same intersection applies to the AI. There is no back door where the agent sees more than the person it works for.
The agent is a second interface, not a second brain
Every person gets an agent bound to their access. It talks to the same scoped APIs as the browser, so it can read what they can read and propose what they can propose, and it cannot invent business truth or bypass an approval. If a person's access changes, their agent's access changes with it, on every device, because the binding is one revision of one record.
In practice this is what makes the AI usable in a company that has never had one. The production lead dictates a field operation in plain words and it lands in the same book the web form writes to, with the same registers and the same validation. A manager asks for the state of today's orders and gets the board, not a guess. The model never sits between a person and the record; it sits beside them, with the same rights.
The universal gate. Anything that touches money, an official record, stock or a person is proposed, then approved by a named human (green), then recorded with its audit row in the same transaction.
The existing systems keep their authority
The ERP stays the ERP. The state e-invoicing portal stays the legal record of invoices. The archive keeps the originals. The HQ translates their complexity into native workflows without exposing provider screens or generic tool consoles, and it does not write into any of them until an explicit, separately accepted cutover.
Finance is the clearest case. Normal dashboard reads never open a session on the ERP at all; they read an immutable snapshot taken outside working hours, validated before it is promoted, with the previous generation kept. How that works is its own page: how finance reads the ERP without touching it. The same rule holds for e-invoices (browse and reconcile, never issue from here), payroll (deterministic calculations and review files, never a posting without a human) and documents (the archive owns archival records, the file service owns working files, the HQ stores its own evidence).
What is in it today
The modules are the ones the company actually runs on, added one working slice at a time: work orders with planning, a board, execution, review, evidence and audit; the digitized paper archive; working files mirrored from people's own cloud drives, with in-browser document editing; finance, ledger, bank, tax, settlement, inventory, asset and KPI views from the ERP snapshot; e-invoice browsing and versioned accounting preparation drafts; deterministic payroll review artifacts; field books with registers, season plans, operations, treatments and yield; fleet positions and history, parcel maps, satellite history, weather forecasts and the on-site station.
Each module is granted per person. The data models are the company's own vocabulary, in the company's own language, and every value shows its source and freshness. Unknown or ambiguous data stays visible as unknown instead of being guessed. Three of the slices have their own pages: the paper archive, the field books and worker check-ins by text message.
Owned, self-contained, provable
The HQ runs on the company's own server with its own database, identity, storage, backups and deployment. Git is the code truth; every image carries the exact commit; a deployment pulls a clean approved commit, migrates forward, checks health on a private port and proves that a backup restores, before any public route is switched. Public exposure, real-user activation and provider cutover are separate gates with their own acceptance evidence, never assumptions.
This is also why ordinary work stays fast and usable without an AI response. The deterministic core is the source of truth for the HQ's own work; the model reads scoped context and prepares proposals. If the model is down, the company still runs.
Trust is the currency. Every gate, label and "answer before you escalate" exists so management trusts the system enough to act on it.