← Back to the film

Inside the Operating System

What the film leaves out — for the people who’ll ask.

The film shows a map of the business, a card that fills with live numbers, one item in coral, two ways to fix it and your yes. This page is the same story with the plain words taken off: what the map is, how it gets drafted from the exports you already have, why the AI can be relied on, who is allowed to do what, and what stays yours when you leave. It is written for the CIO, the head of data and the person who has to sign the security review.

What you get

Four products, one map.

Web dashboard

One page per role — CEO, region head, store manager. What's on track, what's drifting, and the one item that needs a call.

Mobile app

The shop floor's app. Today's instruction, a tap to confirm, and a plain-language question answered on the spot.

AI assistant & alerts

Answers drawn from your own numbers, in plain language. Alerts routed to the right role, with the why attached.

Work orders & confirmation

Your yes becomes a written instruction. Store, region and head office each confirm back, and blockers are routed.

Where it runs

Three ways to run it. All inside your boundary.

Your own servers

AIM runs on infrastructure you already operate.

Your own cloud tenant

AIM deploys inside your AWS, Azure or GCP account.

Hosted by AIM, single-tenant

AIM hosts it for you — still single-tenant, still yours.

Reads first; writes only with your yes. Your data never leaves your boundary. Open-format export — the map is yours.

The map of your business

The map of your business is an ontology.

An ontology is a live model of the business as a set of things, the facts on them, the relationships between them, and the operations you are allowed to perform on them. It is the same structure Palantir calls an ontology. Microsoft Fabric IQ, the SAP Knowledge Graph, and the semantic views in Snowflake and Databricks are the same idea under other names. The concept is no longer unusual; what differs is where it starts and how far it reaches.

Ours is retail-native: it ships with a reference ontology of the objects every retailer already has, it is learned from your exports rather than a data-platform project, it is confirmed by your people in a workshop, and it reaches the store floor — the store manager’s page runs on the same model as the CEO’s.

Term in the filmTechnical termWhat it is
The things on the mapObject typesThe nouns of the business — Store, Product, Supplier, Warehouse, Promotion, Roster, Order, Customer. Each has an identity and a lifecycle.
What's written on the cardPropertiesThe live facts attached to each thing — sales today, days of cover, roster gap, promo uplift — refreshed from the source systems.
The threads between themLinks (relationships)Store stocks Product. Warehouse serves Store. Named once in the model, never buried in a join again.
The buttonsActionsGoverned operations attached to a thing — transfer stock, reorder, markdown, re-roster, escalate — that write back to the system that owns the record.
The numbers behind the two optionsModels (functions)The logic that costs an option on live numbers — a replenishment engine, a forecast, a margin calculation — bound to the objects it reads.

Drafted, then confirmed

AIM drafts the map. Your team confirms it.

Bring the exports you already have — a POS extract, a stock-on-hand dump from the ERP, the planning spreadsheet, a replenishment report. AIM reads their columns and, using a language model for schema matching, proposes how each one maps onto a reference ontology for your industry: store_id is a Store, sku_code is a Product, vendor_no is a Supplier, and the foreign keys between them become links.

Then one workshop. Your people walk the draft, rename what you call differently, delete what you don’t track, add the link the file never held. They sign it off and it goes live. The confirmed map is the contract every later step is checked against.

Why not fully automatic? Because fully automatic ontology learning is still a research problem, not a product — no vendor sells it, and the semantic-view generators that exist scaffold for a human to finish. Manual mapping, at the other extreme, runs 40 to 80 hours per dataset. Mapping onto a known industry target with an LLM doing the first pass, and people doing the last, is the honest middle: fast because the target is known, right because someone who knows the business said so.

Why the AI is reliable

The AI drafts. Something deterministic checks it.

A language model on its own is a fluent guesser. AIM never lets a guess touch your operation directly. Every answer, mapping and action goes through the same four steps.

  1. 1

    The AI drafts

    A language model proposes the answer, the mapping or the action — fast, fluent, and on its own not to be trusted with your stock.

  2. 2

    Deterministic rules and the map check it

    Every draft is validated against the ontology and a rule set with no randomness in it: does the object exist, is the link legal, does the number reconcile, is this action allowed for this role.

  3. 3

    If it can't be sure, it asks the person

    When the checks can't decide — "Store 112 is also on promo Friday, still transfer from there?" — AIM stops and asks. It does not guess.

  4. 4

    The answer is stored as a rule

    "Never transfer out of a store in the week before its own promo." AIM remembers it as an explicit, readable rule you can inspect and revoke. Nothing is silently retrained.

Why the map matters to the AI — answer accuracy on enterprise questions

16%

GPT-4 straight onto enterprise SQL

54%

The same questions over a knowledge graph

72%

Ontology-checked, with query repair

data.world · Sequeda, Allemang, Jacob 2023–24. These measure answer accuracy, not action reliability — which is why every role assistant in AIM carries its own evaluation set and is tested against the map before it is trusted with a button.

Three levels of trust

Suggest · do with your yes · do and tell.

Every action on the map carries a trust tier, set per action and per role, and shown on the button itself. You choose where each one sits and you can move it either way.

Suggest

AIM shows the option and the cost. A person does the work in the system they already use.

Do with your yes

AIM prepares the action and shows its impact — what moves, from where, what it costs. Nothing executes until someone with the right role confirms.

Do and tell

For low-risk, well-tested actions the role owner has chosen to delegate, AIM executes and reports back. Every one is logged and can be pulled back to the tier above.

Two things sit under all three. An impact preview: before any confirmation, AIM shows what will change — which records, in which system, at what cost — so a yes is a yes to something specific. And an audit trail: who or what proposed the action, which rules checked it, who confirmed it, what the system replied. The trail feeds back into the card, so the loop that started with the coral item visibly closes.

This is not decoration. Gartner forecast in June 2025 that more than 40% of agentic-AI projects will be cancelled by the end of 2027, for cost, unclear value and inadequate risk controls. The governance is not a layer added to the operating system; it is the operating system.

One layer over what you own

Your systems stay. AIM connects to them.

AIM does not replace the ERP, the POS or the WMS; it sits above them. Each connector is an MCP-style tool with an explicit permission — read, propose or write — granted per system and reviewed by you. Most connections start as read. A write is only ever performed by an action that has passed the checks and the trust tier above.

PackTypical sources and systems
RetailPOS · ERP (Pronto, SAP) · WMS · replenishment · email · chat · spreadsheets
ManufacturingERP · MES · SCADA tags · maintenance · email · chat · spreadsheets
HospitalityPOS · bookings · stock · rostering · email · chat · spreadsheets
LogisticsWMS · TMS · telematics · ERP · email · chat · spreadsheets

Nothing ripped out · your systems stay · every connection starts as read

Your page builds itself

Not a fixed screen. A page scoped to a role and a goal.

There is no dashboard to design. Each person gets a page generated from the map, scoped to their role and the goal they own — the same objects, seen from where they stand. It is one responsive application, and it is the same one on the office monitor and the phone on the store floor.

CEO

All stores · today

Sales against plan, the six stores driving the gap, the one decision waiting on them.

Region head

26 stores

Ranked by gap, the weekend conversion dip, the two stores worth a visit.

Store manager

Store 214

Best Tuesday this quarter, today's front-table swap, Saturday's roster gap.

Yours, in the open

Your map is yours. Open format, portable.

The confirmed ontology exports to Open Semantic Interchange, the vendor-neutral format published as v1.0 in January 2026 and now carried by the Apache Ossie incubator project since June 2026, with more than sixty member organisations behind it. Objects, properties, links, the rules AIM has remembered — all of it leaves with you in a file you can read. No lock-in by design.

Where it runs today

Live in retail. Built for the rest.

The retail operating system is live, in daily use, inside a national retail network: the map is rebuilt nightly from the exports, and the briefs land every morning. The manufacturing, hospitality and logistics packs shown in the film are built for those industries — their reference ontologies, sources and actions exist — but they are not yet live at a customer. We say so because you will ask.

AIM runs inside your network. Your exports, your map and your rules stay on infrastructure you control; your data never leaves.

Share one data file

Start with one file.

A month of POS, a stock-on-hand dump, the planning sheet — any one of them. We draft the map from it and bring the first morning brief back to you.

Share one data file — first brief in two weeksBack to the film