SHAPESHIFTER
Case study · food manufacturing · Ontario · August 2026

A bakery group gave every employee a computer with an AI in it. Then a manager built the company's own software on it.

A multi-company bakery group with no software developers on staff runs ShapeShifter on its own instance. Seven weeks in, one manager has built ten purpose agents, published nine internal apps and wired the group's ERP into chat. The platform's fix loop closes his bug reports without a ticket queue.

Customer details and references on request Download as PDF
At a glance

Live on the customer's own instance since July 2026

13
people on the instance, each with a private workstation and their own AI
10
purpose agents built by one manager through conversation, no code
9
internal apps published, three promoted to the whole company
3
ERP company databases connected to an agent that answers from live data

Figures read from the customer's instance on 26 August 2026. The customer is a bakery group with several operating companies and a 22-person People directory on the platform; its name is withheld here and available on request.

The situation

Several companies, one small office, no developers

The group runs a handful of food brands out of a shared head office. Its information lived in three separate ERP company databases, a bakery production system, shared mailboxes for vendor paperwork, and the usual spreadsheets. Nobody on staff writes software, so every process ran the way the vendors' software allowed it to run.

The buying question was the one most mid-sized companies are asking this year: how do we give our people AI without pasting company data into consumer chatbots, and without buying another point system we do not own?

What they did

One instance, one builder, and everything else followed

01 · OWN THE INSTANCE

A private ShapeShifter instance for the group

One install, holding only their data, with a persistent private workstation for every employee. Staff sign in with their existing Microsoft accounts. The company's domain, mail and identity stay in the company's hands.

02 · ONE MANAGER, BUILDER RIGHTS

A manager, not a developer, builds the company's tools

He was granted builder rights on two system apps. Working in chat, he shipped an org-mapping canvas, bulk manager tools and other People features through an AI-reviewed gate: five submissions merged and deployed in the first two days, none rejected.

03 · AGENTS ON REAL DATA

Ten agents, made by asking

An ERP agent that signs in to three company databases and answers from live records. Agents for the bakery production system, payroll data, brand, roadmap and a strategy agent per operating company. Each one has a plain-language page anyone in the company can read.

04 · PUBLISH AND PERMISSION

Apps shared to chosen colleagues, enforced at the link

Nine internal apps published from the build studio; three promoted into the company sidebar. A sales-intelligence app is restricted to named viewers, and the restriction is enforced at the link itself, not just hidden in a menu.

05 · PEOPLE AS THE CONTROL PLANE

The org chart runs onboarding, offboarding and access

A 22-person directory with onboarding and offboarding workflows running on the platform. Who sees what follows the reporting tree, and the AI inherits the same boundaries.

06 · THE PLANT FLOOR

Reaching the production machine

A machine on the plant floor is enrolled through the ShapeShifter bridge, so the workspace can reach files on it with default-deny grants the company controls. Line tracking with a camera is the next experiment.

The bugs he finds go into chat. The platform triages, fixes, reviews and deploys the fix into his instance. Updates flow down daily, canary first.
How it works

Why this was possible without a development team

PieceWhat it means for the customer
A real computer per employeeEach person has a persistent Linux workstation on the company's server, running Claude Code as its engine, next to the company database, the org chart and the work board. It remembers everything and can build databases, apps and automations in that person's account.
Your own AI accountEach person connects their own Claude account for interactive work. Scheduled and background work runs on a per-tenant key. The company never becomes the AI billing middleman.
The org chart is the security modelAccess follows the reporting tree; grants are the single permission kernel, and the AI cannot step outside them.
Nothing leaves without a humanAnything that leaves the company's control, an email, a payment, a publish to the internet, waits for a person to confirm. Every agent action lands in a ledger the company holds.
Build, publish, promoteDescribe the tool in the studio, get it wired to real data, share it over a zero-trust link, promote it to a company app. The code is the customer's.
Software that fixes itselfBug reports from chat go to a desk that fixes, reviews and deploys. The customer's instance takes updates daily from a control plane the customer never has to run.
What's next

The pattern, not the project

The group is extending the same loop to vendor paperwork intake from its shared mailboxes, deeper ERP reporting, and a site-to-site link so the platform can reach the plant's other machines. Every one of those is the manager describing what he wants and the platform building it inside his own account.

The interesting part is not any single app. It is that a company with no developers now changes its own software at the speed the business changes, on a computer it owns.

ShapeShifter

An agentic operating system: a private computer for every employee, on a server your company owns, where AI builds the software you run.

Every employee gets a persistent workstation with a frontier agent in it. The apps, chat, files, a shared browser, a build studio, an AI chief of staff on iPhone, iPad and Mac, and a work board for people and AI, are windows onto that same computer.

Customer details, references and a live walkthrough on request
Figures verified on the customer's instance, 26 August 2026 · PDF