Skip to content
instructed intelligence

For IT teams

How it works, and how to run it.

The components, the data flow, identity and access, data handling, models, deployment and rollout.

Book a technical call (opens in a new tab)Read the docs

01 · The parts

Six components, one tenant, one sign-in.

Cornerstone
The knowledge layer. Your documents are distilled into typed facts (rule, decision, process, concept, reference), each carrying source, author, date and scope, held in a per-tenant store. Superseded facts are dated and kept; only the current one is read.
Foundry
Ingestion. Runs on your own hardware, CPU only, and distils documents into facts at near-zero token cost. Nothing leaves your environment at ingestion.
Coordinator
Work records. Tickets, who holds them, what was decided, and outcomes written back to Cornerstone as facts. Every session starts from the current record.
Vista
Administration. People, teams, licences, roles and access, in one console. Nobody else needs to open it.
Workbench
The client. An extension for VS Code and Cursor that connects the user's own AI (Claude or Codex today) to Cornerstone, the Coordinator and support through the Model Context Protocol. One sign-in; tools appear in the next chat.
Habitat and Validator
Enterprise. Habitat is the live view across every project and initiative. Validator checks delegated agent work before anyone relies on it, with each verdict on the record.

Your organisation’s knowledge

  • Documents

    OneDriveGoogle DriveDropboxBoxNotion+ many more

  • Meetings

    Microsoft TeamsZoomGoogle MeetSlackLoom+ many more

  • Tickets

    JiraZendeskSalesforceGitHubLinear+ many more

  • Decisions

    SlackMicrosoft TeamsGmailConfluenceNotion+ many more

  • Policies

    Microsoft 365ConfluenceNotionGoogle DriveAdobe+ many more
  • Turned into cited facts

Instructed Intelligence

The platform

Cornerstone

The knowledge layer

What your organisation knows, held as typed, cited facts. Each one carries its source, its author, its date and what it applies to.

  • Rule

    EU client data stays in EU regions

    Legal · 12 Mar 2026

  • Decision

    Frankfurt is the default EU region

    Architecture · 4 Jun 2026

  • Current facts, with sources
  • Decisions and outcomes, written back as facts

Coordinator

Keeps work and knowledge in step

What is being worked on, by whom, and what was decided. Every session starts from what the last one learned.

  • Q3 board narrative

    In progress · Finance

  • Milan region: residency rule

    Decided · 4 Sep 2026

Your AI with Cornerstone

  • Cited facts for each question
  • A clear no, when there are none
  • Questions

Your AI with the Coordinator

  • What is in progress, and by whom
  • What was already decided
  • Work produced
  • Decisions made

Workbench

Connect your AI

  • Claude
  • ChatGPT
  • Codex
  • Gemini
  • Copilot
  • Llama

+ many others

If you use multiple LLMs, our Router helps choose which is best for the task.

How data moves

  1. 01

    Documents in

    Foundry reads your documents where they are and writes facts to your tenant. Ingestion involves no transfer to a third party.

  2. 02

    Questions

    A user asks in their own AI. Workbench requests the facts the question needs from Cornerstone. The AI receives those facts, not the archive; that is where the token saving comes from.

  3. 03

    Answers

    The user's AI generates the answer from the facts and cites them. Where that AI runs is the one transfer point in the hosted service, and it is your choice of provider, on your key.

  4. 04

    Decisions and work

    What the user decides is written back as a fact with their name and the date. Work goes onto the Coordinator as a ticket and its outcome is recorded.

02 · Identity and access

Who can reach what, and how that is enforced.

One sign-in
OpenID Connect with PKCE. The session lives in the operating system's keychain, never in a file. The extension is a public client and holds no secrets; signing out removes the connections.
Roles in Vista
Role-based access on every seat. Single sign-on and SCIM provisioning with Enterprise.
Access follows the source
A fact carries the access of the document it came from, through distillation and retrieval. A user is never given knowledge from a document they could not have read.
Classification is declared
Content is marked none, internal or restricted by a user or by the assistant, in the document or on the fact. Restricted content is withheld from the search index and the embedding, and access-controlled by category. Person-identifying or unconfirmed content is held for human review before it is indexed. Nothing is auto-detected, and we say so.

03 · Data

Where it lives, how it is protected, how it leaves.

Where it lives
Hosted by us in the United Kingdom by default, with backups in the same jurisdiction. If your policy needs another jurisdiction, you choose it and we host there. Your cloud or your own hardware if the boundary must be yours.
Encryption
Source documents and originals are encrypted at rest with a key unique to your tenant. Search indexes sit on encrypted disk. Restricted material is encrypted under a separate root and never enters the index.
Backups
Encrypted client-side before they leave the host, held in the same jurisdiction, and restore-tested.
Retention and deletion
Configurable retention with a dry-run purge, so you see what a policy would delete before it does. Per-tenant soft-delete, recover and shred. Per-subject lookup, export and deletion for data-subject requests. On termination, deleted or returned, per the DPA.
What we never do
Train any model on your data, other than within your own tenant at your instruction. See your AI provider credentials. Route your prompts. Run third-party analytics; ours is self-hosted, and the extension collects no telemetry.

04 · Models

Yours, not ours.

Bring your own
Any model, including one you host yourself. The platform works with Claude, Codex, ChatGPT, Gemini, Copilot and others; Workbench for VS Code and Cursor supports Claude and Codex today. Your key, your provider agreement, no markup.
The Router
Where a team uses more than one model, the Router helps choose which is best for the task in hand.
The transfer points
Two, and both are yours to control: the model that answers, and, only if you choose a hosted distillation model instead of Foundry, the model that distils. Run both in-boundary and nothing leaves it.

05 · Deployment

Four ways to run it, and what each one means.

Software as a service
Hosted by us, UK by default or your chosen jurisdiction. Scale Up and Enterprise both run here.
Your cloud
The same platform in your own tenancy. Enterprise.
On-premises
On hardware you own, inside your network. Enterprise.
Air-gappable
Every part of the platform deploys inside the boundary. It is air-gapped only when Foundry does the distillation and the answering model also runs inside; we will not describe your estate as air-gapped for you.

06 · Rollout

A plan that works, in six steps.

The order is deliberate. Value first, policy second, scale third.

  1. 01

    Pilot with one team, on the trial

    Five people, thirty days, their own documents. No procurement and no integration work: install Workbench, sign in, load a folder they already search. The first trusted answer is the milestone that matters.

  2. 02

    Standardise the knowledge guide

    Workbench writes a short behaviour guide into a project's guide file (CLAUDE.md or AGENTS.md): check recorded knowledge before answering and cite it, say so when nothing is recorded, offer to save decisions, treat work as tickets. Put it in your repository templates so every team gets it.

  3. 03

    Decide the classification policy

    What is restricted, what is internal, and who declares it. The controls exist on every edition; the policy is yours to set before the corpus grows.

  4. 04

    Widen team by team

    Each team joins the same platform with its own documents. Access follows the source, so nothing has to be re-permissioned. Habitat gives leadership the estate view as it grows.

  5. 05

    Move to Enterprise when it asks

    Single sign-on and SCIM, your own deployment, a higher document limit, named support with an SLA. A security review we take directly.

  6. 06

    Measure what matters

    Activation: the share of people reaching a first trusted answer. Second user in each workspace. Use on three or more distinct days. Citations opened. The product emits these events; the docs and the in-chat support tool carry the rest.

Requirements for the client: VS Code 1.102 or newer on macOS or Windows, or Cursor with Claude Code; a Claude or Codex account per user. Foundry runs on CPU on machines you already own. Documents per workspace: 10,000 included, more by arrangement.

07 · What it changes

What your organisation gets that it did not have.

One sanctioned route for AI to reach company knowledge
Instead of each team pasting documents into whichever tool it has. The same controls, the same audit trail, whichever model a team prefers.
Fewer confidently wrong answers reaching the work
On 49 questions our test corpus could not answer, Cornerstone declined 41. Every other method we tested guessed. Every answer that is given carries its provenance.
Token spend you can forecast
Around 98% fewer tokens than reading whole documents, with cost per question held within 1.1× where file search swung 232×. Your provider's bill, on your key.
Knowledge that survives turnover
What people decide is recorded once, with a name and a date, and stays readable after they leave. New joiners ask the platform instead of interrupting the person who knows.
A record for audit
Who wrote a fact, when, what it applied to, and every decision that rested on it. When a fact turns out to be wrong you can find the consequences instead of reconstructing them.

08 · Disclosure

What we will tell you, and what we will not.

We will: take your security questionnaire directly and answer it in writing; provide the data processing agreement on request; publish the sub-processor list and keep it current; publish our benchmark with the eight questions it got wrong; show you retrieval and Validator behaving on your own documents in a technical session.

We will not: publish how retrieval or Validator work internally. They are the subject of patent filings, and we protect them the way you would. You get their behaviour, their outputs and their provenance, and a contract that says what they do.

FAQs

What network access does Workbench need?+

Outbound HTTPS to your workspace and to your own AI provider. Nothing inbound, and nothing to install on a server for the client.

Can we run Foundry ourselves and use the hosted platform for the rest?+

That is the standard shape. Foundry runs on your hardware in every deployment; the hosted platform holds the facts it produces.

What audit logging is there?+

A per-tenant log of what was asked and what was shown, work records of who did what and when, and the provenance on every fact. Retention on each is configurable, with a dry-run purge.

Can we run a proof of concept without production data?+

Yes. Start the trial on a corpus that is already internal-only, and use classification for anything restricted. Nothing you load is used beyond your own tenant.

How do updates work?+

The extension updates through the Marketplace; the download route is manual. The hosted platform is updated by us. In your own cloud or on-premises, updates are part of the Enterprise arrangement.

Is there an API beyond Workbench?+

Workbench is the supported client. Your AI reaches the platform over the Model Context Protocol, the open standard for connecting assistants to tools. If you have another client in mind, ask us.

Bring the questions. We will answer them on your documents.

A technical call with the people who built it, or start with one team on the trial and assess it running.

Book a technical call (opens in a new tab)Start a trial