Internet SI OS

Investor overview · October 2026

The operating system for enterprise superintelligence

By Internet Labs

A long records-room aisle at blue hour, a lamp lit at its far end

Contents

The argument

Seven chapters, one system

01
The problem
Enterprise AI stalls on the company’s own knowledge
03–04
02
The system
One install compiles what a company knows and puts it to work in every department
05–07
03
The technology
Compile, share, route, protect and learn, inside the client’s walls
08–13
04
Build and control
Every team builds in plain words; two people approve every change; IT sees it all
14–18
05
The business
A paid Proof, a setup from $5M, a $304B market
19–22
06
The plan
Twelve months to build the OS, gates inside each client, Valor after
23–25
07
The pre-seed
$5.5M to build the OS and reach Gate 1
26

The problem

Frontier models have read the internet

The knowledge that runs an enterprise
was never on it

$173M
average planned AI spend over the next 12 months at $1B+ US companies, up 33% on a year
37%
of companies see any EBIT impact from AI
72%
of US leaders say they lack unified, accessible data

The main constraint is the enterprise’s own knowledge: fragmented, ungoverned and unsafe to hand to autonomous agents.

The problem

Spending rises; results don’t

Four leaks between the budget and the P&L

Planned AI spend

$173M

per $1B+ US company, over the next 12 months (KPMG, Sep 2026)

Knowledge
60%
of AI projects without AI-ready data abandoned through 2026 (Gartner)
Compile the foundation
Trust
65%
have seen agents act beyond their scope; 34% check authorisation when an action runs (EMA)
Keep it permissioned and cited
Ungoverned building
77%
say AI adoption is outpacing their governance (IBM)
Let every team build, under IT’s control
After the vendor leaves
70%
of agentic AI built by vendors’ forward-deployed teams abandoned by 2028 (Gartner)
Leave the client able to evolve it

What the platform must do

Reaching the P&L

37%

of companies see any EBIT impact from AI (McKinsey, Aug 2026)

42%

abandoned most AI initiatives in 2025, up from 17% (S&P Global)

Internet SI OS closes all four, inside the client’s own environment.

The system

One system for the whole company

Four technologies at its core

Internet SI OS compiles what a company knows into governed memory, puts expert agents to work in every department, and learns from every day of work, in the company’s own environment and under human approval.

01
Memory
vault
Every document, mail, ticket, recording and scanned archive, compiled into cited, permissioned pages.
02
Pragmatic
model routing
Sensitivity decides first, then difficulty: every request reaches the model that serves it best.
03
Reversible semantic
pseudonymisation
Identifiers become realistic twins, so models learn without seeing a real identifier.
04
Continual
learning
Finds its own gaps, tests fixes on real work and keeps what passes.

One install: the whole company
Unlimited departments: added as data
Any environment: cloud, on-prem or air-gapped
Any model: on no single vendor

The system

How it works

Knowledge in, intelligence out, every change approved

Daily work · read bottom to top

Company material
Scanned paper
Knowledge Compiler a source on every page
Memory vault
Orchestrator pragmatic routing
Expert agents
The Emissary
Approved models
People, in one workspace

Improvement · read top to bottom

Capture
Agents propose
The owner approves
Build or learn
Tested on the living benchmark
A person releases it

The system

One install, every department

Nine presets ship ready; any other is added as data, not code

HR
Leave requests, onboarding checklist, policy questions
Finance
Invoice approvals, expense checks, month-end checklist
Sales
Proposal drafts, discount approval routing, account research
Marketing
Brand review requests, campaign briefs, launch checklists
Design
Asset requests, design review intake, research summaries
Engineering
Incident review drafts, runbook lookups, on-call handover
Legal
Contract intake against the playbook, NDA requests
IT
New-hire provisioning, access requests, leaver offboarding
Security
Alert intake and triage, phishing triage, endpoint isolation with analyst sign-off
Any other department
Added as data: no code change, no redeploy

The technology

The Knowledge Compiler

From raw material to cited, permissioned pages

Raw material

A desk covered in loose papers under a lamp

Documents
Decks
Wikis
Mail
Tickets
Recordings
Paper

The compiler

Map owners and permissions, before reading anything
Label sensitive files before they are taken in
Screen every source for planted instructions
Compile versioned pages, a source on every claim
Recompile only what changed, with one-step rollback

A compiled page Example · Finance

finance / wiki / expense-approvals.mdFinance · all staff

Expense approvals

Claims above a manager’s limit go to Finance for approval.¹

Travel booked outside the travel tool needs a written exception.²

Every claim carries its receipts.^{1 3}

1 Expenses policy.pdf · page 4
2 Travel guide · Finance wiki
3 Month-end checklist.xlsx

v14 · approved by the Finance ownerunder Git

Every claim opens to its source; access is fixed before the page exists.

~10 days

to compile an enterprise wave (target)

90%

of each department’s in-scope material reached (target)

$7,600

estimated model spend for a whole company wave, on an assumed corpus

The technology

Why compiled beats retrieved

Knowledge that compounds

	Search and retrieval (RAG)	The memory vault
Assembled	On every question	Once, then kept current
What accumulates	Nothing	Every source and every approved answer
Citations	Chunks chosen by similarity	Every claim carries its source
Permissions	Filtered at answer time	Fixed before compiling
Change control	None: the index changes as the files do	Owner approval, one-step rollback
Where it lives	In a search index	Plain files in the client’s environment

The test, in every Proof: the vault against plain retrieval on the same model.

The technology

The Emissary

Answers across departments, never their bases

A person

The asker · in the example, a Sales lead

06Authorisation checked again when an action runs

01

The question

Released after the check

The Emissary

Carries the asker’s identity on every hop

05A second model, from another family, tries to refute the answer

02

Under the asker’s rights

04

Derived material only

The other department’s agent

In the example, Finance’s expert agent

03Answers from its own vault; the base is never exposed

Example

“Can we offer this client a deeper discount?”

“That is above the standard limit; it needs Finance sign-off.”

Finance’s approval limits never leave Finance; only the answer does. And only people start these requests, never agents.

How it is proven

Leak tests
across every pair of departments, on every build.
Identity on every hop,
so an answer can never carry more than the asker may see.
A second department live
is the test the Emissary is switched on against, in v2.

The technology

Pragmatic routing

Sensitivity decides first, then difficulty

A request

01
Sensitivitysets the ceiling
02
Difficultypicks the first rung

Regulated data stops here

Department adapterInside the walls · tried firstChecked against cited sources
Open modelThe client’s GPUsChecked
Check fails: escalate
Frontier modelAn approved providerChecked
Only if the data may leave

up to 98%

lower cost at GPT-4-level accuracy on benchmark tasks (FrugalGPT, Stanford, 2023)

85%+

lower cost at 95% of GPT-4 quality on MT Bench (RouteLLM, 2024)

The OS reports its own mix from the first pilot.

Every route, cost and outcome is logged; the router changes only after passing the living benchmark. No single model or routing vendor sits on the critical path.

The technology

Tested on 500,000 financial-services cases

Models learn without seeing a real identifier

	Keeps meaning?	Restorable?
Redaction
Masking with placeholders	partly
Tokenisation
Reversible semantic pseudonymisation

96.2%

answer fidelity after twinning and reversal against 41.3% for masking, 12.8% for tokenisation

1.5%

leak rate under an active extraction attack against 7.8% and 22% for two open-source detectors

100%

of substitutions restored exactly

One identifier, on its way through · an example

In the client’s material
Priya Nair
account 40-1187-22 · Leeds
Its realistic twin
Hannah Cole
account 73-5520-09 · Bristol
What the model sees
Only the twin: the meaning holds, the person does not
Restored at the boundary
Priya Nair
account 40-1187-22 · Leeds
with the client’s key

The technology

Continual learning

Memory first, weights last, behind two human approvals

Memory

Every day · undone in one step

Every day

01
Agents log the day’s work to the department’s memory folder.
02
Overnight consolidation proposes cited updates to the wiki.
03
Approval 1The department owner approves.
04
Each update becomes a commit in the vault.

+39%

agent performance with file-based memory and context editing, with 84% fewer tokens (Anthropic, Sep 2025)

Only approved, stable knowledge

Weights

Only when stable · versioned, removable

01
Pseudonymised first.
02
Trains a department adapter.
03
Beats the current one on the living benchmark, with no regression on a general test.
04
Approval 2A person releases it.

−89% · −71%

held-out scores after full fine-tuning and after LoRA on new facts (Meta, Oct 2025): why weights come last

The vault stays the source of truth; every adapter can be removed.

Build and control

Build a workflow in plain words

Described, drafted, tested, inside a governed build space

Workflow studioFinance build space · Invoice approvalsExampleOn a branch · tests passing

1 · Describe it

“When an invoice arrives in the accounts-payable inbox, match it to its purchase order, flag duplicates, send anything over the controller’s limit for approval, and post approved invoices to the ERP.”

Plain words, from a Finance builderDraft the workflow

Or start from a template

Invoice approvals
Expense checks
Month-end checklist

Put a front end on it

Form
Page
Dashboard
Case tracker

Routes decide who can reach each one, inside or outside the company.

2 · See it on the canvas edits stay in step with the code

TriggerAn invoice arrives in the AP inbox
Tool · ERPFind its purchase orderschema
CodeFlag duplicates, reconcile the totalsnumbers
AgentThe Finance expert reads the approval limitcitations
ApprovalThe controller signs off anything over itsign-in
Tool · ERPPost the approved invoiceonce

Each step’s output is checked before the next step reads it.

3 · Or as code, in Git

id: finance-invoice-approvals
space: finance
consumers: [ap-team]
trigger: { mail: accounts-payable }
steps:
  - { id: po,    kind: tool,  tool: erp.find_po }
  - { id: dupes, kind: code,  check: [reconcile] }
  - { id: limit, kind: agent, agent: finance-expert }
  - { id: ok,    kind: approval, for: send,
      role: controller }
  - { id: post,  kind: tool,  tool: erp.post_invoice }
# caps on steps, spend and time, set by the owner

Who can start it: The AP team
Connectors: ERP and the AP inbox, as the space allows
Credentials: Injected at run time, never shown
Tests: Sample invoices, then the living benchmark

Build and control

Two people for every change

And a person for every action that sends, spends or publishes

Every change

01
On a branch
A workflow, app, skill, agent or fix
02
Tested
Sample data, then the living benchmark
03
The owner approves
With a fresh sign-in
04
A person releases
One-step rollback
05
Live on its trigger
06
The console watches
Runs, failures, spend, savings

Agents propose and draft.
They never approve or release anything.

Autofix and Autotune
propose a fix on a branch

Every action

Approval neededInvoice approvals · step 5 · example

Post this invoice to the ERP?

Above the controller’s limit, so it waits for you.

Supplier: Northgate Facilities Ltd
Amount: £18,400.00 · matches its purchase order
Limit: Approval limit, Expenses policy p. 4
Acts as: You, through one short-lived token

Approve and sign inDecline

Exact arguments. Values from document text are marked; the approval matches them and is used once.
Standing approvals only as owner-signed rules with limits on amount, recipients and volume, logged on every run.

Build and control

Who can do what

Roles come from the client’s own groups;
agents propose, people decide

An open key cabinet and a sign-in ledger under a lamp

The client’s identity groups

Finance-UK
AP-Team
Controllers
Finance-Builders

Bound by group id,
never as a list of people

Roles in the Finance department

Member
Builder
Approver
Releaser
Owner

Plus, install-wide, the install owner in the client’s IT.

	Member	Builder	Approver	Releaser	Owner	Install owner	Internet Labs	Agents
Use the OS, run allowed workflows
Build workflows, apps, agents
Approve a change to go live
Release, or roll back
Approve a send, spend or publish of their own
Connect a system, bind groups
Read a department’s sources
Set AI budgets

Decides, with a fresh sign-in
Does it
Proposes or drafts
Only within limits: own rights, a time-boxed setup grant, or as the person
Never

Build and control

IT and security see everything

Every workflow, app and agent:
who built it, what it touches, what it costs

Operations consoleInside the client’s environmentExample data

Inventory
Monitoring
AI spend
Access map
Duplicates
Savings ledger
Dependencies
AI gatewayv2

Inventory

Every workflow, app and agent across the company

All spaces
Finance
Sales
HR
Legal

Name	Kind	Space	Built by	Touches	Last run	State
Invoice approvals	Workflow	Finance	A Finance builder	ERP · AP inbox	2 min ago	Live
Finance expert	Agent	Finance	Preset	Finance vault	1 min ago	Live
Expense checks	Workflow	Finance	Preset	Expense policy · Jira	14 min ago	Live
Month-end checklist	Workflow	Finance	Preset	ERP	Autofix proposed a fix	Awaiting owner
Discount approval routing	Workflow	Sales	Preset	CRM	38 min ago	Live
Onboarding checklist	Workflow	HR	Preset	Mail	Yesterday	Live
Contract intake	App	Legal	A Legal builder	SharePoint	Not yet	On a branch

Build and control

Inside the client’s walls

Designed for a Fortune 500 security review

Approved model providers
Only the ones the client approves
Model calls
Valor API
Zero retention, if the client chooses it
Requests
Internet Labs
Holds no client data; no shared control plane
No line in

The client’s environment

Its cloud account · on-prem · air-gapped with the Sovereign edition

Memory vaults
One per department, plain files under Git
Workflows and apps
Code in Git, shown as a canvas
Logs and audit trail
Every write, with one-step rollback
Keys and the twin mapping
In the client’s key store
Adapters and local models
Regulated data reaches only these
The operations console
IT and security see everything built and run

Permissions before content
A source on every claim
Two human approvals
Every step sandboxed
Credentials never visible

Certifications

01
SOC 2 Type I
before the first paid Proof
02
SOC 2 Type II and ISO 27001
within 12 months of it
03
Compliance packs on request
HIPAA, ISO 42001, FedRAMP Moderate

The business

How a deployment lands

A paid Proof first, a training month last

10–12 months · value from the first department

Discovery
Shown in person, under NDA
The Proof
Paid, fixed scope, one department; credited against setup
Waves
One department at a time, each live before the next
Handover
About a month training the client’s team
Every year after
License and subscription; departments added at their setup price

Inside the Proof · one department, paid, fixed scope, credited against setup

01Its compiled base
02Its expert agent
03Its first workflows
04A side-by-side test against the current tool
05A baseline for the savings ledger

Thresholds fixed in writing before the test runs; every deployment reports its lift over plain retrieval with the same model.

A pod of four delivers it

Forward-deployed engineer
Forward-deployed engineer
Data and platform engineer
Delivery lead

About one deployment per pod a year.

A ring of old keys beside a closed folder on a wooden table — Handover closes only when the client’s team does every job unaided.

The business

Priced off one number

The setup fee: from $5M, then 70% of it every year

Setup, once

$5M$10Monce P&L case studies exist

From $5M for the first ten deployments

Every year after

70% of setup$3.5M a year on a $5M setup

20% license50% platform subscription

Setup per department, by edition

Core$300K
The shell, Knowledge Compiler, expert agents, routing, memory vault
Enterprise$450K
Core plus the Emissary, the build loop, adapters, governance and audit
Sovereign$675K
Enterprise plus on-prem or air-gapped, data residency, compliance packs

On top, as it is used

Seats: About $25 a seat a month, realised
Agent actions: $0.07 above the included volume
Voice: $0.12 a minute
Desk resolutions: $1 per request an IT, HR or finance desk resolves
Third-party models: List price plus 5%
Valor: What it costs to serve, plus a margin
Adapter training: Compute at cost

Seats are a planning assumption. Support, compliance, residency and services are priced as add-ons.

The business

A $304B market for the OS

Every company with $1B+ of revenue, at our price

Revenue	Buyers	Package a year	Market a year
$10B+	2,000	$70M	$140B
$5–10B	2,000	$22M	$44B
$1–5B	13,750	$8.7M	$120B
All tiers	17,750		$304B

Price: Setup $5M+, then 70% of setup a year
First vertical: Financial services: regulated, document-heavy, and where the benchmark ran

The business

Built to win

And why an early lead compounds

A compiled, cited foundation
Retrieval is table stakes; the compiled vault is not.
The Emissary
Departments share intelligence without sharing their files.
Sovereign, on no single vendor
Any cloud, on-prem or air-gapped.
Evolvable after handover
The improvement loop keeps running after the engineers leave.
Every deployment advances the base
The next one starts further along.
Daily use is the switching cost
Leaving means moving the entire company.

Buyers compare us with Palantir AIP, Glean, Cohere North, Microsoft Copilot and the labs’ own deployment arms.

The plan

Twelve months to build

Three teams in parallel, then one system

OS core
Continual learning
RSP

Integrate and test

Month 13First paid Proof
Month 18Gate 1
Proof targets met and a second department signed
On Gate 1The next round
Sized to the evidence Gate 1 produces

Month 3
Compiler and vault, shown on a real department’s corpus
Month 9
Three financial-services design partners; lift over plain retrieval measured

Internet SI OS is pre-revenue. Product brief, technical design and MVP plan complete; this round builds the OS.

The plan

Twenty-three features, three gated phases

Each client moves up only after passing its gate

13 v1 · Core
- Knowledge Compiler
- OS shell
- Company knowledge base
- Hard-copy digitisation
- Department expert agents
- Access that travels with the knowledge
- Multi-model routing
- Memory vault and living benchmarks
- Workflow studio
- Secure runtime
- Operations console
- Build spaces and skills
- IT and Security presets
Gate 1
7 v2 · Connected
- The Emissary
- Agents that suggest skills
- Graph engineering
- A shared agent space
- Apps and cases
- Autofix and Autotune
- AI gateway
Gate 1 opens v2: the Proof’s targets met
and a second department signed.
Gate 2
3 v3 · Self-building
- Self-improving build loop
- Department model adapters
- Partner and MSSP tenancy
Gate 2 opens v3: no permission breach,
and every living benchmark holding.

The plan

It compounds

Every installation that opts in makes Valor stronger

01
Daily work
in the OS
02
Golden data
demonstrations, preferences and tests, approved by owners
03
Valor
Internet Labs’ model line, post-trained inside each client’s walls
04The five-year aim
The default model
for daily corporate work

Frontier models stay in the mix for the hardest tasks; pragmatic routing decides, request by request.

The pre-seed

Raising · pre-seed

$5.5M

Eighteen months to the first paid Proof and Gate 1, plus three months of breathing room.

Instrument and terms open,
to agree with the lead investor

hi@internetlabs.ai · internetlabs.ai

Reaches

Real-corpus demo
Three design partners
OS built
SOC 2 Type I
First paid Proof
Gate 1

Use of funds

People$2.6M
Compute$0.6M
Go-to-market$0.3M
Company, SOC 2$0.5M
Hiring$0.2M
Contingency$0.6M
Breathing room$0.7M

The founding team

Abraham Vincent Philip

Founder and CEO

Ankush Sharma
Founding CTOPreviously Google, PayPal
Ayush Goyal
Founding principal engineerPreviously Microsoft
Vipul Sharma
Founding principal engineerPreviously Meta
Stacie Chan
Founding CMOPreviously Google

First hires

Founding ML research lead · Enterprise sales lead, financial services · Forward-deployed engineer

Internet SI OS

The argument

The problem

The system

The technology

Build and control

The business

The plan

The pre-seed

Frontier models have read the internet

Spending rises; results don’t

Knowledge

Trust

Ungoverned building

After the vendor leaves

One system for the whole company

Memoryvault

Pragmaticmodel routing

Reversible semanticpseudonymisation

Continuallearning

How it works

Daily work · read bottom to top

Improvement · read top to bottom

One install, every department

HR

Finance

Sales

Marketing

Design

Engineering

Legal

IT

Security

Any other department

The Knowledge Compiler

Expense approvals

Why compiled beats retrieved

The Emissary

A person

The Emissary

The other department’s agent

Leak tests

Identity on every hop,

A second department live

Pragmatic routing

Tested on 500,000 financial-services cases

Continual learning

Memory

Weights

Build a workflow in plain words

Two people for every change

Who can do what

IT and security see everything

Inventory

AI spend · this month

AI gateway

Inside the client’s walls

Approved model providers

Valor API

Internet Labs

The client’s environment

Memory vaults

Workflows and apps

Logs and audit trail

Keys and the twin mapping

Adapters and local models

The operations console

SOC 2 Type I

SOC 2 Type II and ISO 27001

Compliance packs on request

How a deployment lands

A pod of four delivers it

Priced off one number

A $304B market for the OS

Built to win

A compiled, cited foundation

The Emissary

Sovereign, on no single vendor

Evolvable after handover

Every deployment advances the base

Memory
vault

Pragmatic
model routing

Reversible semantic
pseudonymisation

Continual
learning