Security

Your data never leaves your control.

We don’t ask for trust. We build systems you still control — and we build the controls first, before any workflow touches production.

—: Your perimeter — the dashed boundary; lake, harness and model access sit inside it
✕: Customer data leaving — stopped at the edge

Your company’s tenant as a perimeter. Data flows in from your systems and stays inside; a packet of company data trying to leave is stopped at the edge.

The controls, in plain terms.

Zero data retention: Model calls run under zero-retention terms: nothing stored, nothing trained on. Protection at the infrastructure, not by redaction.
Your tenant, your perimeter: The lake, ontology and harness run inside an environment you own. Your data doesn’t leave it.
Least privilege: Every connector and agent gets the narrowest access that works, and nothing inherited by accident.
Permissions mirrored: If someone couldn’t open a record in the source system, they can’t reach it through AI.
Need-to-know walls: People data, contracts and board material live behind walls enforced on every request, not by convention.
Staging, not back doors: Systems that refuse AI access export to a governed staging area. The core system is never opened up.
Checked in a fresh context: Outputs are re-read against their sources by an independent check that didn’t write them.
Evaluated before release: Every workflow and model change runs your eval suite before your team sees it.
Audited end to end: Every prompt, source, user and policy decision is logged and exportable for review.

Redaction protects the data and ruins the answer.

Strip out every name, figure and client before AI sees it, and you pay for a capability you then blind. Zero data retention keeps the data private where it matters — at the infrastructure — so the AI can use the context it needs.

Redacted before AI sees it

Illustrative

How likely is [redacted] to renew? Their contract with [redacted] ends [redacted], and [redacted] raised[redacted] escalations this quarter.

I can’t assess renewal risk without knowing the customer, the contract or what the escalations were about. Could you share more detail?

Private — and useless

Zero data retention

Illustrative

How likely is Northwind Logistics to renew? Their contract with Lumen Co. ends 30 June, and their operations team raised 3 escalations this quarter.

At risk. Two of the three escalations are the same integration fault, open for five weeks, and weekly usage has fallen since March.

Suggest a fix date before the renewal call, and a check-in with their operations lead this week. Sources: 3 tickets, CRM notes, usage report.

Nothing retained
Nothing trained on
Logged in your tenant

The same question, asked two ways. Names and figures are fictional.

Some things stay need‑to‑know.

Salaries, reviews, contracts, board material — and, at investment firms, the wall between public and private sides. Walls are enforced on every request, by people and by agents, not left to convention.

4.3a: Company-wide — every team, working on information anyone may see
4.3b: Need-to-know — the People team, holding salaries, reviews and personal records
4.3c: Shared context — where company-wide requests reach the model
4.3d: People context — kept separate, never shared across
4.3e: Need-to-know wall — enforced on every request; the red reference is stopped and logged

A need-to-know wall. Company-wide knowledge flows freely; People data is reachable only by the people and agents cleared for it.

The walls go up before anything goes live.

Guardrails, walls and audit are phase three — before a single workflow reaches production — so everything we build afterwards inherits them instead of retrofitting them.

Guardrailspolicy v14

Illustrative

Need-to-know: People data stays in People workspacesEnforced
Zero data retention on every model callEnforced
Customer contracts: Legal and account owners onlyEnforced
Answers must cite a sourceEnforced
Irreversible actions need sign-offEnforced
Evals pass before releasePassing

Caught in one morning · illustrative

08:14:02Redacted
Prompt pasted a customer’s personal details — redacted
08:31:47Blocked
Sales user asked for salary data held by People
09:02:15Held
Answer lacked a citation — returned for retrieval
09:40:26Held
Agent drafted a supplier cancellation — held for sign-off

The policy layer and what it caught in one morning. Every later workflow inherits these rules.

Audit logexportable · retained per policy

Illustrative

Time	User	Team	Action	Sources touched	Policy
09:31:04	am.s2	Sales	Renewal risk question	CRM · 3 tickets	Passed
09:28:51	ops.o1	Operations	Supplier renewals	Documents · ERP	Passed
08:31:47	am.s5	Sales	Asked for salary bands	—	Blocked
08:14:02	agent.t2	Support	Support triage	Ticketing · 2 docs	Redacted
07:52:19	lead.c1	Leadership	Board update draft	Documents · 214 files	Passed

The audit trail your security team can pull without asking engineering. Users and actions are fictional.

Access matrixmirrored from source permissions

Illustrative

Role	Customer records	Contracts	People data	ERP records	Support tickets
Leadership	Allowed	Allowed	Summary only	Allowed	Summary only
Sales	Allowed	Allowed	Blocked	Blocked	Summary only
Operations	Summary only	Summary only	Blocked	Allowed	Allowed
Legal	Summary only	Allowed	Summary only	Blocked	Blocked
People	Blocked	Blocked	Allowed	Blocked	Blocked
Support	Allowed	Blocked	Blocked	Blocked	Allowed

Access mirrored from your source systems, by team and data domain.

Bring your security team to the discovery call.

We’d rather answer their questions before we build than after.

Book a discovery call