AI security that holds up in review
Your security, privacy and legal teams will have hard questions. We design for them from the first sprint: where data lives, who can reach it, what the AI may do on its own, and how every action can be traced later.
- Where is our data stored and processed?
- In your own cloud account, in a Canadian region by default.
- Is our data used to train models?
- No. Private model endpoints, with training on inputs turned off.
- Who can access production?
- Named people, least privilege, through your identity provider. Every access is logged.
- What can the AI do without a person?
- Only what you approve, within limits you set. Everything else waits for a human.
- Are we locked into one model vendor?
- No. Models are chosen by evaluation and can be swapped without a rebuild.
A reference architecture, inside your boundary
Most of what we build follows this shape. The details change with your stack and your policies, but the boundary does not: the system runs in your account, under your identity provider and your keys.
What each part does
- 1
Identity. Staff use the accounts they already have. Leaving the company removes access to the AI the same day.
- 2
Scoped service identity. The application holds only the permissions its job needs, defined in code and reviewed with your team.
- 3
Approval gate. Actions with money, customers or records attached go to a person first, with the evidence attached.
- 4
Permission-aware retrieval. The assistant can only cite documents the person asking is already allowed to open.
- 5
Scoped connectors. Read-only by default. Write access is granted per action, never as a general key.
- 6
Audit log. Every prompt, source, tool call, approval and output, tied to a person or a policy.
- 7
Evaluation harness. Real cases with known answers, re-run on every change and every model update.
- 8
Your keys. Data is encrypted at rest with keys you control, and in transit with TLS.
- 9
Private endpoint. Model traffic stays on the provider’s private network path, under terms that exclude training.
Canadian regions we deploy to include AWS ca-central-1 and ca-west-1, Azure Canada Central and Canada East, and Google Cloud northamerica-northeast1 and northeast2. If a model you need is not offered in a Canadian region, we tell you before any data moves and record the decision in the privacy impact assessment.
Twelve controls, built in from the first sprint
Security is cheaper to design in than to bolt on before launch. These are standard on every system we build, not options on a price list.
Data
Canadian data residency
Storage, processing and logs in a Canadian region by default, in your account or on your own infrastructure.
Encryption everywhere
TLS in transit, encryption at rest with your keys, and secrets kept in a managed vault, never in code.
Data minimization
We collect and keep only the fields the use case needs, with retention rules agreed in discovery.
No training on your data
Private endpoints under provider terms that exclude training. Your data does not improve anyone else’s model.
Access
Least privilege
People and services get the narrowest access that works, reviewed at each gate.
Your identity provider
Single sign-on and MFA through Microsoft Entra ID, Okta or Google, with roles you manage.
Separated environments
Development, test and production are isolated. Production data does not land on laptops.
Revocable vendor access
Our engineers work through accounts you issue, can see and can switch off at any time.
Accountability
Full audit trail
Who asked, what the AI saw, what it did and who approved it, kept in your region for as long as your policy requires.
Evaluation and red-teaming
Measured accuracy before launch, adversarial testing for injection and leakage, and regression checks after.
Human in the loop
Autonomy is set per action. Anything consequential waits for a named approver.
Model choice without lock-in
A thin model layer and a portable evaluation suite, so a better or cheaper model can be swapped in safely.
You decide what the AI may do alone
Autonomy is a setting, not a leap of faith. You choose the level for each action, and you can lower it at any time without a release.
- 01
Suggest
The AI drafts, a person decides and acts.
A reply to a customer, ready for review
- 02
Act with approval
The AI prepares the action, a named person approves it.
Posting an invoice matched to its PO
- 03
Act within limits
The AI acts alone below thresholds you set. Every step is logged and reversible.
Routing documents under a set dollar amount
- 04
Escalate
Low confidence or a policy rule sends the case to a person with the evidence attached.
An unusual claim, flagged with sources
Every system ships with a switch that returns all actions to human review in seconds.
Evaluation and red-teaming, before and after launch
We don’t ask you to trust a demo. Each system ships with tests that measure it on your own cases, and attacks we run against it on purpose.
PIPEDA and Québec Law 25, designed in
Canadian privacy law shapes the architecture, so we deal with it in discovery rather than in the last week before launch.
PIPEDA and provincial laws
- Purposes identified and documented before data is collected
- Consent and notice reviewed with your privacy team
- Safeguards matched to the sensitivity of the data
- Access and correction requests that can be answered from the system
- Breach records and the facts you need to assess reporting
Québec Law 25
- A privacy impact assessment for the project, as a standard deliverable
- An assessment before personal information is communicated outside Québec
- Notice and human review when a decision is based solely on automated processing
- Privacy-protective settings by default
- Retention and destruction rules built into the system
What the privacy impact assessment covers
- 1Data inventory and data flow diagram
- 2Purposes and the basis for each
- 3Where data is stored and processed, and why
- 4Model providers and subprocessors
- 5Risks, mitigations and residual risk
- 6Retention, destruction and access rights
We are engineers, not your lawyers. Your privacy officer and counsel make the calls; we build to their interpretation and give them the evidence they need to sign off.
When something goes wrong
Every system ships with an incident runbook agreed with your team, so the first hour is about acting, not deciding who to call.
01Detect
Alerts on failed sign-ins, unusual data access, spikes in model usage and evaluation scores falling below thresholds.
02Contain
Revoke keys, pause agent actions or return everything to human review with one switch, without taking the system offline.
03Notify
We tell your named contact promptly, within the window in our agreement, with the facts you need to assess reporting under PIPEDA or Law 25.
04Learn
A written root cause, the fix, and new test cases so the same failure is caught automatically next time.
Built to pass your vendor review
We answer your security questionnaire line by line, with evidence about the actual system rather than generic policy text.
Your own questionnaire, SIG Lite, CAIQ or a custom format from your procurement team.
What this page does not claim
We make no certification claims here, and no vendor can guarantee your compliance. What we give your reviewers is evidence about the system they are approving, running in an environment they already control.
The review package your team receives
- 01Architecture and data flow diagrams
- 02Threat model for the system
- 03Access matrix: who and what can reach which data
- 04Privacy impact assessment
- 05Evaluation and red-team report
- 06List of model providers and subprocessors
- 07Incident response plan and runbooks
- 08Completed vendor questionnaire
Security questions
Bring your security questions to the first call
Invite your IT or security lead. We’ll walk through how the system would run in your environment and what your reviewers will receive.
30 minutes with a senior engineer. No sales pitch, no obligation.
