Security & Trust

The receipts.
Behind the manifesto.

Confidentiality is the floor of the product. This page is how that floor is built — encryption, residency, retention, sub-processors, the AI use policy, and the door we leave open for your security team.

SOC 2 Type II

Planned · 2027

On the compliance roadmap as the customer base scales.

ISO 27001

Planned · 2027

Information security management. Roadmapped to follow the SOC 2 audit cycle.

UK GDPR & DPA 2018

In effect

Processed under UK GDPR & DPA 2018. ICO registration is on our compliance roadmap.

UK DPIA

In preparation

A platform data-protection impact assessment is being prepared as part of GDPR readiness.

01 · Encryption

How your data is held.

Coachee threads, reflections, and session notes sit in a single-tenant database schema per coach, encrypted at rest and in transit, and scoped to the coach’s relationship. Nothing co-mingles across practices.

At rest: AES-256-GCM via AWS KMS. Encrypted with an AWS-managed Customer Master Key (CMK) scoped to the environment. Per-coach key sharding is on the roadmap.
In transit: TLS 1.2 minimum, TLS 1.3 where supported · forward secrecy (ECDHE) · HSTS enabled.
Application-layer: Sensitive fields (session notes, journal entries, AI messages, direct messages) encrypted with the CMK before they reach the database. Plaintext is never written to disk.
Object storage: Every S3 object written under a per-tenant prefix with SSE-KMS using the environment CMK. No client ever talks to S3 directly — uploads and downloads stream through the API.
Region: UK (London, eu-west-2). Data does not leave the UK. Single-region today; no cross-region disaster-recovery failover is provisioned yet.
Key custody: Keys are managed by us in AWS KMS. Customer-managed keys (BYOK) are on the roadmap, not yet available.

02 · Retention & deletion

What stays. What leaves.

You hold the record. Coachees can export their thread at any time and request erasure of their data. Data is retained while the coaching relationship is active and removed on request.

Coachee export: Full record — goals, sessions, reflections, messages — as a structured JSON export, on demand. The coachee owns the record.
Coachee delete: On erasure, a coachee’s uploaded sources and AI suggestions are deleted and derived coaching notes are flagged for redaction. Cryptographic erasure by key destruction is on the roadmap (the CMK is currently shared across the environment).
Conversation summaries: AI-generated summaries are retained alongside the thread for as long as the coaching engagement is active, then deleted with the thread.
End of engagement: Coachee data is retained while the coaching relationship is active and removed on request (see right to erasure). Audit log retained 7 years (UK regulatory).
Backups: Encrypted, in-region, with a 7-day point-in-time recovery window. Backups age out on the same 7-day cycle.
Right to be forgotten: Erasure requests honoured in line with UK GDPR Art. 17. Production data is removed on request; backups age out on the 7-day recovery cycle.

03 · The AI use policy

What the model does.

The AI is a draft assistant for the coach. It is never an autonomous agent. Below is the same table we hand to security teams — the yeses and nos, in plain English.

Model: Claude (Anthropic), accessed via AWS Bedrock in eu-west-2.
Retention: Inference runs through Anthropic UK (Claude via AWS Bedrock, eu-west-2). The contractual retention arrangement is being finalised; this row will be updated with the verified specifics once signed.
Safety layer: Crisis detection runs before every model call. Flagged content is never sent to the model and is escalated to the coach with pre-written resources.
Logging: Token counts and metadata logged for usage metering. Decrypted message content is never written to logs.

Yes

Drafts replies in the coach’s voice, using the coach’s past notes and persona — for the coach to edit, send, or discard.

Yes

Summarises threads on demand, so the coach can walk into the next session already briefed.

Yes

Surfaces themes across a coach’s own practice, for the coach only. Sponsor reporting is aggregated separately, never linked to an individual.

Never autosends a coaching reply. No reply enters the coaching thread without the coach’s edit-and-send step — enforced in the message pipeline, not just the UI. Automated reminders are clearly labelled as system-generated, never as your coach.

Never trains public models. Inference only. Coachee data, session content, and reflections never leave our tenancy for model training.

Never replaces the coach. There is no “AI coach” mode. A safety prompt and crisis-detection layer run before every model call; flagged content is never sent to the model and is escalated to the coach.

04 · Sub-processors

Who else touches your data.

Short list, by design. We use trusted infrastructure for the boring layers and run the sensitive layers ourselves.

Sub-processor	Purpose	Region	Data accessed
AWS	Hosting (Aurora PostgreSQL, ECS, S3), KMS, backups	eu-west-2 (London)	Encrypted blobs only · environment CMK
Anthropic (via AWS Bedrock)	Drafting model inference (Claude via Bedrock)	UK / EU	Prompt + coaching context · not used for model training
AWS Cognito	Identity provider · MFA · OAuth	eu-west-2 (London)	Email, hashed credentials, JWT claims
Resend	Transactional email (sign-in, notifications, summaries)	US	Email metadata only · no thread content
Stripe	Billing & payouts (live once the billing surface ships)	UK	Coach billing data only · no coachee data

A current sub-processor register is available on request.

05 · Authentication

Who gets in.

Identity is the perimeter. Two paths in, the same cookie model out, and no surface for account enumeration.

Identity: AWS Cognito (eu-west-2). JWT claims (sub, email, tenant, role) verified on every API request against cached JWKS.
Email + password: SRP password verifier — the password never crosses the wire in cleartext. Browser signs the challenge client-side.
Single sign-on: Email + password is the sign-in method today. Google SSO (via a custom Cognito domain with OAuth state-nonce and Origin validation) is on the roadmap, not yet enabled.
MFA: TOTP MFA is available to every account (opt-in); enforced MFA for admin accounts is on the roadmap. MFA challenge sessions are httpOnly, 5-minute TTL, cleared on every auth state transition.
Session: Tokens stored in httpOnly, secure, sameSite=strict cookies. Refresh tokens revocable per device by the user.
Account enumeration: Login errors are deliberately generic (“invalid credentials”). Forgot-password always returns success regardless of whether the user exists.

06 · Disclosure & incident

What happens if.

Where to send a report, what to expect back, and how soon. The same commitments we make to your security team in writing.

How to report: Email security@amplivate.ai. PGP key available on request. Safe-harbour for good-faith research.
Response SLA: We aim to acknowledge within 1 working day and triage within 3 working days. Affected parties are notified within 72 hours of a confirmed breach (UK GDPR Art. 33).
Status & post-mortems: A public status page is on our launch checklist. We commit to publishing incident post-mortems within 14 days, with root-cause and the action items we are taking.
Bug bounty: Private programme. Scope and severity table available on request once the production surface is live.

For your security team

Vendor pack, on request.

As we onboard our first customers we’re assembling the pack your security team expects — sub-processor register, data-flow overview, AI use policy, and our answers to a standard security questionnaire. Tell us what you need and we’ll share what’s ready, under NDA.

By joining, you agree we may email you about Amplivate. No spam; unsubscribe any time. See our privacy notice.

Or email security@amplivate.ai.

We’re preparing

Data protection impact assessment (DPIA)
Sub-processor register
Network and data-flow diagram
AI use policy
Standard security questionnaire (CAIQ-Lite)