PlatformSecurity
Safe, Secure & Private
Harold protects your knowledge and your data with enterprise‑grade security and compliance.
AI processing in the EU
Every model call goes through an EU router to EU deployments: Claude on AWS Bedrock in the EU and embeddings on Azure in the EU. Requests to a non-EU model are refused in code.
Zero retention, no training
Model providers are used only on terms with zero data retention and no training. Your documents, emails and appraisals are never used to train a model.
Isolated by organisation
Every record belongs to one organisation. Row-level security in the database and organisation checks on every API route keep it there.
The right people only
Private files, personal mailboxes and restricted contact groups stay with the people allowed to see them. Someone a mailbox is shared with can read it; only its owner can send or change it.
Encrypted connections
Sign-in tokens for connected apps are encrypted with AES-256-GCM, refreshed on the server and sent only to that vendor. Each person connects their own accounts.
Short-lived file links
Files are served through signed links that expire within minutes, so a copied link stops working.
Guarded against prompt injection
Text from emails, documents and web pages is fenced as untrusted data, never followed as an instruction. Anything that sends, shares or changes a record waits for a person to approve it.
A record of what happened
Stage moves, approvals, task changes and activity in client spaces are written to history, so you can see who did what and when.
Built to stay up
Per-person rate limits, background work with its own time limits, protection against requests to internal addresses, and a switch to turn off any outside service at once.
Reviewing Harold for your firm?
We will give your security and IT teams what they need to sign off.
Request our security packSecurity pack
An overview of how Harold is built, where data is processed and the controls described on this page.
Your questionnaire
Send us your firm's security questionnaire and we will complete it.
Data processing terms
Ask us for our data processing terms and the list of providers that process your data.
Secure by design
Security is part of how Harold is built, not a layer on top. Every query is scoped to your organisation and the person asking. Every model call names its purpose, is logged against your organisation and runs on an EU deployment. Every outside service can be switched off on its own without taking Harold down.
Harold does the reading and drafting, and people make the decisions. Emails, documents and web pages are treated as material to read, never as instructions to follow, and nothing leaves Harold, changes a record or reaches a client without someone saying go.
Where is our data processed?
AI processing runs on EU deployments of the models (AWS Bedrock in the EU and Azure in the EU), reached through an EU router. Harold refuses to send a model request to a non-EU model.
Is our data used to train AI?
No. Harold only uses model providers on zero retention, no training terms.
Can another organisation see our data?
No. Every record belongs to one organisation, and both the database and every API route check that the person asking belongs to it.
Can Harold send an email or change something without us?
No. Harold reads and drafts on its own, but sending, sharing or changing a record waits for a person to say go.
Who on our team can see what?
Each person's mailbox and private files are theirs. Admins can restrict contact groups to named people, and client guests see only the space shared with them.
Do you hold SOC 2 or ISO 27001?
Talk to us about where we are with formal certification. We will share our security pack and complete your security questionnaire.
How do we run a security review?
Get in touch. We will send our security pack, answer your questionnaire and take your team through how Harold handles your data.