Trust
What Halden does with your data
Halden uses your data to run your agents, and for nothing else.
SOC 2 Type II · in progress
GDPR · DPA available
Data stays in region
Status as of November 2026. Reports are available from security@halden.ai.
The short version
Six commitments that hold for every workspace
If you only read one part of this page, read this one. The sections below show how each one works.
Your data never trains models
Tickets, documents and run logs are used to run your agents and nothing else, by us or by our model providers.
Agents only see what you connect
Each agent gets scoped access to the tools and fields its job needs. The rest of your workspace stays out of reach.
Every action is logged
Each read, tool call and decision is recorded with its inputs and outputs. Export the log as CSV or JSON.
Humans approve risky steps
Payments, outbound email and deletes wait for a person by default. You set the thresholds per agent.
Delete anytime
Remove a run, a connection or your whole workspace. A hard delete clears every copy, backups included, within 30 days.
Region choice: US or EU
Choose where your workspace lives at sign-up. Data, run logs and backups stay in that region.
Data flow
How an agent touches your data
Follow one invoice run from connection to approval. Tickets, CRM updates and every other task work the same way.
01
Connect with scoped access
Pick read or write per object when you connect a tool, such as invoices read-only and ledger write. Keys are encrypted and rotated every 30 days.
02
Agent reads only the fields it needs
A run asks for named fields, like vendor, amount and due date, not whole records. Every read is logged with the run.
03
You approve before anything leaves
Payments, outbound email and writes above your limits pause for a person. Nothing is sent until someone approves.
Connected Payments and Ledger
0.4s
Scoped to invoices (read) and journal entries (write)
Read 226 open invoices
1.9s
Fields: vendor, amount, due date · no attachments
Matched 214 invoices to the ledger
3.2s
12 set aside for a closer look
Email the summary to finance
queued
Runs after your approval
Each step lands in the run log with the fields read and who approved it.
Models and providers
Where each model runs and what it keeps
Swift runs on Halden’s own servers. Steady and Deep run with a model provider under a contract that rules out training and limits retention.
Swift
Fast, low-cost calls
Steady
Default for agent runs
Deep
Careful multi-step reasoning
Hosted where
Halden servers, your region
Model provider, your region
Model provider, US (EU on request)
Retention of prompts
0 days
30 days, 0 with ZDR
30 days, 0 with ZDR
Used for training
No
No
No
Zero data retention option
Always on
Configurable
Configurable
ZDR is zero data retention: prompts and outputs are processed in memory and never stored by the provider. Turn it on per workspace in your data settings.
Evaluation method
How we measure the numbers on this site
Every accuracy figure we publish comes from one eval, run the same way each time. Here is the method, so you can judge the numbers yourself.
Tickets resolved without a human
Internal eval · 1,200 real tickets · Oct 2026
Halden (Steady)
72%
Halden (Swift)
51%
Baseline agent
44%
Rules engine
38%
What the numbers mean
Sample: 1,200 support tickets drawn at random from three customer queues, shared with permission, October 1–28, 2026. Resolved: the ticket closed with no human edit and stayed closed for 7 days. Every system got the same tickets, tools and time limit.
We rerun this eval after every model or playbook change and keep past results in the changelog.
Sub-processors
Who else handles your data
Four providers process data for us, each under a DPA with the same limits we accept. We email customers 30 days before adding one.
Sub-processor
Purpose
Data it receives
Location
Cloud hosting provider
Runs the Halden app, databases and backups
All workspace data, encrypted at rest and in transit
Your region (US or EU)
Model provider
Runs Steady and Deep model calls
Prompts and the fields an agent reads
US or EU
Email delivery
Sends approval requests and notifications
Recipient name, email address and message text
US
Error monitoring
Records app errors so we can fix them
Error traces with customer content removed
US
Last updated November 2026. The DPA lists each provider’s legal entity and certifications.
FAQ
Questions security teams ask us
Missing one? Write to security@halden.ai and someone on our security team replies within two business days.
Not entirely, because the run log is how you audit what an agent did. You can shorten retention, mask fields such as card or account numbers before they’re stored, and export or delete logs at any time.
In the region you pick at sign-up, US or EU. Workspace data, run logs and backups stay there. Deep model calls run in the US unless you ask us to keep them in the EU.
No. We don’t train or fine-tune any model on your data, and our model provider’s contract rules it out too. We improve Halden from product usage metrics, never from the content of your runs.
Deleting a workspace removes it from the live system immediately. Backups roll off within 30 days, and we confirm in writing when the hard delete is complete.
Not today. Enterprise workspaces can run single-tenant in the region you choose, with your own model provider keys. If you need a deployment inside your own cloud account, talk to us.
Need our security pack?
We’ll send the DPA, the sub-processor list and a completed security questionnaire, usually within one business day.