the principles
Six commitments
A person decides Live
The agent proposes. It cannot pay, post or send, and every action waits for approval.
Your data stays yours Live
We use your files to do your work and nothing else. We don’t train on them. Any tuning is opt-in and written down.
Nothing kept at the model In setup
We run open-weight models and configure the service so prompts and outputs aren’t stored after processing.
Deletion on request Live
Ask us to delete everything and we confirm by email. Configurable retention periods are planned.
Evidence for every decision Live
What the agent saw, what it decided and why, in a log where any edit can be detected.
Honest about gaps Live
We list what isn’t done yet, and we complete your security questionnaire before touching live data.
the status
Where things stand
Most security pages list certificates. This one lists what is live, what we are setting up and what is planned. Until the AWS hosting is verified, pilots use anonymised files only, and we tell you which AI provider processes them.
| Practice | Status | What it means |
|---|---|---|
| A person decides | Live | The agent proposes. Nothing is paid, posted or sent without a person. |
| Tamper-evident audit trail | Live | Every decision is logged in a hash chain, so any later edit to the log is detectable. |
| No training on your data | Live | We do not train on your files or outputs. Any tuning would be opt-in, in writing. |
| Deletion on request | Live | Ask and we delete everything, then confirm by email. Done by hand during pilots. |
| Security questionnaire before live data | Live | We complete yours before touching any production data. |
| Open-weight models on Amazon Bedrock, US region | In setup | No OpenAI, Anthropic or Google API in the path. |
| Zero data retention at the model layer | In setup | Amazon Bedrock does not store prompts or outputs and does not use them for training. Model invocation logging is off in our account. |
| Ephemeral processing mode | In setup | Files are processed in memory and deleted once your results are delivered. We keep only a hash and a decision record. |
| Encryption in transit and at rest | In setup | Standard cloud encryption, enabled and checked. |
| Multi-factor authentication and least-privilege access | In setup | On every account that can reach customer data. |
| Configurable retention periods | Planned | You choose how long results are kept. |
| SOC 2 Type I | Planned | We start when the first customers need it for production data. |
| SOC 2 Type II | Planned | After an observation period following Type I. |
| Independent penetration test | Planned | Before we handle live production data at scale. |
| Single sign-on and IP allow-listing | Planned | For larger teams, once customers ask. |
the details
Questions security teams ask
The answers describe our policy. The status table shows which parts are live today.
What counts as your data?
The files you send us (invoices, purchase orders, contracts, supplier statements) and what we derive from them: extracted fields, findings and drafted emails.
Where do the models run?
Open-weight models run on Amazon Bedrock in a US region. Your invoices are never sent to OpenAI, Anthropic or Google, and Bedrock does not store prompts or outputs or use them for training. Until that is live, pilots use anonymised files and we name the provider in writing.
Do you keep prompts and outputs?
At the model layer nothing is stored: Amazon Bedrock keeps no prompts or outputs, and invocation logging is switched off in our account. In the application we are building an ephemeral mode: files are processed in memory and deleted once your results are delivered, and we keep only a hash and a decision record. Until it is live, evidence packs contain the fields we extracted, and we delete them whenever you ask.
Do you train on our data?
No. Not on files, not on outputs. If you ever want a model tuned on your data, that is an explicit opt-in, in writing, with a description of exactly how it would be used.
Who can see our files?
During pilots, a single person at Loopclose handles your files. Access to the environment that can reach customer data is protected with multi-factor authentication and least-privilege access, as the status table shows.
Can the agent pay or send anything?
No. It proposes, and a person decides. It cannot pay, post to your ledger or send email on its own.
Do you have SOC 2?
Not yet. We plan to start SOC 2 Type I when customers need it for production data, then Type II after an observation period. Until then we complete your security questionnaire before touching live data.
What happens when we stop?
Tell us and we delete everything, then confirm by email.