Trust & Data
Your email never reaches us. Your systems keep the data.
Here's exactly what LevelOneAI sees, what it keeps, and what it never touches. Everything on this page about Operations Workbench Starter is how the product works today.
How the data flows
Starter, step by step
Your inbox
Gmail or Outlook
Your Claude or ChatGPT
Reads the mail with your connector
Short labels
Ref, a few words, company, state, dates
LevelOneAI
Keeps the work record. Never sees the email
The full email stays between your inbox and your assistant. Only the labels are kept. The short details your dashboard shows for a run, like a status line and links back to the emails, are used for that run and never stored.
What we can see
- The labels for each item: reference, a few words, company, state, dates
- The fields it extracts from documents you add, encrypted
- Your mailbox address
- Which themes you asked it to track or ignore
- When a run happened
What we never see
- The full text of your emails
- Email attachments, unless you add one as a document
- The document files themselves, only the fields it needs
- Your email password or credentials
- Your contacts list
- Anything your assistant doesn't send for that run
Built today
How Starter protects your data
Your email never reaches us
Your mail is read by your own Claude or ChatGPT, through the email connector you already use. LevelOneAI never receives your mailbox or your messages.
Short labels only
We keep a reference, a few words, the company, the state and the dates for each item. The details shown on your dashboard are used for that run, then dropped.
Read-only on your email
Starter never sends, drafts, replies to, forwards, deletes, labels or moves an email, and never changes your systems.
Your mailbox address, nothing more
We store the address so the queue knows which inbox an item came from. We never hold your mail password or credentials.
No business content in our logs
Our logs don't record what your work items or emails say.
Secure sign-in you can revoke
Sign-in uses OAuth with PKCE. Codes and tokens are stored as hashes, refresh tokens rotate on every use, and a reused token ends the session. You can disconnect at any time.
Documents: fields, not files
From an export you add, Starter keeps only the fields it needs, such as PO numbers, dates, quantities and amounts, encrypted with a key unique to your workspace. The file itself is not kept.
No AI calls on your content
The reading and reasoning happen in your own Claude or ChatGPT account. Our server doesn't send your content to an AI model.
Built today
Retention and processing
- Your email content is processed by your own Claude or ChatGPT account, under your terms with that provider
- From email, dates are stored and quantities and prices are kept only as a fingerprint. From documents you add, the fields it needs are stored, encrypted
- Trial data is deleted 30 days after a trial ends without a paid plan. After you cancel, you can export your history, and your data is deleted 30 days later
- Source messages are referenced by a fingerprint of their ID, not the ID itself
- Each item's history is content-free: status changes and times, no names or text. It's kept for 90 days after the item is deleted
- Each workspace's data is kept separate from every other customer's
- Removing a team member deletes their queue, mailbox records and sign-in tokens
For an Operations Workbench, retention, where it runs and what's stored are agreed with you during setup.
When it isn't sure, or the data is incomplete
Real email and records are messy. It's built to say what it doesn't know.
If it's not sure, it asks
When an email could belong to more than one order or PO, it shows the candidates with their evidence and asks a person to confirm. It doesn't guess.
If data is incomplete, it says so
If a source is unavailable or a record can't be found, the item is marked incomplete, so your team knows the picture isn't complete.
Every fact is labeled
Confirmed, conflicting, inferred or unknown, with its source. A guess is never shown as a fact.
Starter and an Operations Workbench
The same principles, at two scales.
| Workbench Starter | Operations Workbench | |
|---|---|---|
| Where it runs | In your own Claude or ChatGPT | Agreed with you during setup |
| What it reads | One mailbox, through your own connector, plus exports you add | The email, ERP and CRM sources the workflow needs, through access you grant |
| What it keeps | Short labels, references, dates and the fields extracted from your documents | References, extracted facts, work state, decisions and an audit trail |
| Changes to your systems | None. It's read-only | Read-only by default. Write-back is optional, and every change is approved |
| Who sets it up | You, in a few minutes | Your team and ours, with IT |
How a Workbench stays consistent with your systems
Your systems stay the record. The Workbench compares what they say with what was promised, shows any conflict, and leaves the decision to a person.
Read-only by default
A Workbench starts by reading. It changes nothing in your systems.
Write-back is optional, and always approved
Write-back is optional. Whether approved updates go back to your systems, and which ones, is decided during setup. Each change is shown to a person, with its evidence, before it's made.
Access you grant, scoped to the workflow
It connects through the access you give it, limited to the systems and data the workflow needs, and you can revoke it at any time.
Every step on the record
What it found, what it compared, what was recommended and who approved it.
Your systems stay the source of truth
When your ERP and an email disagree, the Workbench shows the conflict. It doesn't quietly pick a winner.
Where it runs is decided with you
We agree where it runs and what's stored during setup, based on your security and IT requirements.
Questions your IT team will ask
Does LevelOneAI receive our email?
Not with Starter. Your assistant reads it and sends us short labels only. For a Workbench, what it reads is agreed during setup and limited to the workflow.
Can it change our ERP or CRM?
Not by default. Write-back is optional, decided during setup, and every change needs a person's approval.
What happens if we stop?
Disconnect the connector or revoke the access you granted, and export your work history and audit trail if you want them.
Which systems does a Workbench connect to?
Only the ones the workflow depends on, through access your IT team grants and can revoke. It can connect through the ERP's API, a read-only database view or scheduled file exports, whichever your IT team prefers. The method is agreed during setup.
Where does our email go with Starter?
To the employee's own Claude or ChatGPT account, which reads it through their email connector. For work email, use a business or enterprise plan from that provider, so supplier pricing and confidential terms stay under your company's agreement with it.
Do you have a SOC 2 report?
No SOC 2 report yet. We'll walk your security team through the data flow and complete your security questionnaire on request.
Which AI provider, hosting region and subprocessors does a Workbench use?
These are agreed during setup and listed in the security walkthrough, before anything is connected.
What happens to our data if we leave?
You can export your work history and audit trail. Your ERP and CRM keep their own data, because they were the record all along.
What about what it learns about how we work?
The rules and patterns your team confirms (which source wins, supplier patterns, workarounds, who approves) are kept for your company only, encrypted with your own key, never used to train models or to inform another customer's results, and included in your export. LevelOneAI staff can't see them without time-limited access you approve, and every access shows in your activity log. Nothing becomes a rule until someone on your team confirms it.
For enterprise IT and security teams: how we fit your IT review.
Want a security walkthrough?
We'll walk your IT and security team through the data flow for your workflow, before anything is connected.