Security overview
You're trusting Nagata with your vendor list, your spend and your contracts. This page explains how we protect that data, in plain terms, and what we don't claim yet.
What Nagata does with your data
You give Nagata three kinds of data:
- Vendor spend records uploaded as CSV: vendor name, amount, date, and optional category or department.
- AI usage records uploaded as CSV or synced from a provider: provider, model, workflow, cost, and optional user.
- Software contracts uploaded as PDF or DOCX: we extract renewal dates, notice deadlines, annual values and seat counts.
We use it to build your vendor inventory, flag renewals and notice deadlines, find duplicate tools, ownerless vendors and AI spend spikes, and produce reports your team can act on.
Nagata does not:
- Cancel contracts, downgrade plans, or negotiate with vendors on your behalf
- Push changes into your SaaS providers
- Sell your data or share it with anyone outside the subprocessors on our privacy page
- Use your data to train any AI model
Tenant isolation
Each customer gets its own workspace, and workspaces are hard isolation boundaries.
- Every business table is keyed to a workspace and protected by PostgreSQL row level security.
- Every API request checks that the signed in user belongs to the workspace in the URL. Anyone else gets a "not found" response, so we never even confirm another workspace exists.
- Automated checks in our test suite scan every API route and fail the build if a query isn't scoped to a workspace.
- Privileged database access is limited to our backend. The browser can never bypass row level security.
Files
Spend files and contracts go to private storage buckets that are never publicly addressable.
- Uploads are checked by content, not just the file extension: contracts are validated by magic bytes, and CSVs are rejected if they aren't text.
- Size limits: 10 MB for CSVs, 25 MB for contracts.
- Contract downloads use signed links that expire after 5 minutes.
- Every upload, download and signed link is recorded in the audit log.
How we use AI
Nagata's AI features run on Anthropic's Claude models:
- Contract extraction reads the text of contracts you upload and pulls out key fields. A person confirms every field before it affects any calculation.
- The chat assistant answers questions using read only lookups that are locked to your workspace, plus web search for market and pricing context. It cannot change your data or reach another workspace.
- Procurement research uses web search plus your vendor list and procurement policy to research tools.
- The support assistant helps you describe an issue. It sees only what you type.
What stays out of the model: API keys, credentials, audit logs and other workspaces' data. Recommendations, alerts, ROI figures and reports are rule based. No model decides what to recommend, and every finding shows its formula and the data it used.
Defenses against prompt injection: fixed system instructions, clearly delimited document text, request size caps, and explicit instructions to treat document and web content as untrusted reference material. Prompts and contract text are never written to our logs.
Neither we nor Anthropic train models on your data. Each plan includes a monthly AI budget, and workspaces can be given hard spend caps so usage can't run away.
Humans make every spending decision
- Recommendations are surfaced with their rule and formula. Nagata never executes them.
- Approving, rejecting, deferring or completing a recommendation requires an owner, admin or finance role. API keys and AI integrations are blocked from these actions at two separate layers.
- Every decision creates a permanent record with who acted, what they decided, and when.
Roles and permissions
| Role | View | Decide on recommendations | Edit settings | Invite members |
|---|---|---|---|---|
| Owner | Yes | Yes | Yes | Yes |
| Admin | Yes | Yes | Yes | Yes |
| Finance | Yes | Yes | No* | No |
| IT | Yes | No | No | No |
| Viewer | Yes | No | No | No |
| AI agent (API) | Yes | Blocked | No | No |
*Finance can edit the procurement policy, but not workspace settings or billing.
Sign in and sessions
Passwords are handled by our authentication provider and are never stored or seen in plain text. Session cookies are HttpOnly, Secure and SameSite, and access sessions are short lived.
What we don't claim
We'd rather tell you upfront than have you find out in a security review.
- No SOC 2, ISO 27001 or HIPAA attestation yet. Nagata runs on infrastructure providers that hold SOC 2 Type II reports, but Nagata itself has not been audited. If your process needs a certification, talk to us before you start.
- No standard DPA published yet. If your legal team needs a signed data processing agreement, email us and we'll work through it.
- No customer managed encryption keys. Data is encrypted at rest and in transit by our providers.
- No on premises deployment. Nagata is SaaS only.
Report a vulnerability
Email security@nagata.ai with steps to reproduce. We acknowledge within 1 business day and aim to fix critical issues within 7 calendar days.
See also our privacy page for what we collect, how long we keep it, and who processes it.