AI Privacy

Where does your data go?

When Nexus uses AI, where does our data go, and could it end up training someone's model? It is a fair question, and the honest starting point is to say exactly what the AI sees. When you ask a question, three things leave your system: the question itself, a description of your database's structure (table and column names, not data), and the small result being summarised, at most 40 rows. That is the entire exposure, it is encrypted in transit, and everything below exists to keep it that small and that controlled.

Your data is not used to train AI models

• Nexus uses Anthropic's commercial API under commercial terms. Under those terms, data submitted through the API is not used to train their models. You do not have to take our word for it: the terms are published on Anthropic's website, and your lawyer can read the same document ours did.
• Our API accounts run under zero-data-retention handling, so prompts and responses are not stored on the provider's side after they are answered.

The AI never has open access to your systems

• When you ask Nexus a question about your business, the AI writes a database query; it does not browse your data. That query then runs inside your own database, through a separate, read-only database account that cannot modify your ERP data. That is not a policy, it is a permission grant inside your own database, and your own IT or DBA can list exactly what that account is allowed to do.
• Every query is checked in code before it runs, every single time, including on saved reports someone shared with you: a single read-only SELECT only, business tables restricted to the data categories you have been granted, and results capped at 1,000 rows and 20 seconds. A report saved yesterday cannot show you more than you are allowed to see today.
• Your ERP permissions follow you into the AI. Nexus reads each user's entitlements from their linked Spire account and strips any right Spire denies, at save and again at every sign-in. Generated SQL touching the ERP's cost, gross-profit, markup or margin columns is rejected in code for users without that right, and the question is refused, not answered. Try it: ask a question you are not entitled to and read the refusal.
• Only the minimum necessary leaves your system: your question and the small result being summarised. Documents in the Vault are answered from passages retrieved from your own library, with citations to the source document, so you can always see where an answer came from and check it. When your library has nothing on a topic, the assistant says so plainly instead of guessing.

You stay in control

• Every question you ask Ask Nexus is logged: who asked, what was asked, and the exact query that ran, including refused attempts, not just successful ones. Flagged misuse is additionally recorded in a separate alert log. The log lives in your own database, and your administrators can read it and clear it themselves.
• One switch in Settings pauses every AI feature for your whole company, instantly, without a support call: Ask Nexus and BI, the Vault librarian and document OCR, inbox triage, drafting and AI search, CRM drafts, summaries and follow-up sequences, campaign and collection drafting, booking and rental assists, and the writing assistant. It is a setting your administrators hold, not a request you make.
• Sign-in requires multi-factor authentication by default (one-time codes by email or SMS; a device you choose to trust can skip the code for 30 days), and your deployment runs on its own database: no shared tenancy, no shared schema, no other customer in the same data.

Check it for yourself

A privacy page is a claim; these are the checks. Every one of them takes minutes on a demo or on your own deployment:

• Read Anthropic's commercial terms yourself; the no-training commitment is theirs in writing, not our paraphrase of it.
• Ask us to flip the AI switch in front of you and watch every AI feature stop.
• Ask a question your login is not entitled to, read the refusal it gets instead of an answer, then open the Ask log and find that refusal recorded against your username.
• Open the Ask log as an administrator and see your own questions recorded there, in your own database.
• Have your DBA inspect the read-only database role the AI's queries run under and confirm what it can and cannot touch.
• Ask the document assistant something your documents do not cover and watch it decline instead of inventing an answer.

If a vendor's privacy page does not survive that kind of inspection, that tells you something too.

Why not a privately hosted LLM instead?

Many businesses ask whether a model running on their own server would be more private. It is a fair instinct, and for most businesses it is the wrong trade. Here is the honest comparison.

• The privacy gap is smaller than it looks. Your business data already lives in your own ERP and database, on your own infrastructure, under your own permissions; that never changes under either approach. What the AI sees in the commercial-API model is transient: a question, a schema, a small result sample, encrypted in transit, not used for training under commercial terms, and not retained under zero-data-retention handling. A private model removes that transient exposure and nothing else.
• Self-hosting adds a security surface; it does not remove one. A privately hosted LLM is a new server estate: GPU machines, an inference stack, model files, and their patches, all of which must be secured, monitored and updated by somebody, forever. Most mid-market IT teams are not staffed for that, and an unpatched inference server inside your network is a worse exposure than an encrypted API call leaving it.
• The quality difference is the product. Nexus's AI writes correct SQL against a real ERP schema, reads scanned documents, and cites its sources. Frontier commercial models do this reliably; the models that fit on affordable private hardware are years behind them at exactly these tasks. A private model that writes worse queries does not just answer more slowly, it answers wrongly, and a wrong number in front of a controller costs more than any subscription.
• Hosted models improve monthly; a private model is frozen the day it is installed. When the provider ships a better model, Nexus tenants get it as a settings change. A private deployment gets a procurement cycle.
• The guardrails live in Nexus either way. The read-only database role, the query inspection, the per-user permission clamps, the audit log and the off switch are all in the application layer. They protect you identically under both approaches; self-hosting adds none of them.
• Where a private model is the right call. A hard regulatory mandate that specific data may never transit to a third party is a real constraint, and we design for it case by case. For document ingestion specifically, Nexus supports a fully on-premise OCR option, so scanned documents can be rasterized, recognized and indexed entirely on your own server where that is required.

For everything else, the commercial API with contractual no-training, zero-data-retention handling, and the application-level guardrails above is more private in practice, more capable, and dramatically cheaper to own.

This page covers how Nexus's AI features handle your data. For how we handle personal information generally, including on this website, see the Privacy Policy.