AstorX Insights

If the AI leaves your cloud, so does your tender data

A tender bundle plus your draft response is a map of your commercial strategy: what you can supply, at what price, with what margins and constraints. Before asking what a procurement AI can do, ask where it does it.

What tender data actually reveals

Teams sometimes treat tender documents as low-sensitivity because the buyer's side is public. But your working copy is not the public tender. It is the tender plus your annotations, your pricing worksheets, your win themes, your internal assessments of which requirements you barely meet. Aggregated across bids, that corpus describes your cost structure, capacity ceilings, market priorities, and negotiating patterns. It is, in a very literal sense, competitive intelligence about you, already compiled and neatly formatted.

The deployment spectrum

Procurement AI vendors offer roughly three models, and the differences matter more than feature lists:

• Shared SaaS: your documents are processed in the vendor's cloud alongside other customers'. This can be run responsibly, but you are trusting the vendor's tenant isolation, their subprocessors, their retention discipline, and their security culture, continuously and invisibly.
• Dedicated hosting: your own instance, still in the vendor's infrastructure. Better isolation; same fundamental question of whose keys, whose logs, whose jurisdiction.
• Your own tenant: the platform deploys into your cloud subscription. Your network rules, your encryption keys, your audit logs, your data residency choices. The vendor operates software; the data never changes hands.

None of these is dishonest. But for data that encodes pricing strategy, the third model removes entire categories of question rather than answering them: there is no vendor retention policy to audit when the vendor never holds the data.

The model-training question

Separately from where processing happens, ask what happens to your data afterward. "We may use customer data to improve our services" is a sentence that should end any procurement conversation until it is clarified in writing. The acceptable answer is contractual and unambiguous: customer data is never used to train models, the vendor's or anyone else's. The AI improving from your team's review decisions within your own instance, meaning configuration, terminology, and rules, is a different thing entirely from your documents improving a model that serves your competitors.

Questions that sort vendors quickly

1. Where exactly are documents processed and stored? Which cloud, and whose subscription?
2. Who holds the encryption keys, and can we bring our own?
3. Is any customer data used for model training? Show the contract clause.
4. What leaves our environment when the AI calls a model, and what is logged when it does?
5. When we offboard, what is deleted, how fast, and how is it evidenced?

A vendor with good answers will have them ready. This is the conversation security-serious buyers always open with. A vendor who treats these as unusual questions is telling you who their usual customers are.