The Core Distinction: Consumer Tier vs. Enterprise/Business Tier
The single most consequential decision in using AI tools with client-confidential material is which tier of the tool you're using. Free, consumer-facing AI chat tools frequently retain submitted content and may use it to improve future models unless a user has explicitly opted out — and opt-out settings vary by provider and can change. Enterprise or business-tier accounts from the same vendors typically come with contractual no-training-on-inputs guarantees and, often, a formal data-processing agreement (DPA). These are not the same product wearing a different price tag — the data-handling terms are materially different, and only the enterprise tier should be considered for genuinely confidential client material, and only after someone has actually read the current terms rather than assumed they match a prior version.
What "Enterprise Tier" Does and Doesn't Guarantee
An enterprise-tier no-training guarantee addresses one specific risk (the vendor using your data to train future models) but does not automatically address others: the vendor may still log inputs for abuse-monitoring or debugging purposes for some retention period, subprocessors involved in running the service may have access, and a security breach at the vendor is still possible regardless of contractual terms. Reading the actual current data-processing agreement — not a marketing summary of it — is the only way to know what's actually guaranteed for a specific tool.
Practical Steps Before Submitting Anything to an AI Tool
- Confirm the tool and tier are approved under the firm's AI-use policy before submitting anything client-related — this should already be settled, not decided ad hoc per submission.
- Strip identifying information — remove client name, project address, project number, and any other identifying metadata from drawings, specs, or text before submission, even to an approved enterprise-tier tool, as a baseline habit rather than a reaction to a specific sensitivity level.
- Redact or exclude PII specifically — names, contact information, and any personal data belonging to individuals (not the firm or client organization) deserve extra caution, since PII carries separate legal exposure (privacy law) beyond ordinary business confidentiality.
- Avoid uploading full drawing sets or specification packages when a smaller excerpt would answer the question — submitting only the specific detail, section, or clause relevant to the query reduces the blast radius if something goes wrong, compared to uploading an entire project file.
- Check contract language — some client contracts include specific confidentiality clauses that may restrict or require notification before using third-party AI tools on their project material; confirm this before assuming firm policy alone governs.
Local and Self-Hosted Alternatives
For firms with genuinely high sensitivity requirements (government, defense-adjacent, or contractually restricted work), self-hosted or on-premises AI models avoid the third-party data-handling question entirely, at the cost of more setup and typically less capable models than the leading cloud-hosted options. This is a real trade-off worth weighing against actual sensitivity requirements rather than defaulting to either extreme — most ordinary commercial engineering work does not require this level of isolation, but some does.
What to Do If Confidential Data Was Submitted to an Unapproved Tool
Report it promptly per the firm's incident-reporting process rather than treating it as a minor slip to quietly move past — early disclosure gives the firm the chance to check the specific tool's actual data-retention terms and, if needed, notify the client per contract requirements, which is far better than the client discovering it independently later.