A deal memo. An engineering design. A client file. An unpublished research draft.
AI can make that work faster. The question is whether your organization understands the information route it has created: who receives the material, what is retained, which account controls apply, and what evidence supports the answer.
The right response is not to ban useful tools. It is to make those choices deliberately.
What the research dispute does, and does not, establish
On September 8, OpenAI announced a claimed Navier-Stokes result. Its account acknowledges concurrent work by Tristan Buckmaster and Levent Alpoge. OpenAI denies accessing their specific work before public release, but says it cannot rule out de-identified product-usage data helping improve its models. That statement leaves a provenance question; it does not establish that the researchers' drafts were used, or that a confidentiality breach occurred. OpenAI's published account
The mathematical claim also needs careful wording. A forced result is not automatically outside the Millennium formulation: Clay's statements C and D include external forcing. OpenAI claims to meet those statements and says it will not claim the prize. Neither the announcement nor this article independently validates the proof. Clay's official formulation
For a business, the useful question is simpler than the dispute: what can we establish about the handling of our own confidential work today?
The account matters as much as the tool
Two employees can use the same AI brand under different data-handling conditions. One may use an individual account. Another may work inside an approved business workspace or through a contracted API.
OpenAI's published policy distinguishes individual services from business products. Business, Enterprise and API inputs and outputs are not used for training by default. Individual-service data controls and voluntary sharing choices matter, and Codex full-environment sharing has separate controls. Check the actual account and current settings rather than inferring protection from the product name. OpenAI's data-use policy
Training, retention and access are different questions. A no-training commitment does not by itself answer how long files remain available, who can access them, or what a connected application receives. Retaining data is not, by itself, proof of training use either.
Follow the information, not just the prompt box
A review limited to the chat window can miss the surrounding workflow. Start with two or three high-value questions your team actually uses AI to answer. Then identify the information and services involved.
- Which tools and account tiers process the work?
- Which files, repositories or knowledge sources can connectors access?
- Where do prompts, outputs, logs and backups go?
- Who can share results or approve an exception?
- Which conclusions come from contracts, configuration records or observed tests?
- Which questions remain unanswered because evidence is unavailable?
A tool registry helps classify the services you encounter. It cannot, on its own, establish employee usage or prove what information was transmitted. Use available administrative records, procurement evidence and interviews, with appropriate authorization and privacy boundaries.
Private processing is a control choice, not a magic label
Running an open-weight model locally can remove a cloud model provider from that inference route. But the application around it may still use external connectors, remote support, telemetry, cloud backups or a fallback model.
A defensible private deployment specifies those routes, restricts unnecessary outbound access, controls administration, and tests the agreed boundaries using safe material. Exceptions are explicit. Logs and configuration evidence support the conclusion. No one should promise that owning a box eliminates every confidentiality risk.
For some workflows, approved cloud AI with suitable controls remains the practical choice. For others, private processing is justified by the information involved. The decision should follow the evidence and the work, not a blanket rule.
Turn uncertainty into an implementation plan
Aeon's AI Confidentiality Assessment starts with a bounded scope: one business unit, up to five priority tools and two sensitive workflows. Access, evidence, exclusions and a fixed fee are agreed before kickoff.
The outputs are a data-flow map, ranked findings, a vendor-terms summary and a practical correction plan. We distinguish what is verified from what is documented or unknown. We do not promise to reconstruct a provider's historical training pipeline or put invented dollar figures on exposure.
Agreed low-risk configuration changes can be included where authorized. Larger integrations, migrations and Aeon Private Stack deployments are separately scoped. The goal is to fix the workflow, not merely describe it.
Make the controls usable by people
Employees need safe handling practices. Managers need an exception process. Executives and boards need clear ownership, evidence and escalation, not another list of AI tools.
Applied workshops connect those responsibilities to real, sanitized scenarios. Our companion books support the sessions; they are optional, not a prerequisite.
If confidentiality is the defined problem, start with the AI Confidentiality Assessment. If the wider investment priority is unclear, the AI Control and ROI Assessment is an alternative starting point, not another compulsory step.
Use AI. Know the route. Control what happens next.