← Blog
Security & Confidentiality

Paid AI Is Not Always Private AI.

Your science is your competitive advantage and it is always protected on Streamline. How AI model governance works at the API level, and why this matters more than ever in the age of generative AI.

NDA on every engagement API-level AI access only Explicit consent before corpus inclusion
Scroll
01 / Consumer AI vs. enterprise API

The distinction most researchers overlook

There is a meaningful and widely misunderstood gap between consumer AI products and enterprise API access. Both may use the same underlying model. Their data handling policies are not the same.

Consumer products (ChatGPT.com, Claude.ai, Gemini, Grok, others)

Consumer AI products are designed for general-purpose use and are typically governed by terms that allow the provider to use conversation data to improve the product. Opt-out mechanisms exist for some providers but are not universally available, not always retroactive, and not always applied consistently across all features. A researcher who uses a free or personal-tier consumer AI tool to draft, review, or brainstorm a grant application may be voluntarily submitting proprietary pre-publication research into a training pipeline. The provider bears no liability for this. The researcher and their institution do.

Enterprise API access (how Streamline operates)

Enterprise API access operates under different contractual terms. When an organization accesses Anthropic or OpenAI via the API under enterprise or developer agreements, those terms explicitly exclude input and output data from model training by default. The processing is inference-only: the model reads, responds, and the data is not retained or used further. This is the only appropriate mode for processing sensitive research content, proprietary technology descriptions, or competitive strategic information. It is the only mode Streamline uses. Both providers confirm this in their published API policies:

Anthropic (Claude API)

Anthropic’s API Terms and Privacy Policy state that inputs and outputs submitted via the API are not used to train Anthropic’s models unless you explicitly opt in. Enterprise API customers receive zero-retention data handling by default.

anthropic.com/legal/privacy →
OpenAI (API & Enterprise)

OpenAI’s API data usage policy confirms that data submitted via the API is not used to train or improve OpenAI models by default. This applies to all API users; consumer ChatGPT traffic is governed by a separate, different policy.

openai.com/enterprise-privacy →
Risk awareness

Many SBIR and STTR applicants routinely use consumer AI tools to draft grant narratives. If those drafts contain patentable claims, novel mechanisms, unreported clinical findings, or competitive intelligence, the applicant may be inadvertently disclosing that content under terms that afford limited protection. Before using any AI tool with proprietary scientific content, verify whether you are using a consumer product or an enterprise API integration, and confirm the provider’s data usage policy in writing.

02 / SOC 2 Type II certification

The highest standard for enterprise security

SOC 2 Type II is the gold standard for enterprise security compliance. It is defined by the American Institute of CPAs (AICPA) and requires an independent third-party auditor to verify that an organization’s security controls not only exist but operate effectively and continuously over an extended audit period, typically six to twelve months. It is not a questionnaire. It is not self-certified. It is a rigorous, evidence-based audit conducted by a licensed CPA firm.

Streamline is SOC 2 Type II certified. Achieving that certification required a substantial investment of time, infrastructure, and process discipline. We made that investment because the organizations we serve, including life science startups, university spinouts, and defense technology companies, operate in environments where security posture is not optional. Their funders, partners, and federal program officers expect enterprise-grade data handling. So do we.

What SOC 2 Type II actually covers

The audit evaluates controls across the Trust Service Criteria defined by the AICPA. Our certification covers the three criteria most directly relevant to client data:

  • Security. Systems are protected against unauthorized access, both physical and logical. This covers access controls, encryption in transit and at rest, vulnerability management, and incident response procedures.
  • Availability. Systems are available for operation as committed. This covers uptime monitoring, backup and recovery, and infrastructure redundancy.
  • Confidentiality. Information designated as confidential is protected in accordance with policy and commitment. This criterion maps directly to client IP, unpublished research, and grant content shared with us during engagements.
SOC 2 Type II is not a snapshot. The audit period spans months, verifying that security controls work consistently under real operating conditions, not just on the day of inspection. That distinction matters for clients whose data is with us across a multi-month grant development cycle.

The certification is renewed annually, meaning our controls are not a one-time project. They are maintained, tested, and re-verified every year by an independent auditor. Clients and their institutional compliance teams may request a copy of our SOC 2 report through our standard NDA-gated data governance process.

03 / Corpus governance

What enters our training data and under what terms

Streamline’s grant corpus is the foundation of our AI system. Understanding what is in it, and how it got there, is part of understanding our data commitments.

Source 1: Public record (FOIA)

The largest single source is a Freedom of Information Act request (case #60904, submitted 2023, fulfilled early 2026) that produced 5,910 SBIR/STTR grants and contracts awarded to companies across the country. This data was provided by the federal government under FOIA and is public record. It carries no confidentiality obligations.

Source 2: Streamline-authored grants

Grants written by Streamline’s principals for their own startup companies form a core part of the training baseline. This is first-party data. Streamline owns it outright and uses it under no external confidentiality constraint.

Source 3: Externally provided benchmarks

We have received grants from companies that submitted proposals we did not write and chose to share them as benchmarks. These are included only when they meet our quality threshold and only when the sharing party has confirmed their authority to share the document and grant us permission to use it for training purposes.

04 / Confidentiality by construction

How we operationalize IP protection across every engagement

Confidentiality is not a policy we added when AI became prominent. It has been an operational requirement of this business since the first client engagement. The AI era has made it more important to articulate clearly, not more important to start practicing.

  • NDA before access. Every client engagement begins with an executed mutual non-disclosure agreement before any scientific content is shared. This is not optional and is not waived for any client tier or engagement type.
  • No cross-client contamination. Information shared by one client is never surfaced, referenced, or used in the work of a competing client. Our platform and our human team operate with strict engagement isolation.
  • Pre-publication data is treated as unpublished. Preliminary data, novel efficacy signals, unreported clinical findings, and patentable claims shared with us during grant development are treated with the same confidentiality as attorney-client privileged communications. We do not discuss, reference, or extract insights from one client’s unpublished science for any other purpose.
  • Team access is scoped. Within Streamline, access to a client’s materials is limited to the team directly serving that engagement. Content is not circulated broadly, stored in shared repositories with open access, or referenced in internal knowledge bases in identifiable form.
  • No public use of client work product. We do not publish, excerpt, reference, or use client-developed grant content in marketing materials, case studies, or public-facing communications without explicit written approval. Win announcements and client references require client permission.
05 / Sensitive and regulated content

SBIR/STTR-specific considerations for high-sensitivity proposals

SBIR and STTR proposals occupy a particular sensitivity tier that most generic AI guidance does not address. They frequently contain elements that carry legal, regulatory, or competitive consequences if improperly disclosed.

Controlled Unclassified Information (CUI)

DoD SBIR solicitations sometimes involve topics designated as Controlled Unclassified Information under the NIST SP 800-171 framework. When an applicant is responding to a CUI-relevant solicitation, the proposal itself, and the technical approach it describes, may require handling under specific controlled-access protocols. Streamline does not process proposals containing marked CUI through any external AI system without first confirming that the processing environment meets the applicable handling requirements. When working on DoD engagements that may involve CUI, we discuss handling requirements explicitly with the applicant before beginning work.

Patent-sensitive content

Disclosing a novel invention publicly before filing a patent application can constitute prior art that invalidates a future patent in some jurisdictions. An SBIR proposal is not a public disclosure in the legal sense, but the handling of that proposal is. If your grant describes patentable technology that has not yet been filed, tell us upfront. We will ensure that our handling protocols for that engagement are consistent with your IP strategy and, if relevant, coordinate with your IP counsel on appropriate information handling practices.

Pre-publication clinical and scientific data

Many SBIR applications describe clinical findings, biomarker results, or efficacy signals from ongoing or unpublished studies. This data is frequently embargoed pending publication or journal review. We treat all preliminary data shared in the context of grant development as embargoed by default, regardless of whether it is explicitly labeled as such. It is not extracted, aggregated, or referenced beyond the immediate engagement.

Export control and ITAR/EAR considerations

Some technologies relevant to SBIR/STTR applications, particularly in the defense, dual-use, and advanced materials spaces, may be subject to export control regulations under ITAR (International Traffic in Arms Regulations) or EAR (Export Administration Regulations). If your technology falls under these frameworks, we recommend discussing your export control classification with your compliance counsel before sharing detailed technical content with any external party, including Streamline. We can structure our engagement to minimize technical depth until appropriate controls are confirmed.

06 / The new risk landscape

Why the AI era changes the calculus for researchers

For most of the history of grant writing, confidentiality risk was concentrated in a narrow set of scenarios: a careless email, a shared folder with wrong permissions, a conversation at a conference. The threat surface was manageable because the number of parties who could see your content was small and largely known.

Generative AI changes that. The threat surface is now invisible and structural. Researchers and teams routinely use AI-powered writing and editing tools without considering where the data is going, how long it is retained, and what the provider’s terms permit them to do with it. The risk is not hypothetical.

What can happen when consumer AI tools process your grant

Consumer AI products process your input to generate a response. They may also log it, analyze it, use it to fine-tune future model behavior, or make it accessible to safety and trust reviewers inside the company. The extent to which any of this occurs varies by provider, product tier, and current policy, and policies change. A researcher who used a consumer AI tool to draft a Specific Aims page two years ago may not be able to determine today whether that content was used in any training process, or what, if any, retention policies applied at the time.

The most protective stance is simple: do not paste proprietary scientific content, unreported clinical data, or novel mechanism descriptions into any consumer AI product. If you want AI-assisted grant writing, use a platform that can confirm it operates under enterprise API terms with zero training use and contractual confidentiality protections. Verify this in writing before sharing anything sensitive.

Institutional responsibility is evolving

Research institutions, technology transfer offices, and sponsored programs administrators are increasingly aware of these risks and are beginning to issue guidance on AI tool use for federally funded research. Some NIH funding mechanisms now include language around data management and AI disclosure. DoD solicitations are likely to become more explicit over time. Getting ahead of institutional policy by establishing principled AI use practices now is considerably easier than retrofitting them after an incident.

The researchers most at risk are not the ones being careless. They are the ones who have no reason to believe a tool they use every day is operating under anything other than the terms they assumed. The responsibility to be explicit about those terms sits with the platforms, not the researchers. We take it seriously.

If you have specific questions about how Streamline handles your materials, our data agreements, or the AI tools we use in your engagement, we welcome the conversation before work begins. Reach us through the engagement intake form and specify that you have data handling questions. We will connect you directly with the appropriate team member.

Your science is your competitive advantage. Protecting it is our baseline commitment, not an upgrade.