AI Use Policy
How we use AI in PR and content production: what it touches, what it never touches, who reviews it, and how we handle mistakes.
Last reviewed: July 1, 2026
1. Data handling
We use AI to assist with research, outlining, drafting, and editorial review. Client data is handled with a clear line between what AI tools may process and what they never see.
What AI may process
- Public information about the client (company name, public filings, press history, public bios).
- Draft copy the client has already approved or requested feedback on.
- Research queries about the client's industry, competitors, and topic area.
What AI never processes
- Non-public financials, unreleased deal terms, or unannounced product information.
- Personally identifying information about executives beyond what is already public.
- Anything the client has marked confidential in writing.
- Material non-public information for securities-regulated clients (see section 4).
Client documents live in our own backend infrastructure. They are never pasted into consumer AI interfaces (free ChatGPT, Claude.ai, Gemini consumer).
2. Training opt-out
Every AI vendor we use is accessed through their enterprise or API tier with a written no-training-on-inputs commitment. Inputs we send (drafts, research questions, summaries) are not used to train future models.
We review vendor data policies on a recurring schedule. If a sub-processor materially changes its terms, we re-evaluate the relationship and notify affected clients.
3. Human review and publication approval
AI may assist with research and writing, but a generated draft is not a verified article. Our publication standard requires two editorial passes: a whole-draft review during preparation and a separate final recheck of the saved version before publication.
Both passes cover the article body, FAQs, takeaways, tables, captions and examples, together with titles, search descriptions, image descriptions and structured data supplied to search engines. After a repair, the changed claims and any repetitions elsewhere in the article must be checked again.
Factual claims must be supported by sources that actually substantiate them. Each statistic must be checked against its named, dated primary source, including the population, period, denominator and context. A working link does not prove a claim is accurate. Unsupported statistics are removed. Review also challenges absolute legal claims, misleading comparisons and promises of outcomes.
Unresolved factual issues or missing verification prevent an article being marked publication-ready. Evidence limits are recorded and reported; uncertain claims must be removed or accurately qualified without disguising a sourcing gap. Automated checks can assist this work but cannot certify accuracy or replace editorial judgment.
Publication requires explicit human approval. Saving a draft, passing automated checks or approving its tone is not a substitute for the final factual review. Errors can still occur; when they do, we investigate and correct them rather than claiming our checks are infallible.
Reader value and originality
We use AI to support research, outlining and drafting—not to mass-produce pages primarily to manipulate search rankings. Each article must answer a distinct reader need and add useful analysis, practical guidance, clearly labelled illustrative examples or verified original information. We do not publish lightly rewritten source material or overlapping articles simply to target different keyword phrases.
We do not add filler to meet a word-count target or publish an incomplete article to meet a posting schedule. Article length follows the topic and the evidence needed to answer it. A shorter, complete article may be considered for an explicitly approved exception to our internal depth requirements; those requirements are not a Google ranking rule.
We do not invent first-hand experience, expert credentials, quotations, research findings or client results. Illustrative examples are identified as illustrative and are not presented as evidence of real outcomes.
Authorship and explaining AI assistance
Authorship and reviewer information must accurately describe responsibility for the content. We do not use invented expert identities or imply that someone personally wrote, tested or experienced something they only reviewed.
When AI substantially assists an article and readers would reasonably expect an explanation, we provide a clear account of how it was used and who is responsible for editorial review. Any review statement must reflect work actually completed. Our AI Use Policy explains the broader process; it is not a certification that every statement across the website has been verified.
4. Securities & regulated-client carve-out
Securities-regulated clients, including Reg A+ issuers and any client subject to SEC, FINRA, or FTC oversight, operate under a documented carve-out from our general AI workflow.
- Material non-public information never enters a general-purpose AI tool.
- Drafts pass through a human editor before any external transmission.
- The workflow is isolated from our general content pipeline, with a documented chain of custody.
- All copy is reviewed against the client's securities counsel guidance prior to publication.
See our Reg A+ Issuer PR page for the regulated-client engagement structure.
5. AI sub-processors
The following AI vendors may process client-related inputs as part of our workflow. Each is on an API or enterprise tier with a no-training-on-inputs commitment.
| Vendor | Purpose | Training |
|---|---|---|
| Lovable AI Gateway | Platform-level AI access layer | No training on inputs |
| OpenAI | Drafting, research, reasoning | No training on inputs |
| Anthropic (Claude) | Long-form drafting, editing, reasoning | No training on inputs |
| Google AI (Gemini) | Research, reasoning, multimodal review | No training on inputs |
| Perplexity | Sourced research and citation discovery | No training on inputs |
Research, diagnostic, and infrastructure tools that only process public web data (search APIs, crawlers, analytics, hosting) are not listed here as sub-processors, because no client-identifying inputs are sent to them.
6. Incident & correction protocol
When AI-assisted output contains a factual error that reaches publication, we follow the same correction process as any other editorial mistake:
- The error is corrected as soon as it is verified.
- Material corrections are logged on our public Corrections page with a timestamp and explanation.
- An internal review identifies the workflow gap that allowed the error and updates our review checklist.
- Affected clients are notified before any public correction is issued on content carrying their name.
7. Client opt-out (zero-AI workflow)
Any client may request a zero-AI workflow on their account. Research, drafting, and fact-checking are handled by humans only. The option is added as a written addendum to the engagement agreement.
Expect longer turnaround times and higher hourly cost; human-only production takes more hours, but the option is available without friction.
Questions about our AI use
For questions about how AI is used on your account, or to request a zero-AI workflow addendum, contact the editorial team.