Responsible AI

Last updated: 7 August 2026.

SortedResearch's AI assistant, Graham, helps researchers build studies, check quality, name groups, and write insight reports. This page explains how we use AI, what data it touches, and the guardrails around it.

What powers Graham

Graham is powered by Anthropic's Claude models via Anthropic's API. We use smaller, faster models for quick tasks and more capable models for analysis and report writing. Model choices are centralized and reviewed, so upgrades are deliberate, not accidental.

What the AI sees, and what it doesn't

Graham processes your study data (cards, groupings, choices, screener text, aggregate results) and your conversations with him in order to help you.
Study participants are anonymous by default, so the data Graham analyzes typically contains no names or contact details.
Graham conversations are scoped to your account and organization. Tenant isolation (database row-level security) applies to AI features exactly as it does everywhere else.
No training on your data: under Anthropic's commercial API terms, inputs and outputs are not used to train Anthropic's models.
We log AI usage (which feature, token counts, cost) for billing, abuse prevention, and reliability. This logging records usage only, and is never a second copy of your content.

Guardrails

Human in the loop: Graham drafts, suggests, and summarizes; researchers review and decide. AI-generated reports are editable and clearly attributable to Graham.
Spending and abuse limits: per-plan daily caps and burst limits on AI usage, monitored on an internal dashboard.
Accuracy: large language models can make mistakes. Our Terms of Service describe the limits of AI-generated output; we design features so Graham's work is verifiable against your underlying data (for example, reports cite the actual numbers from your study).
Availability: if the AI provider is unavailable, core research features (collecting responses, viewing results, exporting data) keep working. Graham assists your research, and never gates it.

Data retention

Graham and support conversation data is retained for up to 24 months (see the Privacy Policy), and participant study data that Graham analyzes follows your instructions as controller under the DPA.

Questions

We are happy to walk security and procurement teams through our AI architecture: sales@sortedresearch.com.