Pursor uses large language models (currently Claude, from Anthropic) to draft communications — emails, text replies, proposals, review responses, and follow-ups — in your voice and based on your business context. AI is an assistant; you remain in control of what is sent.
Human-in-the-loop is the default. Unless you explicitly enable auto-approval for a specific message type in Settings, every draft is shown to you before it is sent. You can approve, edit, or reject any draft.
Some features run with limited or no human review (for example, the auto-text sent within 60 seconds of a missed call). These are clearly identified in Settings, and you can disable them.
Known limitations include: outdated pricing or scheduling info if your business profile is stale; confident statements about facts that may not be true ("hallucinations"); tone misjudgments in edge cases; missing context the AI was not given. Pursor builds guardrails to reduce these but cannot eliminate them.
A growing number of state laws require businesses to disclose when communications are AI-generated. These include — without limitation — the California BOTS Act (Bus. & Prof. Code §§ 17940-17943), the Colorado AI Act (effective Feb 1, 2026), the Utah AI Policy Act, New York's bot disclosure law, and various proposed bills in other states.
You are responsible for determining which disclosure obligations apply to your business based on the states where your customers reside. Pursor provides the following to help:
Pursor's AI is not licensed to give legal, medical, financial, tax, mental-health, or other regulated professional advice. Several states (Illinois, Nevada, Utah and others) restrict the use of AI for therapeutic services or other regulated practice. If your business is in a regulated field, you must configure Pursor to avoid generating output that could be construed as professional advice or unlicensed practice, and you must always include appropriate disclaimers.
If your customer expresses distress or mentions self-harm, Pursor's safety guardrails will redirect to crisis resources (988 in the US) instead of producing an AI reply. Do not disable this guardrail.
Pursor does not use Your Content or your customers' personal data to train general-purpose AI models. Our contracts with AI providers prohibit them from training on your inputs and outputs.
We may use anonymized, aggregated, and de-identified data — such as latency metrics, error rates, and feature usage patterns — to improve the Service's reliability, prompts, and safety guardrails. This data does not identify you or your customers.
When you use Pursor's AI features you agree: