VX / How we use AI
AI drafts. People decide.
Vector UX uses AI in two ways: to build software faster (Claude Code writes much of the code, tests, and first drafts; a person designs, reviews, and decides) and to power features on this site and in client software. On this site, the scoping assistant, the brief summary, and the audit readout are written by Claude from fixed tables and checks. Nothing you type is used to train models, and inputs are deleted after 30 days.
01On this site
What is automated, exactly
| Feature | What the model does | What it cannot do | Who checks |
|---|---|---|---|
| Scoping assistant | Reads your description and picks one lane and one scenario from our pricing table; writes a 40–60 word summary and three questions. | Invent a price. Every number comes from the table on the pricing page. | Jared, before any quote. |
| Brief summary | Restates your project brief as a scope card so you can see what we heard. | Commit us to anything. It is a draft. | Jared reads every brief and replies within one business day. |
| Audit readout | Turns 25 deterministic checks on your page into a plain-English summary and three fixes. | Run the checks. Those are code, not a model. | You. The full check list is shown under the readout. |
| Everything else | Nothing. Pages, prices, cases, and articles are written by people. | — | — |
02Your inputs
How we handle what you type
| Training | Never. Inputs go to the model provider only for the request itself, under terms that exclude training. |
|---|---|
| Retention | Deleted from our tools after 30 days. Briefs and estimates you asked us to email are kept as normal correspondence. |
| Sensitive data | Do not paste passwords, health records, or card numbers. We never ask for them. |
| Logs | Anonymized questions are reviewed to improve the scenario table. Never published. |
| Limits | Rate limits per visitor and a daily spend cap on the model account. |
03Models
What runs where
| Building | Claude Code for scaffolding, components, tests, and drafts. Every line is reviewed by a person. |
|---|---|
| This site | Claude Opus 5 for the three features above, with structured outputs so the answer is always a valid scope or readout. |
| Client software | Chosen per feature for accuracy and cost. We tell you what each feature runs on and publish the accuracy before launch. |
| Fallback | When the model is unavailable, the estimator and audit still work from rules, and say so. |
04Limits
What we do not let AI do
- Quote a price. A person does, after a call.
- Publish a case study or a number we cannot source.
- Send email on its own. Every email you get from us was triggered by something you asked for.
- Make a design decision for your project without you seeing it first.
- Run without a review step in any client feature that moves money, data, or a commitment.
Next step
Want an AI feature that works like this?
Constrained to real data, sources shown, a person at the end. From $8,000.