How we use AI
One feature in this product sends text to a language model run by a third party. This page names it, says exactly what leaves our server and what deliberately does not, and — where we have not yet done something a reviewer would like us to have done — says that too.
Where AI is used
Nothing else in the product calls a model. Every call below goes through one gate that checks consent, then the organisation’s monthly limit, then whether a provider is configured at all.
Rep suggestion
For administrators, when assigning a lead to a sales rep.
- Sent
- A lead’s skills, position and interests, and up to five reps as anonymous letters A–E with their skills, interests and current load.
- Never sent
- Every name, e-mail address and identifier on both sides. The letters are matched back to real people on our own server, after the answer comes back.
- Consent
- None is asked, because nothing identifying anyone is sent.
Who processes it
AI requests go to Anthropic’s API and its Claude models. Anthropic acts as a sub-processor for this processing and for nothing else. The operator of this instance holds the API key and chooses the model; with no key, every feature above is off and hides itself.
Model training
We do not train, fine-tune or evaluate any model on your data, and we never will without asking first. We are an ordinary API customer on the provider’s published commercial terms; we have not signed a separate data-processing agreement or a zero-retention addendum with them, and we would rather tell you that here than imply a guarantee we do not hold.
Saying no
The feature above sends nothing that identifies anyone, so it asks nobody for consent. Nothing else in the product depends on it — no core flow needs AI, and with AI switched off every feature you already had keeps working.
Switching it off for a whole organisation
An administrator can set the monthly AI limit to zero, which switches every feature above off for everyone in the organisation whether or not a provider key exists. A deployment with no key configured is off by default. In both cases the features hide themselves rather than showing a broken button.
A person decides, always
No model sets a pipeline stage, accepts or rejects anyone, makes an offer or changes an account. The rep suggestion is ranked first by a plain rule — shared skills, then who has room — and the model only re-orders that shortlist and adds one sentence of reasoning; an administrator reads it and chooses.
What is kept
Neither what is sent nor what comes back is stored. Generated text is shown in your browser and is gone when you reload. For each call we store a single metering row — which feature, when, which account — with no prompt, no output and no excerpt of either, so the monthly limit can be counted without keeping anyone’s data.
How to recognise generated text
Everything a model wrote carries a ✨ marker labelled “AI”, a description for screen readers and — where it is prose you might act on — a line reminding you to check it. If a piece of text in this product is not marked, a person or a plain rule produced it.
What we do not have yet
A page that lists only the good parts is worth nothing. As of today:
- We hold no SOC 2 or ISO 27001 certification for this application.
- We have signed no separate data-processing agreement and no zero-retention agreement with the AI provider; we use their published commercial terms.
- We cannot promise a processing region — requests are served wherever the provider serves them.
- We have published no evaluation measuring how accurate or how biased the rep suggestion is.
The technical version
The same facts written for engineers and kept next to the code — every task, every cap, every failure mode — are in the repository. If this page and that document ever disagree, the document is the one being checked against the code.
Read docs/ai.md on GitHub →