How ShiftPilot uses AI.
Two kinds of automation, kept apart on purpose. A deterministic solver builds the schedule under your rules. An assistant answers questions about it and asks before it changes anything.
The schedule is solved, not guessed
Generation uses a CP-SAT constraint solver. It satisfies every hard rule or names the one it cannot, and lists the soft preferences it traded. The result is reproducible and explainable.
The assistant reads, then asks
Avery answers from the live schedule on the web, WhatsApp, SMS and the mobile app. Any change it proposes, to a shift, an availability record or a setting, takes effect only after a person confirms it.
Only what is needed leaves
When someone talks to Avery, the message and the scheduling information needed to answer it are sent to a language-model provider through an API gateway. Nothing else is shared, and nothing is used for training.
People approve every change
Swaps are checked against the rules, availability sent by text is parsed onto the calendar where the sender can see it, and reminders never change a schedule. Every change traces back to a person, not a model.
Open by choice
Administrators can connect Claude or ChatGPT to ShiftPilot through MCP, the Model Context Protocol, with the same permissions and approval steps as Avery. It is off until a group turns it on.
Written into the policies
The privacy policy names the model provider as a subprocessor and sets out the no-training, no-retention terms. The terms of service cover approval and configuration rights.
Questions we get asked
The detail lives in the privacy policy and terms. These are the short versions.
Is the schedule generated by AI?
No. Schedules are built by a CP-SAT constraint solver, which is deterministic and rule-driven: the same rules, availability and credentials always produce a schedule that satisfies the same constraints. No language model decides who works when.
Where does Avery use a language model?
Avery uses a large language model to understand questions and requests in plain language and to phrase its answers. What it knows comes from the live schedule and the group's own configuration. When asked to change something it runs the rules and asks for approval first.
Is our data used to train models?
No. The model provider is contractually prohibited from using message content to train or improve its models and from retaining it beyond the time needed to generate a response. ShiftPilot does not use customer data to train models either.
Can we turn the assistant off?
Yes. Administrators can pause the assistant for any person, and any member can stop its messages with a single word. Nothing about the schedule depends on the assistant being on.
Your next month, already scheduled.
See ShiftPilot on your own rules and roster. Thirty minutes, no slides.