AI and model providers
Harlow uses hosted AI services for two things: turning speech into text, and understanding what a sentence asks for. Here is which service does what, and the commitments around it.
| Purpose | Provider | Terms |
|---|---|---|
| Transcription | Azure AI Speech, within Harlow's Azure environment | Not used for training; inputs not retained beyond providing the service. |
| Understanding a request | Azure OpenAI, within Harlow's Azure environment | Same: no training, no retention beyond the request. |
| Screen automation (specific features) | OpenRouter | Enabled for a customer only under a data processing agreement with zero-retention terms. |
| The Overview's observations | A model over counts and timings only | No names, no transcripts, no guest content. |
Commitments
- We send a model the minimum content needed for the request in front of it.
- We do not train shared, generalised models on your guest data for the benefit of other customers.
- Where we improve prompts or models, we use aggregated, sampled or de-identified data that the contract permits.
- A new provider is introduced only under equivalent no-training terms, and appears on the sub-processor list first.
Automated decisions
Harlow routes and drafts; people decide. A request with consequences in a hotel system waits for a person where the hotel has asked for that, and every automated step is visible on the request's record with a way to undo or correct it.
The binding text is section 5 of the privacy policy.