Our principles
- People stay accountable. AI handles routine administrative work. Clinical decisions, and the approvals an organization asks for, stay with people.
- Rules outside the model. What an AI may do is set by deterministic rules the organization controls, not by the model itself.
- Say when it's AI. People are told when they are talking to an automated assistant.
- Data is for care, not for training. We do not train AI models on customer data.
- Test before trust. Agents are tested against scripted conversations before we rely on them.
Human oversight
In ByteOps, every action an agent can take has an approval rule. The organization chooses one of three for each action:
- Automatic: the agent does it.
- Patient confirms: the patient must agree first.
- Staff approve: a member of staff decides.
The rules are built so that mistakes fall on the safe side:
- When in doubt, staff decide. If something is uncertain, such as an unconfirmed patient, missing consent, or a possible duplicate record, the action goes to staff whatever the rule says.
- Money goes to people. Claims and payment work goes to staff.
- Agents can't change the rules. Only an organization's owners and admins can, and every change is recorded.
- Every decision is explained. Each decision records the rule that applied and why.
- Staff can step in. Staff can take over any conversation from the assistant.
What the AI does
- Answers patients' messages and helps them book, change, and keep visits.
- Registers new patients, within the organization's rules.
- Prepares insurance work: coverage checks, authorization requests, claim batches, and follow-ups, for staff to approve.
- Interviews a new organization about how its services run, and records what it learns for the organization to review.
- Helps ByteCare users find a clinic and send it a request.
- Turns speech into text when someone uses voice input.
What the AI doesn't do
- It does not diagnose, prescribe, or give medical advice.
- It does not make clinical decisions.
- It does not reveal a person's details before that person has given them.
- It does not change approval rules or its own permissions.
- It does not replace emergency services.
Telling people it is AI
Every patient conversation in ByteOps opens with the clinic's notice, which by default says "I'm an automated assistant". Where the clinic asks for consent, the assistant waits until the patient agrees. ByteCare labels who each message is from. When ByteOps contacts an insurer on a clinic's behalf, it says it is an automated assistant.
Data and models
- No training on customer data. We do not use customer or patient data to train or fine-tune models.
- Third-party providers only where allowed. Some models run on infrastructure we run or contract. Others are third-party providers, used only where an organization allows them. All are listed on our subprocessors page.
- Known limits. Each model lists its limitations, such as the languages it handles well, and language coverage differs between models.
- Voice is not kept. Recordings are turned into text and never stored.
Emergencies
ByteCare does not rely on AI alone to notice an emergency: fixed keyword checks run alongside the model. When an emergency is raised, nearby clinics are alerted, and the person is always told to call 112.
Evaluation
Each agent is tested against simulated conversations before we rely on it. The tests check that it takes the right actions in the right order, never takes actions it must not, doesn't reveal a person's details before they have given them, and records the right things in the health record. Scheduling rules are checked by separate fixed tests that don't use AI.
ByteCare and AI
- ByteCare's assistant is an assistant, not a clinician. It helps you find care, send requests, and follow them. It does not give medical advice.
- Your messages to the assistant are processed by AI models. The providers are listed on our subprocessors page. We do not use your messages to train models.
- A clinic's staff can step in to a conversation, and the clinic, not the assistant, decides on your care.
- You can ask for a copy of your conversations, or for them to be deleted: [email protected].
Report a problem
If an assistant gave a wrong, unsafe, or unfair answer, tell us at [email protected] with the time, the product, and what happened. Please don't include more health information than you need to. We review every report and use it to improve our tests.