A draft is not a send: the approval queue behind Lisa's agents

An AI agent can draft a post-visit guidance message in seconds. The question that matters is different: who decides that it reaches the patient?

In the architecture of Lisa's agents, the answer is always a person. Every message an agent produces - a care-instruction summary, a follow-up reminder, a recovery check-in - starts as a draft. It sits in an approval queue until the responsible clinician or team reviews the content and authorizes the send. A draft is not a send. The two states exist separately on purpose.

It may look like an operational detail. It is a design decision.

The World Health Organization's guidance on AI for health places human autonomy at the center: AI systems must not displace people from the centre of health decision-making, and their operation requires mechanisms for oversight and continuous assessment. The approval queue is how that principle shows up in a platform's daily work: the machine speeds up production, but the final word - the word that reaches the patient - remains human.

In practice, this changes three things.

The first is accountability. When every message has an identified reviewer, there is no "the agent sent it." There is a professional who read it, adjusted it when needed, and owns what was released. In healthcare, this chain of authorship is not bureaucracy; it is what makes auditing, correction and learning possible.

The second is the quality of the draft. A good agent does not remove human work - it changes its nature. Instead of writing every instruction from scratch, the team reviews, refines and approves. The time saved on writing returns as attention to the cases that genuinely need clinical judgment.

The third is the right metric. A healthy approval queue is not measured by how many messages are generated, but by what happens between draft and patient: review time, edit rate, rejection reasons. Those numbers show whether the agent is learning to serve the team - or merely producing volume for the team to carry.

Some see human approval as a brake that technology should eventually remove. We see the opposite. In a care journey, trust is the most expensive asset, and it is built precisely at the points where someone answers for what is said. The agent writes fast so the human can decide well.