Human-in-the-Loop AI: Draft, Don't Send
Care schedules, support-session reports, sales outreach, training content: our AI systems propose and draft. A named person decides. Here's why that design wins.
Written byMuhammad Abdullah Chaudhary
Look across the AI systems we have shipped and one pattern keeps repeating: the model drafts, a person decides.
That is a design choice, not a limitation of the technology. This post explains where we use it, what it costs, and how to make the human step fast enough that people actually do it.
Where the pattern shows up
| System | What the AI does | Who decides |
|---|---|---|
| Care scheduling | Proposes shift assignments | Care coordinator |
| Disability-support sessions | Drafts report fields from the transcript | Practitioner |
| Sales research | Drafts personalised outreach | Sales rep |
| Learning content | Generates training material | Named reviewer |
| Lead capture | Fills contact fields from cards and conversations | Sales staff, before saving |
In care settings, a named person stands behind every record the system drafts.
Why not fully automate?
Because an automated record nobody checked is not an improvement. It is a faster way to be wrong at scale. Automated sending scales the wrong thing, and automated approval scales the error.
In regulated settings like care and healthcare there is a simpler reason too: someone has to be accountable for each record, and that someone has to be a person.
What the human step buys you
- Trust. Practitioners adopt tools that respect their judgement. Tools that overrule them get quietly avoided.
- Data quality. Every confirmation is a labelled example of what "right" looks like.
- Accountability. A named person stands behind each decision.
- A safe way to improve. Edit rates show exactly where the model is weak, without customers finding out first.
Making review fast
The human step only works if it is quick. If reviewing takes as long as writing, people stop reviewing.
- Structured drafts. A form to confirm, not prose to rewrite. In the support-session system, the practitioner checks fields, not paragraphs.
- Show the evidence. The transcript excerpt, the constraint that was satisfied, the source cited.
- One action to accept. Confirming a correct draft should be a single click.
- Edits are cheap. Changing one field should not mean starting over.
What to measure
| Metric | What it tells you |
|---|---|
| Acceptance rate | How often the draft is confirmed unchanged |
| Edit rate by field | Which parts the model gets wrong |
| Time to confirm | Whether review is actually fast |
| Rejection reasons | Gaps in data, rules or prompts |
The time saving is still real
In the disability-support project, paper double handling disappeared entirely: sessions are recorded, transcribed and drafted, and the practitioner confirms. The person stays in charge, and the admin burden still drops.
See it in practice in AI transcription for disability support sessions.
Sources and further reading
- People + AI Guidebook, Google PAIR
- Guidelines for human-AI interaction, Microsoft HAX Toolkit
- AI Risk Management Framework, NIST
- Collaborative intelligence: humans and AI are joining forces, Harvard Business Review
Related case study
AI Transcription and Report Drafting for Disability Support Sessions
