Here's a scenario worth sitting with: it's 6:15 PM, and a family physician has 40 AI-drafted replies to patient portal messages waiting for approval. She reads the first ten carefully. By number thirty she's approving after a glance, because every draft so far has been fine. Number thirty-four reassures a patient about a mild headache and passes over the line where the patient mentions it started after a fall.
The draft was fluent and mostly right, which is exactly the problem. She was in the loop on paper, but the loop had been built so that her attention ran out before the messages did. Putting a person in the loop is easy. Designing a loop that keeps that person useful is the real work.
What “human in the loop” means in practice
The phrase covers a wide range. At one end, a person approves every single output before it goes anywhere. At the other, the system acts on its own and a person reviews a sample afterward. Most good setups mix the two: tight control where mistakes are expensive or hard to undo, lighter checks where they're cheap and quickly visible.
So the question for each task isn't “should a person be involved?” It's “at which point, looking at what, and deciding what?”
Five patterns, and when each one fits
1. Draft, then approve
The system prepares and a person reviews and sends. Use it for anything a client or patient will read: engagement letters, portal replies, return-status emails. It's the safest pattern and the slowest, and it works far better when the reviewer sees each draft beside its source rather than on its own.
2. Confidence thresholds
The system acts alone only when it's confident and the case matches rules you've set; anything else goes to a person. A bookkeeping system might post a transaction automatically when the vendor and amount match the last three months, and hold it otherwise. Set the threshold from your own measured trial, not a vendor's default.
3. Exception queues
Items the system can't or shouldn't handle land in one queue with the reason attached: “document missing,” “amount over limit,” “patient mentions a new symptom.” The queue needs an owner and a time by which it's cleared, or it turns into a second inbox nobody reads.
4. Spot checks
Once a task runs on its own, a person reviews a sample, say ten random items a week plus anything a customer complained about. Spot checks catch drift: the slow decline that sets in when your prices, forms or clients change and the system doesn't.
5. Audit trails
Every action is logged: what the system saw, what it did, who approved it and when. The trail lets you answer “why did this happen?” a month later, and it makes supervision visible to a client, a regulator or a managing partner.
| Task | Pattern | Why |
|---|---|---|
| Engagement letter drafted from intake notes | Draft, then approve | A lawyer's name goes on it |
| Bank transaction categorization | Confidence threshold plus spot checks | High volume, and errors surface at review |
| Patient portal messages | Exception queue plus draft, then approve | Anything clinical has to reach a clinician |
| Appointment reminders | Spot checks | Cheap, visible and easily corrected |
| Every automated action | Audit trail | Someone will need to explain it later |
Flowchart
One client email through the loop
- StartClient writes: “Did you get my 1099s? When will my return be done?”
- StepSystem reads the email and checks the engagement's document list and status
- DecisionIs it a factual question the file answers clearly?NoA personPreparer gets the email with the file status attached and repliesYes
- StepSystem drafts a reply: two documents received, one missing, upload link
- DecisionIs this reply type on the firm's approved auto-send list?NoA personStaff member reviews the draft beside the file and sends itYes
- StepReply sent and logged with the records it relied on
- DoneOffice manager spot-checks ten sent replies every Friday
Before and after
Rubber-stamp review vs. designed review
Rubber-stamp review
- 01Forty drafts arrive in one undivided list
- 02Each draft shown without the patient's message beside it
- 03Routine and clinical messages look identical
- 04Approval is one click, so fatigue sets in fast
ResultA person in the loop who can no longer catch errors
Designed review
- 01Symptom messages go to the triage nurse's queue first
- 02Each draft shown beside the patient's own words, with flags
- 03Refills and scheduling batched separately
- 04Anything mentioning a new symptom can't be bulk-approved
ResultLess to read, and attention where it matters
When to loosen review, and when never to
Review should get lighter as evidence builds. A reasonable test for moving a task from approving every item to spot checks: it has run for several weeks at normal volume, reviewers' corrections have become rare and minor, a mistake would be cheap and quickly noticed, and the customer can easily reach a person. Appointment confirmations often pass that test. Client-facing advice almost never does.
Some steps should keep a person however good the system gets:
- Signatures and filings. Tax returns, court filings, contracts and anything signed in a professional's name.
- Clinical decisions. Diagnosis, treatment, and any reply to a new or worsening symptom belong to a clinician.
- Legal advice. Whether a person has a claim worth pursuing, and what to do next, is a lawyer's call.
- Moving money. Payments, refunds and write-offs above a threshold you set.
- Saying no. Declining a client, a claim or a patient's request deserves a person's judgment and a person's voice.
What the rules say about keeping a person involved
Professional rules increasingly assume a human reviewer. The ABA's July 2024 Formal Opinion 512 applies lawyers' existing competence and supervision duties to generative AI; in practice, a lawyer checks AI-assisted work before relying on it. Mata v. Avianca shows the cost of skipping that step: in 2023, a federal court in New York sanctioned lawyers who filed a brief citing non-existent cases generated by ChatGPT.
In health care, California's AB 3030, in effect since January 1, 2025, covers health facilities, clinics and physicians' offices, solo or group, that use generative AI to produce written or spoken patient communications about clinical information. Those messages need an AI disclaimer and instructions for reaching a person, unless a licensed or certified health care provider has read and reviewed them. Scheduling, billing and other administrative messages aren't covered.
Human review, in other words, changes what the law asks of the message. Whether a glance at draft thirty-four counts as “read and reviewed” is a question for your own counsel, but a designed review, with the patient's words beside each draft, is far easier to stand behind. Consider this general background rather than legal advice; requirements vary by state and by profession.
Setting up your own loop: a checklist
- 01List every step the system will take, and mark the ones a client, patient or court will see.
- 02Choose a pattern for each: approve every item, threshold, exception queue or spot check.
- 03Name the reviewer for every queue and the time by which it's cleared each day.
- 04Show reviewers the source beside the output, and highlight anything unusual.
- 05Block bulk approval for high-risk categories, defined in writing: for example, any message that mentions a new or worsening symptom, any amount over your money threshold, and anything addressed to a court or agency.
- 06Log every action, approval and correction.
- 07Set a date to decide whether each step's review should tighten or loosen.
If you're designing this for the first time, deciding which step gets which pattern is part of how we plan a build. And our piece on why AI pilots stall covers the weekly review that keeps a loop honest once it's running.
Sources
- American Bar Association: Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024)
- U.S. District Court, S.D.N.Y.: Mata v. Avianca, Inc., Opinion and Order on Sanctions (June 22, 2023)
- California Legislative Information: Health and Safety Code section 1339.75 (added by AB 3030, Chapter 848, Statutes of 2024)
- California Department of Public Health: AFL 25-07, Assembly Bill 3030, Health Care Services: Artificial Intelligence