Healthcare & Disability /Customer-service agent live field guide · 9 min

Complaint Response Agent

Triages every patient complaint the moment it lands, flags the clinical-safety ones, and drafts a policy-grounded acknowledgement for staff to approve.

Open the live demo

theater/demos/healthcare_complaint-response-agent.html · sandbox · read-only

Open

FIG. 1

The live demo, running on fabricated data. Open it to step through the full flow — every output is shown for a person to approve before anything happens.

How it would work

Reads each complaint as it lands, classifies intent and urgency against the patient record, flags possible clinical-safety events, and drafts a policy-grounded acknowledgement for a person to approve before anything sends.

Input 01

The complaint + the record

An inbound patient complaint — email body or web-form submission — plus the patient record, recent call logs, the current medication list, your complaints policy and your routing map.

Agent 02

Classifies, verifies, flags, drafts

Classifies intent, urgency and sentiment from the text, verifies the patient by name and DOB, cross-references the record and medication list, raises a clinical-safety flag where a symptom matches a known reaction, routes by category, and drafts a cited acknowledgement.

Output 03

A draft, with its working shown

A draft acknowledgement with the policy clause it cited and the records it cross-referenced, plus an escalation banner on clinical-safety cases — staff approve, edit, reassign or escalate before anything is sent.

Where it works well

It reads every complaint the moment it lands, so the clinical-safety one stops waiting its turn in the queue.

The practice manager or complaints officer starts from a classified, verified, routed draft instead of a cold inbox.
Best where complaint volume makes triage a real cost and complaints arrive as readable text: email or web form.
At dozens of complaints a week, the recaptured triage and first-draft time goes back into handling the complaints that need real judgement.

The slow, invisible cost is the lag between a complaint arriving and a human realising it is actually a clinical-safety event. "I changed my medication and now I feel faint, I rang twice and no one called back" sits in the same inbox as a parking gripe and waits its turn.

Where it works badly

It is confidently incomplete when the signal isn't in the text it can read — and the draft looks authoritative either way.

Weak on clinically serious complaints written calmly — urgency partly rides on the words used, so flat polite language can under-trip the flag it deserves.
If your medication list lags the GP's notes, the clinical-safety flag is worse than no flag, because it looks authoritative while pointing at the wrong drug.

The honest test

Read your last twenty complaints: how many would classify correctly from their text alone, and how many needed a phone call or a record you'd have had to open anyway?

A complaint left as a voicemail with no transcript, or one whose substance is in a scanned letter attached as an image, gives it almost nothing. It classifies on the thin covering note and is confidently incomplete.

What it doesn't do — and shouldn't

It surfaces and drafts. A person decides and sends. That boundary is deliberate.

WHAT IT DOES

Classifies intent and urgency, and verifies the patient by DOB

Flags a possible adverse drug reaction and routes to clinical governance

Drafts an acknowledgement citing the policy clause it relied on

WHAT IT WON’T

Send anything — nothing is auto-sent

Decide whether an event is reportable or whether open disclosure is triggered

Decide whether an AHPRA notification is warranted, or give clinical advice

Whether an event is reportable, whether the Australian Open Disclosure Framework applies, and whether a notification to AHPRA or the relevant National Board (the Medical Board of Australia for doctors) or a state health complaints commissioner is warranted are human judgements with legal and ethical weight. The accountable person stays on the decision because the consequence lands on them — not the tool.

What your data has to look like

Complaints as readable text, a record it can verify against, and a complaints policy structured down to the clause.

Typical readiness

across orgs we see, before the first job

Complaints arrive as readable text

Usual weak point

The patient record supports identity verification

Usually ready

Medication list and call logs current to the day

Needs shaping

A complaints policy with clause-level structure

Needs shaping

A routing map by complaint category

Usual weak point

The real first job

Most practices have some of this and not all of it: the policy exists but isn't structured so a clause can be cited; the routing rules live in someone's head; the medication list is current in the clinical system but not reachable from the inbox. Getting that information captured and connected is usually the real first job — larger and more valuable than the AI layer that sits on top of it.

Right fit if…

Complaint volume is high enough that triage is a real, recurring cost

Complaints arrive as readable text — email body or web-form submission

You have a written complaints policy and a routing map it can ground against

Your medication list and call logs are current to the day in a reachable system

Walk away if…

Most complaints arrive as voicemail, phone notes or scanned-image attachments

Your medication list is batch-updated and lags the GP's notes

Your complaints policy isn't structured so a clause can be cited

You need a tool that decides reportability or signs off open disclosure for you

Open questions

The worried-buyer questions, answered straight

It is built to do the opposite of reassure when a message hints at clinical harm. A complaint that mentions a new symptom after a medication change — like a dry cough on perindopril, a known ACE-inhibitor reaction — is flagged High urgency, routed to clinical governance, and the draft is an acknowledgement only that books a clinician callback and gives explicit worsening-symptom safety-netting. It never offers clinical advice or tells a patient whether to keep taking a medication, and the draft is held for staff approval with the escalation flag the loudest thing on the screen.

It copes better with text it can read than with what it can’t. A forwarded email thread or a pasted web-form body classifies well; a voicemail with no transcript, or a complaint buried in a scanned PDF attachment, gives it little to work with and the classification confidence drops. When confidence is low it says so rather than guessing, and the item still lands in front of a person. It accelerates the clear cases so your staff have more time for the genuinely ambiguous ones.

No. It does the first read — classify, verify, route, draft — so the person who owns the complaint starts from a structured summary and a draft instead of a cold inbox. Under RACGP Standards (5th edition) Criterion QI1.2 the practice is responsible for analysing and responding to patient feedback and complaints, and that responsibility stays with the people accountable for it: whether this is an adverse event, whether open disclosure is triggered, whether it goes to AHPRA or a state health complaints commissioner. It recaptures the triage and first-draft time and points it at the complaints that need real handling.

Currency is the whole game for the clinical-safety flag. The medication list and recent call logs need to reflect today’s state, because the flag depends on what the patient is actually taking now. If your clinical system syncs in real time the cross-reference is trustworthy; if the medication list is updated in batches or lags the GP’s notes, the flag can be stale — and a stale flag in this setting is a real risk, so we’d scope data freshness before going near live complaints.

It runs against your own records and your own policy library, and the content is not used to train any external model. Patient identifiers and complaint text are sensitive health information under the Privacy Act and the Australian Privacy Principles, so where the data sits and who can see it is part of the build, not an afterthought. The demo here runs on fabricated data — Margaret Tully is not a real patient. We scope hosting and access with you before any real complaint touches it.

The person who approves and sends it — exactly as today. The tool drafts and cites; it does not send. Because every draft carries the policy clauses it relied on and the records it cross-referenced, the approver can see why it said what it said and correct it before it goes out, which is a better audit trail than a reply typed from scratch under time pressure.

What it takes to build

3–4 weeks · 4 phases

Reused from template~65%

Bespoke to this skin~35%

stack · Claude · private RAG · Outlook

What it would cost

Fixed scope, fixed price, fixed dates.

Bite-sized first piece

One contained change, low risk

Pilot build

Most builds land here

Embedded support

Scale on proof

Considering this for your practice?

The honest place to start is a bite-sized first piece — one contained change, low risk. Tell us where it hurts; we'll play it back, scope it, and show you what's possible.

Book a call How we work