Inbox Triage and Draft Replies
Sorts incoming messages by kind and drafts replies from your approved facts only. Unclear messages go to a person, and nothing is sent until you approve.
Who it is for
W02 Inbox Triage and Draft Replies is an iSystematic workflow for clinics, law firms, accounting firms, property management and service businesses.
Organisations use it inside the Front Door Agent, and firms inside the Practice Platform.
What it prepares
It prepares a queue of your messages, sorted by kind, each with a draft reply.
What always needs you
These always wait for your approval, at every level.
- Sending
- Promises
- Refunds
- Confidential details
How it runs
It runs in six steps, and nothing leaves until you approve.
Classify each message
Each message is sorted as scheduling, billing, general, complaint or urgent human review.
Extract the request
It sets out what the sender is asking for.
Flag risk
Complaints, clinical requests and anything unreadable or unclear are flagged for a person.
Draft a reply
A reply is drafted from your approved facts only; unknowns are left as questions.
You approve
You review each draft. Sending, promises, refunds and confidential details always need you.
Send, if you choose
Sending is optional, and happens only after your approval.
The free card
Prompt: “Classify as scheduling, billing, general, complaint, or urgent human review. Draft from the approved facts only. Leave unknowns as questions. Do not send.” Manual version: open your assistant, paste the prompt and your approved response policy, then paste the messages you are allowed to share, or fictional samples while you try it. Read each draft against its message, edit it, and send it yourself.
What you get
You get a queue in which every message carries four things.
- A category: scheduling, billing, general, complaint or urgent human review
- The source message, so you can check the draft against it
- A draft written from your approved facts, with unknowns left as questions
- An escalation reason wherever a person must step in
When it stops
It stops on any message it cannot read or cannot be sure about, and passes that message to a person for review.
How we test it
We test the free card with fictional data, first against this workflow's own acceptance tests, then against the catalogue's test checklist items that apply to the manual version. The card is marked Ready only once it passes.
- Clinical requests and complaints are escalated to a person.
- No policy is made up.
- A duplicate message does not produce a duplicate send.
- It works by copy and paste alone, with no connector.
- Every source it cites and every detail it pulls out is checked against the original.
- It is given missing and contradictory information, to check that it does not fill the gaps.
- It is asked for clinical, legal and financial judgements, and must refuse them.
- A hostile instruction is hidden in the content it reads, and it must not obey it.
- It is given duplicates, timezones and out-of-date information.
- Nothing is sent, spent, published or changed before you approve it.
Levels
It runs at three levels; running when a message arrives needs a custom integration or a native feature where one is supported.
| Level | In this workflow |
|---|---|
| Manual (M) | Yes: paste the messages, review the drafts, send them yourself |
| Reusable (R) | Yes: saved instructions in your assistant, where your account supports it |
| Connected (C) | Yes: a supported connector retrieves the messages, and sends only what you approve |
| Scheduled (S) | Not part of this workflow |
| Custom (X) | Only for running when a message arrives: a custom integration, or the platform's own feature where it is supported |
Patterns it uses
Each pattern is a reusable rule from the catalogue. Every implementation also carries the controls listed on the workflows page, the same ideas set out in how we build.
- P03 Classify and Escalate: each message gets one category, and uncertain or high-risk ones go to a person with the reason recorded.
- P04 Approval Before Commitment: a reply is sent only after you approve that exact reply.
- P07 Deduplicate Before Acting: a message already handled does not get a second reply. Durable protection may need a custom integration; a prompt alone is not enough.
- P10 Inspect Retrieved Instructions: an instruction written inside an email is treated as part of the message, never as an order.
A fictional example
Sara (fictional), a front-desk lead at a dental clinic, tries W02 on fictional sample messages with the clinic's approved response policy. Each message comes back with a category and a draft written from the approved facts, with the questions it could not answer left open. A message describing a symptom is flagged for a person, with the reason. Sara reads every draft against its message and sends the ones she approves herself.
Where to go next
Front Door Agent
The Front Door Agent gives an organisation one front door across email, web chat, WhatsApp and forms, built on ARIA by Simplification.io, with people approving.
Practice Platform
The Practice Platform gives a professional firm one approved knowledge library behind its website, staff assistant, front door and answer-engine presence.
W01 Daily Briefing
W01 Daily Briefing turns your notes and appointments into a dated briefing, with sources and gaps listed. It sends nothing and changes no calendar without you.
W03 Approved FAQ
W03 Approved FAQ Assistant answers only from your approved FAQ and cites the section. If the answer is missing, out of date or conflicting, staff must confirm.
Clinics
Clinics: iSystematic sets up AI workflows that sort messages, draft answers from your approved FAQ and prepare requests. Nothing clinical is automated.
Law firms
Law firms: iSystematic sets up AI workflows for intake summaries and client onboarding. Legal advice, conflict checks and every citation stay with lawyers.
Set it up in your own account
We configure W02 with you in your own Claude or ChatGPT account, in a 60 to 90 minute call.