A morning worklist from your service backlog

1What it looks like handed off

A request is open in yesterday's export, but a later update says it's finished. Another appears twice. A third has a draft ready and still needs someone to send it. Before sorting the backlog, you need to establish which requests actually need work.

The proposed list keeps one entry per known request ID, with the recorded state, latest supporting update and a next step for review. Unresolved records stay visible in a separate exceptions list. This is a workflow to try with exported copies, not a demonstration of a working agency-system connection.

The inbox guide helps read messages. This job reconciles exported work records across updates before you decide what the team should do. It doesn't draft another round of replies.

2The job, end to end

1. Save dated copies of the exports you're allowed to process. Record the system, export time and timezone. Start with fictional records or a small permitted sample.
2. Write down what each status means: open, waiting on staff, waiting on customer, draft ready and complete. Define which record establishes completion.
3. Ask for one proposed entry per request ID. Include each source filename, row or update ID, recorded owner, state and last update time.
4. Reconcile repeated IDs and later updates. Put missing IDs, conflicting states and uncertain ordering in the exceptions list; don't resolve them by guessing.
5. Account for every source row: included, repeated, recorded complete or unresolved. Review the omissions as well as the proposed work.
6. Check current records before acting. You set priority and assign the confirmed work in the system your staff already use.

A recurring run can prepare a new proposal from the next export. Keep each run's input versions and exceptions. Never let a second run create another assignment merely because the same request appears again.

For a known open thread that needs a nudge, use the follow-up guide after checking for a newer reply. If the export itself has format problems, clean a copy first and retain the original row references.

Try the fictional example

Download the files below. No signup is needed. All names, IDs, dates and documents are made up. The answer key shows the intended review decisions; it is not a measured AI result.

• Instructions and copyable prompt (TXT)
• Request export (CSV)
• Update export (CSV)
• Blank review sheet (CSV)
• Answer key (CSV)

Reconcile the two exports, then compare your proposal with the answer key. Check all nine input rows: the completed request and the duplicate still need an explanation; three entries must stay on hold.

Start with the instructions, give your existing AI tool the inputs and field rules, and keep the answer key aside until you have a proposal. Open the CSV files in your spreadsheet app. Leave approval fields blank until you have checked the sources.

3What stays with you

Priority

Age is one input. You decide which request needs attention first, using your service commitments and the current facts. A model's urgency label doesn't establish a deadline.

Assignment and closure

The agent writes a proposal in a separate file. Start without write access to the live work system. Staff assign, update and close requests.

Sending

A draft or proposed action is not a sent message. Check actual sending evidence before treating a response as delivered.

Uncertainty

Missing owner, unknown timezone or conflicting state becomes a visible question. It doesn't become an inferred fact.

Use only data you have permission to share with the AI provider, and remove personal details that aren't needed. For insurance work, licensed staff still handle policy facts, coverage advice and customer communication. This guide requires no EZLynx or Applied Epic login.

4When it gets something wrong

Finished work comes back

Compare the proposed state with the latest completion evidence. Correct the status rule and regenerate the proposal before assigning anything.

Two requests become one

Match on stable IDs, not a similar customer name or subject. If identity is uncertain, keep both source rows and ask.

An urgent item disappears

Reconcile the source rows against the output categories. A short worklist isn't evidence that the missing work is done.

Yesterday's export looks current

Show the export time at the top. If it is too old for the decision you're making, obtain a fresh copy or check the live record yourself.

A deadline is guessed

Separate recorded due dates from suggested priority. Leave the date unknown until someone establishes it.

Test the review with one fictional closed request, a duplicate, a conflict and an old export. Keep the expected decision beside each. If the proposal hides an exception, stop using it for assignments until the rule is corrected.

5What it costs

Try a small export with the AI access you already have. Model usage, upload limits and any export charges depend on your tools; check them before making the job recurring. You don't need a new computer or a system integration for the first trial.

Your main cost is reconciliation and review. Count the time spent exporting, checking exceptions and verifying current state. Compare that whole job with the open-work view you already have; don't count only how quickly the proposed list appeared.

6When not to bother

Use your system's existing open-work report if it already shows the owner, current state and due date reliably. A saved filter or a short staff check-in may solve the problem.

If the source records don't have reliable IDs or status updates, fix that process first. AI can't establish that work is finished from an empty cell.

For a handful of requests, sorting a sheet yourself is often enough. A recurring agent is worth considering only when reconciliation repeats and the checked proposal leaves you with less work overall.