| You get | A proposed intake record with source references and unresolved fields |
|---|---|
| Stays with you | The required fields, customer follow-up and approval to accept the work |
| You need | A small set of fictional or approved requests and your own intake requirements |
| Rough cost | Start with manual review and existing AI access; count review time and any additional tool charges |
| Skip it if | Your current form already captures the needed details and staff can check them directly |
1What it looks like handed off
You have one proposed record for each request. Each important value points to its source. Missing information stays blank, and competing values stay visible. A complete set of fields means ready for staff review, not accepted work or a confirmed booking.
This is the step before a backlog becomes a morning list. The worklist guide starts from existing request records and updates; this exercise starts with the incoming messages and checks whether a usable record can be prepared.
2The job, end to end
- Choose the fields your staff actually need before deciding what happens next. For this exercise they are service, quantity, requested date and reply address.
- Retain each source message and its reference. The practice pack supplies request IDs; it does not test matching unknown people or accounts.
- Ask AI to propose one record per request, with a source reference for each field. Require blanks for missing facts and both values for a conflict.
- Compare the proposed record with every supplied message. A later-looking row is not evidence that one value supersedes another.
- Have staff resolve the missing information or conflict using the approved process. Keep that decision separate from the AI draft.
- Approve the checked record before it enters live work. Confirming a booking, sending a reply or changing another system needs its own authority and evidence.
Try three fictional requests
The four source messages describe three made-up menu-printing requests. All names, addresses, dates and quantities are fictional. The two messages for I-03 disagree on quantity; neither establishes that the other was withdrawn. The separate key is an authored teaching aid, not output from a model test or customer engagement.
Prepare the blank sheet before opening the key. I-01 has the exercise's required details. I-02 lacks a quantity. I-03 has two quantities that need a staff decision. Even I-01 needs review; none of the supplied messages proves that work was accepted, a reply was sent or a booking was confirmed.
Using only the supplied fictional messages, prepare one intake row per request ID. Record source references for the proposed fields. Leave missing values blank. Retain both conflicting values without choosing one. Use ready for staff review or hold, not accepted or confirmed. Leave approved_by and confirmation_sent_at blank. Do not browse, contact anyone, book work or update a live system.
For an actual trial, use information your business has approved for the chosen tool and account. Keep credentials and customer records out of shared training files. The example addresses use the documentation-only .example domain and are not destinations for messages.
3What stays with you
You decide the required fields, who can see the information and who may accept the work. A customer's requested date is not proof that you have availability. A reply address in a message is not verification of identity or authority. Staff handle those decisions and any customer follow-up.
This practice does not connect to your inbox, forms, CRM or agency management system. Before any live connection, agree the actual access, provider data handling, reviewer and stop procedure. For insurance work, licensed staff retain coverage decisions and advice.
4What happens when it is wrong
Reject a proposed quantity that is not supported by the source. If a result silently picks one of I-03's values, put the conflict back on hold and retain both references. If it describes I-01 as booked, remove that claim: the messages supply a request, not an acceptance.
Keep the initial output, the correction and the reviewed version together. Change the instructions and try another fictional case before deciding whether the process is useful. Re-running the draft is not evidence that the underlying question was resolved.
5What it costs
Compare the same checked records prepared manually and with AI assistance. Count collecting the messages, preparing the instructions, checking every field, resolving exceptions and filing the approved record. Keep setup and upkeep separate from repeat task time. The practice pack establishes no savings or accuracy rate.
6When not to bother
If a short form and a staff check already cover the task, use them. Do not automate the acceptance of work when the inputs do not establish availability, authority or the agreed details. For a single ambiguous request, asking the right person may be the useful next step.
Updated : added related guides. Cost estimates and source-check dates retain their stated dates.
7Want this running on your own computer?
I'll set it up, and keep it running if you like. On a Mac mini, a Windows PC or Omarchy, with the agent you use or want: OpenClaw, Hermes, Claude Code, Codex or another. Approval settings depend on the agent, tools and configuration. Tell me about the job and I'll quote it. If a $20 subscription does it, I'll tell you. You keep the computer, the accounts and the setup.
Tyler Mayberry, who runs Cirlet (formerly Animas)
For independent insurance agencies nationwide, Cirlet offers AI and automation help. The first 30-minute conversation is free. Paid work and ongoing support are scoped and priced before they start.