Helpline Callback Desk

Staff workflow prototype

Independent prototype of Legal Aid Society of Middle Tennessee's helpline intake, by The AI Pipe
Invented calls, live agent

A caller says the thing that matters at two minutes forty one, in Spanish. Here is what a desk does with it.

One real execution on an invented call, recorded end to end: loading. The desk is frozen at 2026-09-16 and nothing on it moves with the day you read it on. Everything explaining it sits below the result.

Simulated helpline activity

The library

You are talking to an AI system. Every live answer below is written by a language model, Claude Sonnet 5, calling the tools of this desk over invented records. Nobody reviews it before you read it, and nothing here is professional advice.

Fictional desk. Do not paste real caller, client or case records. Whatever you type here is sent to the model that answers, and it is not kept in the counts this page keeps of its own traffic.

Enter to send, shift and enter for a new line. Live tool use over ten invented calls, the published offices and counties, and the public rules.

1. Observed public material

The helpline number, the callback promise, the two sentences about what a call is not, the eight problem areas, the eight offices with their counties and their published hours, and the paragraphs of the organization own public telephony solicitation, read on 2026-09-16. Each row carries its status.

2. Simulation

Ten invented calls with their transcripts, invented callers, relatives and attachments, and the invented internal policy of an invented program. Every clause of it is labelled simulation policy, not this organization's policy.

3. Public rules, verbatim

45 CFR 1611, 1620 and 1626 and the 2026 income appendix, quoted from eCFR with the source and the date. Real rules, never paraphrased, never attributed to the organization.

What this shows

The problem: A helpline takes calls all day and publishes that it usually returns them within 24 to 48 hours. Somebody has to listen to each one, work out where the caller lives, what the matter is, who the other party is, and which one somebody should look at first. An ordering by arrival time cannot know that the eleventh thing somebody said, in Spanish, was a month, a day and an hour off a piece of paper.
The result: The run at the top of this page is real and is recorded: the transcript read to the end, the date and the hour read out of the Spanish itself rather than out of anything normalised in advance, the postal routing run in code, the published sentence turned into two moments with the four assumptions behind them printed, the checks run, the public rule quoted verbatim, a structured new matter with its empty fields named, and an ordering of today callbacks that the model cannot touch.
What the recorded run above actually does: The public Get Help page offers one route, a toll-free number, and a sentence saying calls are usually returned within 24 to 48 hours. No intake form was found on it, so every request starts as speech. The run reads a transcript segment by segment in the language it was spoken in, reads the date and the hour out of the words themselves rather than out of anything normalised in advance, runs the postal routing in code, assembles a structured new matter in LegalServer shape with its empty fields named, and proposes an ordering of today callbacks that the model cannot touch. It determines no eligibility, answers no legal question, runs no conflicts check and contacts nobody. The time on it covers the documentary work before an intake specialist picks the call up.
Where it sits: It prepares a matter draft and proposes an ordering of callbacks, and hands over a sheet that can be copied out. The ordering is this prototype own proposal: the organization publishes no statement of case priorities and no queue configuration, and neither is guessed at here. Standalone demonstration. It connects to nothing, writes into nothing, and is not a way to ask for legal help.
What is simulated, and what is not covered: The program, the ten calls, the callers and the internal policy are invented, and that policy is one invented program's own, not anybody's law. The counties, the offices, the published hours and the postal codes are real and are marked row by row. Four things were not published and are therefore not guessed at: the written statement of case priorities, the case service report problem codes, the queue configuration, and the income and asset ceilings the organization adopts. There is no audio here: a transcript stands in for what a transcribing voicemail produces. The English text beside a segment is a working machine rendering, never an interpretation. No connection to any case management system, and no eligibility determination, no acceptance or refusal of a case, no legal advice, no conflicts check, no problem code and no deadline. The tool layer refuses all of those, and refuses a figure that is not written in a section it was cited from.
What a pilot would need from your side: Your blank screening questions, the helpline policy in whatever form it exists today, whatever is written down about how fast a call is promised back, a short conversation with whoever runs the line and one person who can say whether an answer is right, and counts rather than calls: how many arrive in each language, how many are returned inside the promise. Callers and their calls stay invented until your legal and security leads have agreed, in writing, which fields a pilot may see, who holds them, where they sit, how long they are kept and how they are deleted. No call transcript and no caller record are needed to start, and removing names is not anonymisation.
Who built it: Frederic de Lavenne de Choulot, The AI Pipe. Built as a demonstration, not sold as a legal, eligibility or compliance product.