A Voice Ask-Your-Data Assistant and Deposit Reconciliation for a Multi-Location Retailer

Retail Operations Voice AI Financial Reconciliation

The Client

The client runs several retail locations. Its back-office system prints daily sales reports for each location. AppWT does not name the client here.

What the Client Needed

The client needed two things. The first was fast answers from its daily sales reports, asked by voice from a computer or a phone. The second was a daily check that each location's expected deposits match what actually reached the bank.

What AppWT Built

1. A hub on the client's own network. A small always-on Windows PC watches the folder where the back-office system saves its reports, loads them into a local database and answers questions. It is reachable only on the local network and a private encrypted network, never from the public internet.
2. Nothing new on the back-office PC. The reporting system keeps working exactly as before. Its reports also save to a shared folder, and the hub reads them from there.
3. A desktop app with push to talk. Speech turns into text on the computer itself. The question goes to the hub, and the answer comes back as text, a table or a chart, with its source, and is read aloud in the computer's own voice.
4. A phone web app. It installs on the home screen of an iPhone or Android phone. It uses voice where the browser supports it, offers a type box where it does not, and reads each answer aloud.
5. Answers that cannot change the data. The hub only reads. Every question runs as a read-only query against four approved data views, and a SQL guard blocks anything else.
6. An AI model that never sees a figure. When AI turns a question into a query, it sees only the question and the names of the views and columns, never the numbers. Common questions are answered by fixed rules with no AI at all.
7. Every answer names its source. Each answer states the report and the date it came from, so a person can check it.
8. A key for every device. Each computer and phone has its own key, and the hub stores those keys hashed.
9. Deposit reconciliation. Expected deposits for each location and day, across cash, card settlements, benefit cards, delivery apps and other channels, are matched to the bank's export file. Matching runs in three passes: exact, grouped across days, and net of fees, to the cent.
10. Exceptions a person approves. Unmatched items age at 3, 7 and 14 days, and the accountant approves each one. The engine never posts to the accounting system. It exports a CSV file instead.
11. Sales data stays at the client. AppWT's own server holds the code, the builds and the documents. It never holds the client's sales data.
12. An offline install kit. A USB drive carries everything needed to install the hub at the location.

How AppWT Tested It

Every count below was measured on sample data, on two consecutive runs with the same result both times.

• Hub: 260 automated tests passed, covering the query guard, device keys, the AI path, report loading and reconciliation.
• Report loading and reconciliation: 89 tests passed. These are part of the 260.
• Acceptance questions: the hub handled all 40 questions in the acceptance set, with 38 answers and 2 refusals.
• Desktop voice loop: 8 of 8 passed, from a recorded voice file through speech to text to the answer.
• Phone app: 53 of 53 voice checks against the hub and 15 of 15 voice unit tests passed.
• Static checks: 81 of 81 passed across the phone app, the installer and the project rules.
• Windows installer: built and tested under Wine on AppWT's build server.

Measured Outcomes

Honest Status

Every result on this page was measured on sample data. The real report formats, the live microphone and the desktop window on a real Windows PC are not proven yet, and the Mac version is scripted but not yet built. The system stays in test status until it passes against the client's own reports at the on-site install.

What This Means for a Multi-Location Business

If your numbers live in daily reports, a voice assistant can answer from them without moving your data off site. AppWT keeps the AI away from the figures, makes every answer name its source, and leaves the final approval on money with your accountant. See AI consulting, API development and integration and AppWT software and platforms.