Construction / Contracting

WhatsApp AI Assistant Inside a Contracting Company's ERP

Managers of a contracting company in Riyadh ask about projects, send invoices and voice notes, and receive Excel and PDF reports on WhatsApp, while nothing is recorded in the ERP without their approval.

Built by the JK Software Solutions engineering team.

Confidential Client Project
WhatsApp AI Assistant Inside a Contracting Company's ERP

The Challenge

The company already ran its projects, procurement and site teams on an ERP, and its WhatsApp number was on the Cloud API for technician notifications. Managers still had to open the system for every question, purchase invoices were typed in by hand and split across projects one line at a time, and late tasks or plans waiting for approval were noticed only when someone checked.

Our Solution

Message path: Meta keeps delivering the company's webhooks to its ERP. The ERP forwards only messages from linked team numbers to the JK Automation OS assistant through a relay signed with HMAC-SHA256, and pulls the current team list from a signed endpoint. Customers keep talking to the existing bot. The assistant reads and writes through a dedicated agent API on the ERP, authenticated with a key stored only as a hash. People and permissions: each employee links their own WhatsApp with a one-time code valid for 24 hours. A phone number alone grants nothing, and each linked number has its own areas, such as projects, purchasing or payroll. What the assistant can do: list projects, employees and pending documents; produce operations, purchasing, project finance, productivity, payroll draft and VAT worksheet reports as Excel or PDF; and transcribe voice notes with whisper.cpp in a separate worker. Two operations write to the ERP, a purchase entry split by project and the allocation of employees' working days to projects, and both are high risk: the manager approves them from a WhatsApp list, with a PIN when one is set. Invoices: a manager sends a photo or PDF. Claude extracts supplier, lines and amounts in a call with no tools, the server recomputes quantity times price, line totals, subtotal plus VAT and the total, flags an invoice number that was already recorded, and lets one line be split between projects by quantity. After approval the entry is written from the stored, checked extraction and the original image is archived with it. Follow-ups inside the ERP: in Riyadh working hours, from 7 am to 8 pm Saturday to Thursday, a technician who has not accepted a task gets a reminder after 3 hours and the project engineers are alerted after 24. Engineers get a daily reminder for site reports waiting more than 24 hours, owners for execution plans waiting more than 12, and owners receive an operations report as a PDF every day at 6 pm, every Thursday and on the 1st of each month.

Objectives

Let managers and staff work with the ERP from WhatsApp in plain Arabic. Turn a photo or PDF of an invoice into a purchase entry split by project. Never write to the ERP without an explicit approval from an authorized manager. Keep the company's number, its Meta setup and its existing bot unchanged. Give each linked number only the areas its owner should see. Send reminders, escalations and periodic reports without anyone asking.

Outcomes

The assistant has been live for the company since November 2026, on its existing number and without any change to its Meta setup. In live tests from a real phone a number was linked with a code, the assistant answered questions from the ERP, produced a payroll Excel report, read four invoices sent as photos and PDFs, and transcribed a voice note. A read-only recording on 18 November 2026 shows it listing the company's projects from the ERP and sending the weekly operations report as a PDF; a reply on Sonnet took about five seconds and cost under two US cents. The company's name, number, employees and project names are hidden in every image and in the video.

Challenges

Two systems answer on the same number: the ERP's own bot and the new assistant. The relay solves it by forwarding only linked team numbers, except bot keywords and buttons, so neither side answers the same message twice. Money had to stay under human control. Each write is bound to an operation id, a hash of its input, the target's version and a single-use nonce, so a stale or edited approval is rejected, and the amounts come from the checked extraction rather than from the model. The first live recording showed the assistant repeating an earlier answer from conversation memory instead of reading the ERP. Status and number questions now always go through the tools, and older turns carry their date. Whisper inside the API process ran out of memory on the first real voice note, so transcription moved to a separate worker lane.

Our Role

JK Software Solutions built the assistant in JK Automation OS (identity linking, per-number permissions, the ERP tools, approvals, document checks, reports and billing) and, on the company's ERP, the signed relay, the agent API and the follow-up service, then deployed both and ran the live tests.

Technologies

WhatsApp Cloud APIClaude APITypeScriptFastifyNestJSPostgreSQLHMAC-SHA256whisper.cppExcelJSHeadless Chrome PDF

Related Service

AI & Automation Solutions

Project Documentation

System Media Gallery

6 documented assets

Project Media

6

Have an idea but not the technical details?

Send us the idea and we will respond within 24 hours with the next practical step.

The initial consultation is free with no commitment, and for suitable projects it can include discussing an early visual concept.

WhatsApp AI Assistant for a Contracting ERP