An AI Agent That Learned
to Do a Human's Job.

A Personal Assistant WhatsApp chatbot trained on 11 year chat history

The Problem

Human-Dependencies

Hasherut ran a task-fulfillment service entirely over WhatsApp. Every client request still required a person to read it, classify it, price it, and assign it.

Human-Dependency Bottleneck

Every client request required a human dispatcher to read, interpret, and route it manually. The entire operation scaled with headcount, not software.

Classification by Memory

Hundreds of active client threads across 15 service categories. Categorization, pricing, and routing all sat in one person’s head.

Knowledge Trapped in Text

11 years of pricing logic, escalation rules, and edge-case handling existed only as informal Hebrew chat. No database, no knowledge base, no way to onboard a new dispatcher.

The Solution

Personalized WhatsApp AI-Chatbot

Replacing human-dependencies with AI-Assistance that is trained on tasker’s knowledge base and client’s personalized tasks memory.

Architecture

A software that thinks and acts for you

The same architecture now called “agentic AI with tool use”, designed and shipped in 2020–2023, before the terminology existed.

PRODUCTION FLOW WhatsApp message in Hebrew · informal · multimodal references LAYER 01 Understanding Classify the intent · auto-complete · confirm with user LAYER 02 Knowledge Retrieve the matching procedure · inject into prompt context LAYER 03 Execution Run the multi-step workflow · call APIs · gather missing info LAYER 04 Safety Gate 5-minute human override · deterministic fallback if no intervention Task created in Monday.com type · urgency · client · attachments · via API WhatsApp completion message sent through the same client number
LAYER 01

Understanding · Intent Classification

21,102
Labeled Tasks
153
Unique Clients
1,561
Action-Verb Patterns

From 21M+ WhatsApp messages, the agent learned to recognize 1,561 unique intents.

Deep dive

Every inbound WhatsApp message routes through a classifier that places it into one of 15 service categories, auto-completes any missing fields it can infer, and asks the user to confirm before proceeding. The structural backbone the model learned to recognize was extracted from the company’s 11 years of chat history.

Bureaucratic4,018
Insurance3,761
Secretarial3,433
Errands1,602
Vehicle1,490
Computer / IT1,436
Secretarial A.M.941
Purchasing876
Repairs764
Delivery718
Technician672
Professionals671
Escort393
Transportation266
Misc61
How the training data was built
01 · EXTRACT
WhatsApp text exports
Programmatic export of client threads spanning 2011–2020.
02 · TAG ROLES
Client vs. tasker per message
Speaker classification across mixed Hebrew / English threads.
03 · TAXONOMY
15 categories, 1,561 verbs
Hebrew keyword sets aligned to each service category.
04 · Q&A PAIRS
SQuAD-format records
VBA + Excel macros generated context / question / answer triples.
LAYER 02

Knowledge · Retrieval-Augmented Generation

145
Procedures
67
Formal Protocols
97
Unique Workflows
21
Procedure Categories

Constrain the model to the company’s own rules.

Deep dive

Once an intent is classified, the agent retrieves a matching procedure from an authored corpus and injects it into the prompt at request time, constraining replies to the company’s own rules instead of letting the model invent policy.

LAYER 03

Execution · Tool Use & Multi-Step Workflows

Purchasing protocol

Quote → approval → fulfillment → receipt.

Vehicle service protocol

Schedules an appointment with an authorized garage, threading vehicle & client metadata.

Insurance claim filing

Collects evidence and stages the government / insurer form packet.

Monthly billing

Closes the open task ledger, reconciles, generates the period invoice.

Dissatisfied client (5 steps)

Detect → triage → remediate → escalate → close-the-loop.

Follow-up tracking

Scheduled nudges to external parties holding up an open task.

The agent runs the workflow or assigns the task to a human-tasker with all supporting context for human task execution.

Deep dive

Each procedure is a multi-step workflow the agent runs, not a paragraph it reads back. API calls, sequential coordination, asking the user for missing information. Every variant of what to ask and when was mapped and trained.

Model evolution Fine-tuning ran across two platforms over 12+ months: OpenAI (Babbage → Davinci → GPT-3.5 as each became available) and HF AutoTrain. Hundreds of cycles tightened Hebrew fidelity, PII redaction, conversational state, and procedure adherence, not parameter counts.
LAYER 04

Safety Gate · Human-in-the-Loop Override

Five minutes for a human to override any outgoing message from the agent.

HUMAN OVERSIGHT

Human-in-the-loop control

Every agent output passes through a 5-minute human review window. Operators can edit, reclassify, or reject before delivery.

TRANSPARENCY

Explainable decision trace

Full reasoning chain logged per session: detected intent, retrieved procedure, staged reply, and every branch the agent took.

ROBUSTNESS

Deterministic fallback

If no human intervenes, the system defaults to a safe deterministic message. No unreviewed model output ever reaches the client.

ACCOUNTABILITY

Grounded, auditable generation

Every reply is constrained to the 145-procedure authored corpus with a per-session audit trail.

Deep dive

After the agent prepares a reply, a 5-minute human-override window opens. A human operator watches the session through a Debug Dashboard and can edit, reclassify, or approve. If no one intervenes, the system falls back to a deterministic message: “We’ll get back to you soon.”

Reflection

What Three Years Taught

The biggest challenge was getting the AI to pay attention to the exact right details for each person, forcing me to spend countless hours manually cleaning out useless information so it wouldn’t get confused. If we had an automatic way for the system to dynamically filter and focus on what mattered in real-time, it would have changed everything.