WhatsApp Store AI Assistant
AI-powered conversational sales assistant transforming WhatsApp into an automated store with instant replies and product management.
AI agents
AI agent development for multi-step business tasks
An AI agent is software that uses a language model to decide which steps to take and then carries them out through tools: looking up an order, checking stock, updating a record, sending a message or asking a person for approval. I build agents for WhatsApp and chat, document workflows and internal operations, with clear limits on what they may do, human review for risky actions and controls on running cost.
When an agent is the right tool
- Customers ask on WhatsApp or chat for things that need several lookups, such as availability, prices, order status and next steps, before a reply is possible
- Staff spend time reading incoming documents, checking them against your records and filling in another system
- A request has to be gathered over a conversation, validated, and then turned into an order, booking or ticket
- Answers depend on your own policies, catalogue or knowledge base, and the agent must also act on them, not only quote them
- You already have APIs or a database the agent can use, and the rules for what it may change can be written down
What's included
Agent scope and permissions
A written list of the tasks the agent handles, the tools it may call, what it may only read, what it may change, and what always goes to a person. This is the most important design decision and it comes first.
Tool and API layer
Small, well-defined functions the model can call, such as find product, check slot, create draft order or escalate, built on your existing APIs or database with validation on every input.
Conversation and channel integration
WhatsApp Cloud API, web chat or an internal interface, with conversation state, hand-off to a human and templates that follow channel rules.
Document workflows
Agents that read PDFs, forms or emails, extract the fields, compare them with your records and prepare the next action for review.
Retrieval over your content
Retrieval-augmented generation over policies, product data or past answers, so the agent works from your sources rather than its general training.
Guardrails, logs and evaluation
Step limits, spending caps, approval queues for sensitive actions, a full log of every tool call and a test set of real cases to check quality before and after changes.
Agents versus LLM features, and why limits matter
My AI and LLM integration work adds a single language-model step into an application: draft this reply, summarise this document, extract these fields. An agent is different. It runs a loop, deciding which tool to call next based on what it has learned so far, and it can change things in your systems. That makes agents more capable for tasks with several steps, and also more risky, because a wrong decision can become a wrong action. If a task can be done well with one model call or a fixed workflow, I will recommend that instead.
When an agent is the right fit, I keep it narrow. It gets only the tools the task needs, reads are separated from writes, actions involving money, customer commitments or deletions wait for a person to approve, and every step is logged so you can see why it did what it did. Cost is controlled with step limits, smaller models for simple decisions and caching where answers repeat, and measured against real examples before launch.
How the project runs
- Pick one taskChoose a frequent, multi-step task and collect real examples, including the awkward ones, along with what a good outcome looks like.
- Define tools and rulesList the systems involved, design the tool functions and write down which actions need approval or must never happen.
- Prototype and testBuild a working agent against test data and run it through the example set to measure accuracy, step count and cost per task.
- Pilot with reviewRun the agent on real traffic with a person approving its actions, and tighten prompts, tools and rules based on what goes wrong.
- Launch and monitorAllow low-risk actions to run automatically, keep review for the rest, and watch logs, costs and escalations over time.
Technology choices
- OpenAI API and other LLM providersTool calling and structured outputs, with the model chosen per step for quality, speed and cost.
- Node.js and ExpressThe agent loop, tool functions, queues and integrations, run on the server so keys and rules stay private.
- WhatsApp Cloud APIConversational agents in the channel many customers already use, with human hand-off.
- PostgreSQL and vector searchConversation state, action logs, review queues and embeddings for retrieval in one place.
Related projects
Schooliee AI-Based School System
Pakistan's first AI-based school management system featuring automated parent WhatsApp notifications, rosters setup, attendance, and fee alerts.
Banana Arena AI Studio
AI-powered creative studio allowing users to generate high-resolution AI portraits, enhance images, remove backgrounds, and apply style filters.
Clipout Video SaaS
AI-driven video editing and data visualization platform that transforms raw data into animated visuals, presentation videos, and social clips.
What affects cost and timeline
- How many tools and systems the agent needs to use, and the quality of their APIs
- How much of the task can run automatically versus needing approval screens
- Request volume, the number of steps per task and the models used
- Whether retrieval over your documents or data is needed
- Channel requirements, such as WhatsApp business verification and message templates
Every estimate starts from your requirements. Request a quote with your users, must-have features, timeline and budget range, and I will reply with questions and a suggested first phase.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A typical chatbot answers questions. An agent can also act: it looks things up in your systems, fills in forms, creates records or sends messages, deciding the steps as it goes. That is why agents need permissions, approvals and logging that a simple chatbot does not.
How is this different from your AI and LLM integration service?
LLM integration adds a single model step to software, such as drafting or extracting. Agent development is for tasks with several steps where the model chooses which tools to use and takes actions. Many projects start as an LLM feature and grow into an agent once the task is well understood.
Can an agent make mistakes?
Yes. Language models can misunderstand a request or choose the wrong step. Narrow scope, validated tool inputs, approval for important actions and a test set of real cases reduce the risk; they do not remove it. The agent should be designed so a mistake is caught before it costs anything.
How do you keep running costs under control?
By limiting the number of steps per task, using smaller models where they are good enough, caching repeated lookups and tracking cost per completed task. A prototype on real examples gives a cost estimate before you commit.
Will the agent have access to all our data?
No. It only gets the specific tools it needs, each tool checks permissions on the server, and read access is separated from write access. Sensitive data can be excluded entirely or masked before it reaches the model.
Have a project in mind?
Tell me what you want to build, who will use it and when you need it. I will reply with questions and a practical next step.
Discuss your project