Key facts
- Business
- Kota sandstone and natural stone supplier for landscaping, indoor and outdoor use.
- Products
- Kandla Grey, Mandana, White Mint, Jaisalmer Yellow, Modak and quartzitic sandstone.
- Buyers
- Architects, builders, real-estate developers, landscapers, contractors, hotels, importers and exporters.
- Design
- One AI agent with 7 Phase 1 modules and an approval matrix.
- Control
- 8 action types always need owner approval, including ads, budgets and external messages.
- Logging
- 100% of agent actions logged with time, source, result and "AI-done" or "owner-approved".
- Expected
- A daily briefing ready every morning in about 2 minutes of reading (design target).
Problem
What did the business ask for?
The owner wanted one AI assistant to help run the business, not ten separate tools. The owner wanted to stay in control of every important decision.
The business told us:
- They sell to many buyer types, and leads come from the website, ads, calls and personal contacts.
- Follow-ups get missed, and records are spread across places.
- Research on new buyers and competitors takes a lot of time.
- They want to speak to the assistant: ask questions, give tasks and hear a daily briefing.
- In their words: reliability and clear execution matter more than many features.
Their success criteria were simple: save time, keep records correct, miss fewer follow-ups, research faster, get better leads and see the whole business in one view.
Why is this an agent and not a chatbot?
A chatbot answers questions. This agent does work: it reads data, takes steps across tools, and asks for approval before risky steps. You can read more on why this is an agent, not a chatbot.
What can the agent do alone, and what needs approval?
The agent does routine, internal work alone. Anything that reaches a customer, spends money or changes important data waits for the owner. This was the client's own rule, and we built the design around it.
| Action | Automated | Needs owner approval | Always logged |
|---|---|---|---|
| Organize data, classify leads, remove duplicates | Yes | No | Yes |
| Research prospects, competitors and industry news | Yes | No | Yes |
| Write summaries, reports and reminders | Yes | No | Yes |
| Draft emails, WhatsApp and LinkedIn messages | Yes (draft only) | To send | Yes |
| Publish social media posts | No | Yes | Yes |
| Major website changes | No | Yes | Yes |
| Launch Meta ads or raise budget | No | Yes | Yes |
| Delete data or bulk-change records | No | Yes | Yes |
| Financial or legal commitments (prices, terms) | No | Yes | Yes |
Each log entry stores: timestamp, action, result, data source, and "AI-executed" or "user-approved". If a step fails, the agent reports it in the next briefing and on the dashboard. It never fails silently. It never invents a contact, a phone number or a fact. Research results are marked "verified" (with a source link) or "assumed".
Solution
How is the system built?
The design has one agent "brain" connected to six parts. Each part has one job.
| Part | Proposed tool | Job |
|---|---|---|
| Business memory | Notion (via Notion API) | Products, prices, buyer types, rules, past decisions |
| Lead pipeline | A CRM (existing one, or a simple one we set up) | Leads, stages, de-duplication, history |
| Channels | Email, WhatsApp Business API, LinkedIn (manual send) | Send approved messages, receive replies |
| Ads | Meta Marketing API and Lead Ads | Draft campaigns, pull ad leads into the CRM |
| Voice | Speech-to-text and text-to-speech on phone or web | Ask, dictate, hear the briefing |
| Dashboard | Web command center | Approvals queue, pipeline, logs, alerts |
Text description of the diagram: the owner talks or types to the agent. The agent reads Notion first, then acts on the CRM, channels or ads. Every risky action goes to the approvals queue. Every action writes to the log.
How it works
How does a lead move from first contact to follow-up?
A new lead is captured, checked, scored, drafted for, approved, sent and then tracked until the next step. Here is the step-by-step flow.
- Capture. A lead arrives from a website form, a Meta lead ad, an email or a manual entry.
- De-duplicate. The agent checks phone, email and company name against the CRM. A match is merged, not duplicated.
- Classify. The agent tags buyer type (architect, builder, hotel, importer), product interest and city.
- Enrich. The agent looks up the company from public sources and marks each fact as verified or assumed.
- Draft. Using Notion (products, price ranges, tone), the agent drafts a first reply by email or WhatsApp.
- Approve. The draft goes to the approvals queue. The owner approves, edits or rejects it, by tap or by voice.
- Send and log. The approved message is sent. The CRM stage and history update.
- Remind. If there is no reply in the set time (for example 3 days), the agent drafts the next follow-up and puts it in tomorrow's briefing.
- Report. The lead appears in the daily briefing until it is won, lost or paused.
What does the daily briefing look like?
The daily briefing is a short summary the owner can read or hear each morning. Below is a sample with made-up data, to show the format.
Good morning. Briefing for Thursday.
- New leads: 4 (2 from Meta ads, 1 website, 1 email). 1 duplicate merged.
- Follow-ups due today: 6. Drafts ready for your approval: 6.
- Hot lead: a landscaping firm asked for Kandla Grey prices for a villa project. No reply sent yet.
- Ads: Campaign "Mandana for landscapers" cost per lead went up 40% in 3 days. Suggest: pause or change the creative. Needs your approval.
- Competitors: one competitor added a new product page for White Mint.
- Failed: 1 WhatsApp message did not deliver (number not on WhatsApp). Suggest: send by email.
- Waiting for you: 9 approvals.
What can the owner say to the agent?
The owner can give commands by voice or text. Here are sample commands and sample agent replies.
| Owner says | Agent replies (sample) |
|---|---|
| "Give me today's business briefing." | Reads the briefing above, then asks: "Do you want to review the 6 follow-up drafts now?" |
| "Check all follow-ups due today and prepare the messages." | "6 follow-ups are due. I drafted all 6. 4 by email, 2 by WhatsApp. They are in your approvals queue." |
| "Research 20 landscaping and real-estate companies and add qualified prospects." | "I found 20. 12 match your buyer rules. I added them to the CRM as 'Prospect - not contacted'. 5 have verified emails. 7 contacts are marked 'assumed'." |
| "What did competitors change this week?" | "2 of 5 tracked sites changed. One added a product page. One changed its homepage offer. Links are in the report." |
| "Create a lead-ad campaign for this product and show me the plan and budget before launch." | "Draft ready: audience, 3 ad texts, a lead form and a suggested daily budget. Nothing is live. Say 'approve' to launch." |
| "What happened with this lead and what should I do next?" | "First contact 12 days ago. Sample sent. No reply to 2 follow-ups. Suggest: a short call. I can draft a call note." |
You can learn more about the voice assistant side of this kind of agent.
What is in Phase 1, and what is in Phase 2?
Phase 1 is a small, reliable core the owner can use every day. Phase 2 adds more channels and monitoring once Phase 1 works well.
| Phase 1 (MVP) | Phase 2 |
|---|---|
| A. Notion business memory | Website monitoring and approved updates |
| B. Lead management with de-duplication | Website visitor to lead capture |
| C. Follow-ups with approval | Competitor monitoring at scale |
| D. Prospect research with sources | Industry news and trends |
| E. Daily briefing | Social publishing (IG, FB, LinkedIn, X) with approval |
| F. Voice control | Weekly reports and decision support |
| G. Meta lead ads with approvals | Full command-center dashboard |
Phase 1 still includes a simple approvals screen and the action log. Approval and logging are not optional, so they come first.
What is practical today, and what is limited by platforms?
Most of this design is practical today. Some parts are limited by platform rules, so we design around them honestly.
- WhatsApp Business API: Practical. But the buyer must opt in. Messages the business starts outside a 24-hour reply window must use Meta-approved templates, and those are paid per message.
- Meta Ads API: Practical. The agent can draft campaigns and read lead form data. It needs a verified Meta Business account and API access. Meta reviews every ad, so launch is never instant.
- LinkedIn: Limited. There is no open API for personal messages or connection requests. Automation tools can get the account restricted. So the agent drafts, and the owner sends by hand.
- X (Twitter): Limited by cost. The X API is paid, and reading data at scale is costly. We suggest posting only, or leaving X for later.
- Notion API: Practical, but it has rate limits. Large imports run slowly in the background.
- Prospect research: Practical from public websites and search. Some contacts cannot be verified, so they stay marked "assumed".
Expected outcomes
What impact do we expect?
Expected outcomes
These are design targets, not results. Each one depends on the assumption next to it.
| Expected impact | Assumption |
|---|---|
| Every new lead logged in the CRM within 5 minutes | Website form and Meta Lead Ads are connected; no manual entry needed |
| Close to zero missed follow-ups | The owner clears the approvals queue at least once a day |
| No duplicate lead records from connected sources | Leads share a phone or email the agent can match |
| A list of 20 researched prospects in under 1 hour of agent time | Public websites exist for most target companies |
| Daily briefing in about 2 minutes of reading or listening | The CRM and ads data are up to date |
| Poor cost per lead flagged within 1 day | The owner sets a target cost per lead in Notion |
We did not get time or lead volumes from the business, so we do not promise hours saved. We would measure them after launch.
What will this agent not do?
The agent will not act alone on anything that is risky or public. It has clear limits.
- It will not send external messages, publish posts, launch ads or raise budgets without approval.
- It will not delete data or change many records at once without approval.
- It will not make price, contract or legal promises.
- It will not invent contacts, numbers or facts. Unknown facts stay "assumed".
- It will not automate LinkedIn messages or scrape sites against their terms.
Risks and open questions: CRM choice, Meta Business verification, WhatsApp template approval, Hindi voice accuracy for the owner's speech, and the time the owner can give to approvals each day.