WhatsApp + quoting

WhatsApp AI quotes and estimates: answer, collect details, price from your rules, hand off to staff

Yes. An AI agent on WhatsApp can answer the common questions, collect the details an estimate needs, send a quick estimate from your own price rules, and hand serious issues to a person with the full conversation attached. It is built for service businesses and retailers that get a steady flow of "how much would it cost to..." messages on WhatsApp and Instagram and lose jobs because the reply comes the next day. The design rule that makes it safe: the model gathers information and writes the message, but the number comes from your rate card in code, never from the model. Small jobs get a range or a firm price on the spot; anything above a threshold you set waits for a person to approve. I build this as an independent AI systems engineer on Meta's official Cloud API, in your accounts, as a three-week fixed-price sprint with a 60-day operational guarantee.

How does a WhatsApp AI agent collect the details an estimate needs?

Every service you quote has a short list of required fields. A cleaning job needs the address, the size of the space and the frequency. A repair needs the item, the symptom and a photo. An installation needs measurements and access conditions. The agent works from that list, written per service in a configuration file you can read, and asks only for what is still missing. If the customer already said "two bedrooms in Viña, every two weeks" in the first message, it does not ask again.

Photos and voice notes are part of the intake. Voice notes are transcribed to text. Photos are stored against the conversation so the person approving the estimate sees them, and a vision-capable model can flag the obvious cases (wrong item, unreadable photo, a picture that suggests the job is bigger than described). The agent confirms the collected details back to the customer in one short summary before any number appears, which catches most misunderstandings early.

Required fields defined per service, not improvised by the model
Asks only for what is missing, in natural language
Photos and measurements attached to the case
Voice notes transcribed and treated like text
A confirmation summary before any price is shown

How is the estimate calculated, and can the AI invent a price?

It cannot, by design. Your price list or rate card lives in code or in a sheet the system reads: base prices, unit rates, minimums, surcharges for distance or urgency, taxes. When the required fields are complete, a deterministic function computes the estimate from those rules. The model's job is to extract the inputs from the conversation and to write the reply around the number. If an input is missing or outside the rules, the function returns "no price" and the agent says a person will confirm.

You decide what kind of answer each case gets. Well-defined small jobs can get a firm quote immediately. Jobs with real uncertainty get a range ("between X and Y depending on what we find on site"), which is honest and still fast. Anything above an amount you set, or with a discount, a custom item or an unusual condition, goes to an approval queue. A person sees the details, the photos and the computed number, approves or adjusts it with one action, and the agent sends it. Every estimate is logged with the inputs and the rule version that produced it.

What counts as a serious issue, and how does the handoff to staff work?

You define "serious" in writing during the first week, and the agent applies it consistently. Typical rules: a complaint about a previous job, a damage or safety report, a legal or billing dispute, an angry or repeated message, an explicit request for a person, a large or unusual job, and any question the knowledge base cannot answer. Some rules are keyword and pattern checks, some are a classification the model makes, and the strict ones are both, so a safety report is not left to a single judgment call.

When a rule fires, the agent stops replying on that conversation, tells the customer a person will follow up, and notifies the right staff member in WhatsApp, Slack or your CRM with a short summary, the collected details, the photos and a link to the full thread. Your team answers from the same number. The agent stays paused on that thread until a person releases it, so the customer never gets a bot reply in the middle of a human conversation.

Escalation rules written by you, versioned like code
Complaints, damage, safety, billing disputes and frustration go to a person
Summary, details and photos travel with the handoff
Agent pauses on the thread until staff release it

Can the same flow handle pricing questions from Instagram DMs?

Yes. Many businesses get the same "how much for..." questions in Instagram direct messages as on WhatsApp. Meta offers an official messaging API for Instagram professional accounts, and the agent can receive those messages through a second webhook that feeds the same runtime. The required fields, the rate card, the escalation rules and the audit log are shared, so a price is the same whichever channel the customer used.

The differences are practical. Instagram conversations tend to start from a post or a story, so the agent reads that context when it is available. Some customers prefer to move to WhatsApp once a real job is on the table, and the agent can offer that with a link to your number. Each channel has its own messaging rules and windows, which I check against Meta's current documentation before a build, because they change.

What makes a WhatsApp AI quoting setup stable instead of fragile?

The first decision is the channel. Tools that automate WhatsApp Web or the phone app by imitating a person clicking are unofficial, break when WhatsApp updates, and put the number at risk of being banned, which for a business that sells on WhatsApp is the worst outcome available. A stable setup runs on the WhatsApp Business Platform through Meta's official Cloud API or a business solution provider. I run a WhatsApp AI agent in production on the official Cloud API in coexistence mode for a small business in Chile: the business kept its existing number and keeps using the WhatsApp Business app on the phone while the agent answers through the API.

The second is engineering. Meta delivers messages to a webhook that must answer quickly, so the system acknowledges first and processes after, deduplicates repeated deliveries, and retries failed calls to the model or your systems. Every message, estimate and escalation goes to a log in your database. Monitoring alerts a person when the agent stops replying, when error rates rise or when the queue of pending approvals grows. Prompts, rate cards and rules are versioned, so a change can be rolled back.

Official Cloud API or a business solution provider, never WhatsApp Web automation
Coexistence mode keeps your number and the phone app where supported
Acknowledge, deduplicate, retry
Audit log of every message, estimate and handoff
Alerts on silence, errors and a growing approval queue

What does the architecture of a WhatsApp quoting agent look like?

In plain words: a customer message arrives at a webhook Meta calls. The runtime, an n8n workflow or a small Python service, loads that contact's conversation and the case being built, retrieves relevant fragments of your FAQ and policies from Postgres with pgvector, and asks the model (Claude or OpenAI through their APIs) to do one of four things: answer a question, ask for a missing field, hand the completed fields to the pricing function, or escalate. The pricing function reads your rate card and returns a number, a range or "needs a person".

Firm quotes under the threshold go straight back through the API. Everything else lands in an approval queue your team works from WhatsApp, Slack or a simple web view. Approved estimates are sent, recorded in your CRM or quoting sheet, and can trigger a booking link or a payment link from the provider you already use. It all runs in your Meta account and your hosting, with your API keys.

Trigger: inbound WhatsApp or Instagram message at your webhook
Data: rate card and service fields in code or a sheet, FAQ in Postgres with pgvector
Model: Claude or OpenAI, for extraction and wording only
Guardrails: deterministic pricing, approval threshold, escalation rules, audit log
Runs in: your Meta account and your hosting; you own code, prompts and data

How much does a WhatsApp AI agent that sends estimates cost?

The reference build is the AI Intake Agent with knowledge base, a productized scope quoted at a fixed price, with an optional monthly operation for iteration after the guarantee period. Intake, answers from your documents, escalation and the audit log are that scope. A quoting agent with a rate card, an approval queue and more than one channel is usually quoted as a single fixed price after a discovery call, with the AI Operating System as the upper reference point.

What is not inside the fixed price: the messaging fees Meta or the provider charge you, the model API usage and the hosting. Those are billed to you directly, in your name, with a cap declared in the proposal. I do not bill hourly for builds and I do not resell software.

When is a quote form or an off-the-shelf tool enough?

If your prices are a short fixed list, a pinned message, a catalog in the WhatsApp Business app and quick replies cover it for free. If every estimate needs the same five fields and a person always reviews it anyway, a simple web form linked from WhatsApp, feeding a sheet, is cheaper than an agent. If you only need frequent questions and routing to the right person, a WhatsApp chatbot platform you configure yourself is the right purchase. I will say so on the first call, because I do not resell tools.

A custom agent earns its cost when the volume of quote requests is high, the details arrive messy and spread over several messages, pricing has real rules, and the delay between question and estimate is losing you jobs. The 15-minute call is where we check that against your actual conversations.

Frequently asked questions

Is there a tool that answers WhatsApp questions, generates estimates and hands serious issues to staff?

Chatbot platforms cover frequent questions and routing well. Estimates from your own price rules and handoffs with full context depend on your systems, so that part is usually a build on the WhatsApp Business Platform connected to your rate card, CRM and team channel. The customer sees one conversation on one number; behind it, the pricing is deterministic and a person approves anything above your threshold.

Will the AI ever make up a price?

No. The price comes from a function that reads your rate card. The model extracts the inputs and writes the message around the number. If the inputs are incomplete or the job falls outside your rules, the agent says a person will confirm, and the case goes to the approval queue instead of guessing.

Can I keep my current WhatsApp number and the app on my phone?

Often yes, through coexistence mode on Meta's Cloud API, where the business keeps its number and the WhatsApp Business app while the agent answers through the API. Eligibility depends on Meta's current rules for your country and account, which I check with you in the first week.

Should estimates be ranges or firm quotes?

Both, per service. Well-defined small jobs can get a firm price immediately. Jobs where the real scope is only visible on site get an honest range with what it depends on. Anything above your amount threshold, with a discount, or with an unusual condition waits for a person to approve.

Why not use a cheaper tool that automates WhatsApp Web?

Because it is unofficial. Those tools imitate a person using WhatsApp Web, break when WhatsApp changes, and put your number at risk of a ban. For a business whose sales run through that number, the savings are not worth it. I build only on the official Cloud API or through a business solution provider.

What are the limits of this setup?

The agent is only as good as your rate card. If your pricing lives in someone's head and changes job by job, we first write it down, and some cases will always go to a person. It also cannot see a site: photos help, but a complex job still needs a human visit. And I do not build cold outreach on WhatsApp.

Fifteen minutes on the quote requests waiting in your WhatsApp.

A free call. I'll tell you straight which processes today's AI can solve and what your infrastructure needs for them to actually run. If it fits, we move forward. If not, I point you the right way, free.