Why an AI chatbot costs USD 500 in one place and USD 5,000 in another

The difference between a USD 500 chatbot and a USD 5,000 one is almost never the provider's margin or better AI: it is how many of your systems the bot touches. Both use similar models and spend almost the same per conversation. The USD 500 one answers a set of questions someone loaded by hand and is not connected to anything. The USD 5,000 one reads your up-to-date price list, checks real stock or calendar availability, writes the appointment or the lead into your system, hands off to a person with the history in front of them and has someone watching when it breaks. Between the two there are 70 to 170 hours of integration, testing and tuning. The price measures plumbing, not intelligence.

The engine is a commodity; the plumbing is not

A conversation of 8 to 12 messages with search over your database consumes between 3,000 and 8,000 tokens. With a low-cost model that comes to between USD 0.006 and USD 0.015. That number is identical for the provider charging you USD 500 and for the one charging you USD 5,000: neither of them trained a model of its own, both hit the same API.

What does change from one proposal to another is four things: where the bot gets its information, where it writes it, what it does when it does not know, and who finds out when it stops working.

From that comes a practical rule for reading any quote. Zero connected systems is the low tier. One or two read-only systems is the mid tier. Two or three systems where the bot also writes, plus handoff to a human, is the high tier. If an expensive quote does not list concrete systems, it is not expensive: it is padded.

What you get for USD 500 and what it is missing

In the USD 400 to 900 range, what gets delivered is fairly similar across providers:

• Between 20 and 50 questions and answers loaded by hand by someone who read your website.
• One channel: your own number on WhatsApp Cloud API, or a widget on the site.
• Read-only answers: prices, opening hours, address, payment methods, delivery areas, accepted health insurance plans.
• Basic capture of name, phone and reason that lands in a Google spreadsheet or an email.
• A handoff that in practice is a notification to an internal group, with no history and no routing.
• Between 3 and 10 hours of actual work.

That is more than enough for a neighborhood hair salon, a single-location gym or an accounting firm that always answers the same fifteen questions.

It breaks at three specific moments. When you update the price list and someone has to go in and edit the bot by hand. When a customer asks whether size M is still available and the bot has no way to know. And when someone asks for an appointment on Thursday at 4:30 pm and the bot cannot look at the calendar. In all three cases the conversation goes back to a person, which is exactly what you were trying to avoid.

What each step up to USD 5,000 adds

The biggest price jump is between reading and writing, and there is a technical reason. If a read fails, the bot says it cannot check that right now and nothing happens. If a write fails halfway through, you end up with two overlapping appointments in the same professional's calendar, a duplicate lead in the CRM or an order with no stock reserved. That forces idempotency, controlled retries, overlap locking and a rollback path. It is the most boring part of the quote and it is the part that separates a bot that helps from one that creates extra work.

Two quotes for the same practice, with the full math

Dental clinic, three professionals, around 600 WhatsApp inquiries per month, 40% outside business hours. Same request, two quotes.

Quote A, USD 600 setup and USD 60 per month. Thirty-five frequently asked questions with prices loaded by hand, no connection to the calendar, every appointment request handed off to reception.

Quote B, USD 4,200 setup and USD 280 per month. Knowledge base over the list of treatments and accepted health insurance plans, updatable from a spreadsheet; reading from and writing to Google Calendar per professional with overlap locking; reminder 24 hours in advance with rescheduling through the same chat; handoff to Chatwoot with history, to reception during business hours and to a mailbox outside them; 120 evaluation questions run before every change.

Of the 600 monthly inquiries, around 250 are appointment requests. A answers price and hours and passes those 250 to the receptionist anyway. B closes between 55% and 75% of those requests on its own: this is a typical range for this kind of integration, not an audited result from a specific company. That is 140 to 190 appointments booked without anyone stepping in, a good share of them between 8 pm and 8 am.

Now the math almost nobody does. The difference is USD 3,600 once plus USD 220 per month. In the first year that comes to USD 6,240. If the average appointment leaves between USD 50 and USD 70 of margin, with 3 to 5 recovered appointments per month you already cover the extra monthly fee, but to pay off the full USD 6,240 within the first year you need between 8 and 11 per month: 8 if your appointment leaves USD 70, 11 if it leaves USD 50. If your practice does not believe it is losing a dozen appointments a month for not answering in time, quote B is not justified yet, and you can estimate that before signing by counting how many of last month's after-hours inquiries never came back.

On top of that come 12 to 18 monthly hours of reception time freed up, which is not money until the receptionist uses them on something that does not get done today: confirming the next day's appointments one by one, or calling the patients who left a treatment halfway through.

The same structure in a hair salon, where each recovered appointment leaves around USD 15 of margin and not USD 60, would need more than 35 appointments per month to justify quote B. That is why the same chatbot is correctly priced at USD 600 in one industry and at USD 4,200 in another: what changes is not the bot, it is your average ticket.

Seven questions that separate a solid quote from one that is going to fail

1. Where does the bot get its prices? If the answer is that they load them, every price list change is going to be a support ticket.
2. What does it answer when it does not know? If there is no written fallback answer, the model is going to improvise prices and deadlines.
3. Whose name are the WhatsApp Business Account and the number under? If they sit inside the provider's Business Manager, your number is a hostage: switching providers means a migration with Meta and the risk of days without service.
4. Whose name is the model's API key under? If it belongs to the provider, you cannot audit consumption or take anything with you.
5. Where does the workflow run and can I export it? n8n or Make in your account exports to a file. A closed platform gets rebuilt from scratch.
6. How many test questions run before production? If the answer is that they test it by chatting for a while, the first real test is run by your customer.
7. What happens if Meta suspends the number or rejects a template? There has to be a written protocol, not a verbal promise.

The first three go in the contract, not in an email. If they are in writing, switching providers next year costs one exported file and a partner change in Business Manager. If they are not, it costs rebuilding the entire bot and asking Meta to release your number.

Neither quote, moreover, usually includes what is going to reach you anyway: the conversations Meta bills, token consumption and hosting. That breakdown is in how much an AI WhatsApp chatbot costs.

When the USD 500 quote is the right one

There are scenarios where paying for the high tier is throwing money away, and it is worth saying so up front:

• Fewer than 80 or 100 inquiries per month. There is no volume to amortize any integration.
• Low average ticket with no repeat business. At USD 15 per sale, the math above never works out.
• A catalog or list that changes twice a year. Loading it by hand costs less than syncing it.
• You still do not know what people ask you. Before quoting anything, export 30 days of WhatsApp history and count the ten most repeated questions. Sometimes 70% is solved with a well-written welcome message and your opening hours loaded into your Google Business Profile.
• Your management system has no API. Many local ERPs and desktop calendars expose nothing. In that case the USD 5,000 chatbot simply cannot be built: you would pay for the integration and end up with manual data entry anyway. Ask your software provider whether it has a public API or webhooks before anything else.
• The process behind it is broken. If nobody confirms appointments today, a bot that books three times as many only multiplies the no-shows.

In the free 15-minute audit we look at exactly those three variables: real volume, average ticket and whether your systems have an API. If the math points to the low tier, we tell you that and we do not sell you integrations. If it points to the high tier, the Live in 72h model covers the standard stack with hosting and monitoring included, and the custom model is for when your management system does not expose an API and the sync has to be solved another way.

If you want to understand how what you are quoting is put together technically, the WhatsApp automation guide shows the full path. Write to us at info@striqtech.com with your monthly volume and the name of your management system, and we will tell you which tier you fall into before sending you any number.