Will this replace my support team?

No, and we would rather say so before you buy than after. It takes the repeat questions — delivery times, returns, opening hours, whether something is in stock — the ones your team answers over and over in slightly different words. The rest still needs a person, and the bot's job at that point is to get out of the way quickly rather than keep trying.

What happens when it does not know the answer?

It says it does not know and passes the conversation to a human. That is a written rule in the system prompt, tested against a set of questions designed to push it past its limits. A bot that guesses convincingly is worse than no bot, because a wrong answer about a refund or a delivery date reads exactly like a right one.

What does it cost to run each month, honestly?

It depends almost entirely on conversation volume, and the $150/mo in the price above is a low-volume figure — a few hundred conversations a month, not a few thousand. Above that, running cost is billed on actual usage with the report attached to the invoice. We show you the arithmetic before you sign, not after the first bill.

Can it look up a real order rather than quote the policy?

Yes, if your order system has an API we can call. That is a real addition rather than a setting though — it adds roughly two weeks to the build and moves the quote up, because every live lookup needs its own error handling for the case where the order is not found or the system is down.

Can it run on WhatsApp as well as the website?

Yes. Be aware that WhatsApp Business API access has its own approval process on Meta's side, with its own timetable that nobody can compress. We start that application on day one for exactly that reason, so it runs alongside the build instead of after it.

What stops it running up an enormous bill overnight?

Four locks, all fitted before launch — a monthly spend cap on the API account with an alert at half of it, a limit of around thirty messages per conversation, a per-IP rate limit of about ten requests a minute, and a hard output limit on every single call. Without those, one script pointed at your chat endpoint can generate months of normal usage in a night.

AI & Machine Learning

A support bot that answers the repeat questions

A chat widget on your site or a bot on WhatsApp that handles tier-one questions from your own policies, cites where the answer came from, and hands everything else to a person.

From: $800–$8,000 + $150/mo
Typical time: 2–5 weeks
Usually bought by: Any business

The three problems this is actually bought for

Almost every support inbox has the same shape. The same handful of questions keeps arriving in slightly different words, and salaried people are answering each one by hand. Nobody answers anything at night or over the weekend, so the person who was ready to buy at 10pm goes somewhere else. And on the big days — a sale, a holiday weekend — the team is buried precisely when the traffic is worth the most.

The case for buying is arithmetic rather than enthusiasm. A support hire costs what a support hire costs in your market, every month, forever. If a bot absorbs a meaningful share of the repeat questions, that is one fewer person needed on the rota, or the same people finally doing the work that actually needs judgement. What it is not is a replacement for your team, and any supplier who tells you it is has not run one in production. The honest version is that it handles the boring majority well and knows how to stop.

How the build runs

Collect the questions and find where the answers live. We gather the top thirty questions you actually get and trace each one to its source — a policy page, a spreadsheet, or somebody's memory. That last category is normal, and it is why the next step exists. Two to four days.
Build the handover route before the bot. Where does a conversation go when the bot should not be handling it — an email, a ticket, a live chat queue? This is built and tested first, deliberately. It is the part most projects leave until the end and then rush. One day.
Assemble the knowledge base and the retrieval. Your policies are split into passages, indexed, and searched two ways at once — by meaning, so a customer asking in their own words still finds the right paragraph, and by exact keyword, so a product code or a specific term is matched literally. Where your documentation has gaps, we write what is missing and you approve it. Three to five days.
Write the system prompt and its limits. Tone, answer length, the exact sentence it uses when it does not know, the list of situations that trigger an immediate handover, and the things it must never say — no promising a refund, no giving a delivery date that did not come from a live lookup, no discounts, no legal opinions. This is where the quality of the thing is decided, not in the code. Two to three days.
Connect live lookups, where they are in scope. Order status, stock, booking availability. Each one gets its own handling for the case where nothing is found or the system is unreachable. Three to five days per integration.
Test against a written question set, then fit the locks and launch. We run a fixed set of real and deliberately awkward questions, record what came back, and fix what fails. Then the spend cap, per-session message limit, per-IP rate limit and per-call output limit go on, and the widget goes live. Three to five days.

Tools, so you know what you are getting rather than a black box — the Anthropic or OpenAI API, Postgres with pgvector for the knowledge index, a chat widget built with the Vercel AI SDK or written directly, Langfuse for logging every call, and Twilio or 360dialog where WhatsApp is involved. Every call is logged with its full prompt and its cost, which is what makes a complaint two weeks after launch a ten-minute job instead of a two-day one.

What changes the quote

How many kinds of question it has to cover, and whether any need live data. A bot answering from policies is a different build from one that reads your order system. Each live integration is roughly two extra weeks and a step up in price, because the failure cases multiply.

Languages. Every additional language is a second knowledge base, a second set of tests and a second round of tone decisions. It is never a checkbox, whatever the platform demonstrations imply.

WhatsApp. The channel itself is straightforward. The Business API approval is an external process with its own clock, and it is the single most common reason a chatbot project slips past its date.

Where these go wrong

The handover route gets built last, or not at all. All the attention goes into how much the bot can answer, and nobody decides what happens when it cannot. The result is a customer arguing with a machine in a loop and then leaving. A bot that escalates smoothly is worth more than one that answers a wider slice of questions on its own, and that is not a close call.

A flat all-inclusive monthly price with no usage ceiling written into it. Then the Christmas sale arrives, traffic triples, and the month your bot is doing the most good is the month it costs the most to run. The fix is a stated included volume, a per-conversation rate above it, and a notification before the ceiling is reached — so you hear about it before the invoice does.

Anything with a date in it gets typed into the bot's instructions instead of fed to it. Somebody asks for "tell customers the sale ends this Friday", the sentence goes into the system prompt, and weeks later the widget is still telling people the sale ends this Friday. The bot has no clock and no sense of when now is — it reads that line fresh at the start of every conversation and repeats it faithfully, to every customer, until somebody notices. Anything time-bound goes in as a real date, passed in by your server on each call along with today's date, so the bot can work out for itself that the sale is over.

See where the site stands before you brief anyone.

If the site is already live, start from what it scores rather than from a guess. The audit reports what it checked, what it found, and — the part most tools leave out — what it could not check, and why.

See the website audit Talk about this work

Questions people ask before they commit.

Will this replace my support team?: No, and we would rather say so before you buy than after. It takes the repeat questions — delivery times, returns, opening hours, whether something is in stock — the ones your team answers over and over in slightly different words. The rest still needs a person, and the bot's job at that point is to get out of the way quickly rather than keep trying.
What happens when it does not know the answer?: It says it does not know and passes the conversation to a human. That is a written rule in the system prompt, tested against a set of questions designed to push it past its limits. A bot that guesses convincingly is worse than no bot, because a wrong answer about a refund or a delivery date reads exactly like a right one.
What does it cost to run each month, honestly?: It depends almost entirely on conversation volume, and the $150/mo in the price above is a low-volume figure — a few hundred conversations a month, not a few thousand. Above that, running cost is billed on actual usage with the report attached to the invoice. We show you the arithmetic before you sign, not after the first bill.
Can it look up a real order rather than quote the policy?: Yes, if your order system has an API we can call. That is a real addition rather than a setting though — it adds roughly two weeks to the build and moves the quote up, because every live lookup needs its own error handling for the case where the order is not found or the system is down.
Can it run on WhatsApp as well as the website?: Yes. Be aware that WhatsApp Business API access has its own approval process on Meta's side, with its own timetable that nobody can compress. We start that application on day one for exactly that reason, so it runs alongside the build instead of after it.
What stops it running up an enormous bill overnight?: Four locks, all fitted before launch — a monthly spend cap on the API account with an alert at half of it, a limit of around thirty messages per conversation, a per-IP rate limit of about ten requests a minute, and a hard output limit on every single call. Without those, one script pointed at your chat endpoint can generate months of normal usage in a night.

Often taken together.

Start with the brief, not the invoice.

Send what you are trying to achieve. You get a written scope, a price and a date — or an honest “this is not the right service for that”.

Talk about AI customer-support chatbot All AI & Machine Learning services