An AI chatbot development company that builds bots your customers do not try to escape
Grounded in your actual content, honest when it does not know, and wired to hand off to a human before anyone gets annoyed. Deployed on your site, WhatsApp, or wherever your customers already are.
- Grounded in your content; every answer cites its source
- Human handoff stays easy to reach, on purpose
- You own the bot, the content index, and the evals
Why this is worth building, and why us.
Everyone has met the bad version. It answers a question you did not ask, loops back to the same menu, and hides the path to a human behind four taps. It exists because someone bought a chatbot to deflect tickets rather than to answer questions, and the metric it optimised was containment rather than resolution.
We build the other kind. A chatbot grounded in your real documentation that cites what it used, says "I do not know" instead of inventing a refund policy, and offers a human the moment it is out of its depth. Deflection then happens as a side effect of being useful, which is the only way it is worth having.
What we build into every bot
- 01Retrieval over your real content: help docs, policies, product data, kept in sync
- 02Citations, so a customer and your support team can both check the answer
- 03A refusal path: it says it does not know rather than guessing at your policy
- 04Human handoff with the full conversation attached, into the inbox you already use
- 05Channel deployment: website widget, WhatsApp, or embedded in your product
- 06Analytics on what people actually ask, which is usually a roadmap in disguise
Two tiers and a cheap way to test it
A chatbot is easy to scope because the variable is your content, not your imagination. So the pricing can be honest.
Before you buy a bot, find out whether your documentation can actually answer customers. We index a slice of your real content, test it against real questions, and give you a coverage number plus the gap list. Sometimes the finding is that you need eight help articles, not a chatbot. Credited back in full against any build you start within 30 days.
- Starter Bot$1,4971 week
The twenty questions your team answers every day, on one channel.
- Retrieval over your help content, with citations in every answer
- A refusal path: it says it does not know rather than guessing
- Website widget styled to your brand
- Email handoff when it cannot help
- Built on the same foundation as the bigger tiers, so it upgrades cleanly
- Most chosenSupport Bot$3,4972–3 weeks
A bot doing real deflection, with the measurement to prove it.
- Everything in Starter Bot, plus WhatsApp or a second channel
- Live human handoff carrying the full transcript into your inbox
- An eval set built from your real support history
- Cost caps and rate limits so traffic spikes cannot surprise you
- A dashboard of questions asked, answered, and escalated
- Tuning against the first two weeks of real conversations
- Multi-Channel Bot$7,9974–6 weeks
Bots that take actions, not just answer: bookings, orders, account changes.
- Everything in Support Bot, across multiple channels with shared memory
- Tool-calling: look up an order, book a slot, update a record
- Authentication, so it can safely discuss a specific customer’s account
- CRM and helpdesk integration, both directions
- Permissions-aware retrieval: customers see only what they should
- Thirty days of post-launch tuning included
- Grounded in your content; every answer cites its source
- Human handoff stays easy to reach, on purpose
- You own the bot, the content index, and the evals
- Inference billed to your provider, never marked up
The honest caveat: a chatbot cannot answer what your documentation does not say. If the Content Check finds real gaps, fixing them is either your work or a separate scope, and we would rather tell you that for $297 than discover it in week three.
Containment is the wrong metric
Chatbot vendors sell deflection rate because it is easy to move: make the human option hard to find and containment goes up while satisfaction quietly collapses. The customer still had the problem. They just gave up, or found your Twitter.
The metrics worth tracking are resolution rate, escalation quality, and whether people come back. We instrument all three, and we make the human handoff easy on purpose. A bot that hands off cleanly at the right moment beats one that traps people, on every measure a business actually cares about.
Your content is the real constraint
A chatbot is a retrieval problem wearing a conversational costume. It can only be as good as what it can find, and most teams overestimate their documentation until they watch a bot fail on the twenty questions customers ask most.
That is why the cheapest thing we sell is a content check. One day, your real content, your real questions, and an honest coverage number. Sometimes the finding is that you do not need a chatbot yet; you need eight missing help articles, and writing those will deflect more tickets than any bot would have.
Chatbot or agent?
A chatbot answers. An agent acts: it calls tools, changes state, and makes decisions with consequences. The line matters because the engineering, the risk, and the price are all different. Answering "what is your refund window" is retrieval. Actually issuing the refund is an agent with authentication, permissions, and an audit trail.
Our Support Bot tier is the first thing; the Multi-Channel tier crosses into the second. If your requirement is mostly actions, the agent page is the better starting point and we will point you there rather than stretching a chatbot to fit.
Not claims. Shipped work.
- 01ChatberryAn AI WhatsApp marketing platform: conversational AI shipped as a real product.See the work
- 02AI agent developmentWhen the requirement is taking actions rather than answering questions.See the work
- 03Why we ship eval suites with every agentHow we measure whether a bot is actually getting better.See the work
Common questions.
01How much does AI chatbot development cost?
Ours are published: $1,497 for a starter bot on one channel in a week, $3,497 for a measured support bot, $7,997 when it needs to take actions rather than just answer. The $297 Content Check comes first and tells you whether your documentation can actually support a bot, which is the question that determines whether any of the rest is worth spending.
02How do you stop it giving wrong answers?
Answers are grounded in your retrieved content with citations, so nothing is generated from the model’s general knowledge about your business. Where retrieval finds nothing relevant, the bot says so and offers a human rather than improvising. An eval set built from your real support history tracks this over time.
03Can it hand off to a human?
Yes, and we treat that as a feature rather than a failure. The handoff carries the full transcript into whatever inbox your team already uses, so the customer does not repeat themselves. We deliberately keep the human option easy to reach, because hiding it is how bots earn their reputation.
04Which channels can you deploy to?
Website widget and WhatsApp are the common two, and the Multi-Channel tier adds more with shared memory across them so a conversation survives switching. Embedding inside your own product is also straightforward. The retrieval and eval work is shared; the channel is mostly a delivery detail.
05What if our documentation is a mess?
Then the Content Check will say so for $297, which is the cheapest possible way to learn it. Sometimes the right next step is writing the missing articles rather than building a bot, and we will tell you that even though it means not selling you a build. A bot on thin content just fails more articulately.
Find out what your content can actually answer
Point us at your help docs and the questions your team answers most. The $297 Content Check gives you a real coverage number in a day, before you commit to building anything.
- A real person replies, usually same day
- No sales sequence, no automated nurture
- Content Check is credited back if you go ahead