Custom Chatbot Development

Custom chatbot development for support, sales and internal use. Built on your own content, with honest answers, escalation to a human, and no invented facts.

Book a call ↗
4.9 on Clutch, 6 reviews4 international awards since 2022Building from Budapest since 2021
Custom Chatbot Development

Most chatbot projects fail for the same two reasons

They are built on nothing, so they guess. And they have no exit, so a customer who needs a person ends up trapped in a loop and angrier than when they started.

Both are solvable, and neither is solved by picking a better model. The work is in the content underneath, the boundaries around it, and the handover when it runs out. That is where we spend the time.

What we build

Customer support. Trained on your help material, order data and past tickets. Answers the repetitive questions, escalates the rest with context. The realistic goal is a meaningful share of first-line volume, not all of it.

Sales qualification. Answers product questions, works out whether an enquiry fits what you do, and books the ones that do. Useful mainly because it works at nine on a Sunday evening.

Internal knowledge. The one that quietly saves the most time. New staff asking how a process works, and getting the current answer instead of a document from two years ago.

How it goes

We start by looking at what you have. Not a discovery workshop, an actual audit: which documentation is current, which contradicts itself, what your support inbox is really full of. A week of that tells us whether a chatbot is the right answer, and roughly a third of the time we conclude that a better help page or one fixed process would do more.

If it is the right answer, we build a narrow version first, on the highest-volume topics only, and put it in front of real users. Narrow and correct beats broad and unreliable, and it earns the trust that lets you widen it later.

The costs nobody mentions upfront

Language model usage is a running cost, and it scales with conversations. We will size it for your volume before you commit rather than after your first bill.

It also needs maintenance. Your prices change, your policies change, and a chatbot answering from last quarter's material is worse than no chatbot. We will either train your team to keep it current or do it on a retainer, and we will tell you which we think makes sense for you.

How it works

Grounded in your content, not the open internet
The chatbot answers from your documentation, product data, policies and past support tickets, and it cites what it used. That is the difference between a tool your team trusts and a novelty that invents a refund policy on a Tuesday. Getting your content into a usable state is part of the work and it is usually where the real effort goes.
It says when it does not know
The single most damaging thing a support bot does is answer confidently and wrongly. We build in the boundary explicitly: outside the material it has, it says so and hands over. That behaviour is tested with deliberately awkward questions before launch, including the ones your customers actually ask when they are annoyed.
Escalation that works
A conversation the bot cannot finish reaches a person with the transcript attached, in whatever your team already uses. Nobody is asked to repeat themselves. A bot with no exit is a customer service problem wearing a technology costume.
Where it lives is your decision
On the website, inside WhatsApp, in Slack or Teams for internal use, or behind your own login. We have built all of these. The channel changes the interface and the constraints far more than it changes the underlying work.
Measured against something
Before launch we agree what it is for: deflect a share of repetitive tickets, qualify enquiries out of hours, answer product questions before checkout. Then we measure that and show you the conversations it failed at, because those are the useful ones. A bot with no target is impossible to judge and impossible to improve.

Where we have done this

What is and is not included

Included

  • An audit of your existing content and support history first
  • A narrow first version on your highest-volume topics, in production
  • Escalation to a person, with the conversation history attached
  • Adversarial testing before launch, including hostile questions
  • A measurement target agreed in writing beforehand
  • Monthly review of failed conversations for the first quarter

Not included

  • Writing the documentation it answers from, unless quoted separately
  • Language model usage, which is billed to your own account
  • Human support staffing for escalated conversations
  • Voice or phone channels, which are a different build

What we build it with

  • Python
  • TypeScript
  • Next.js
  • OpenAI
  • Anthropic
  • PostgreSQL
  • pgvector

Questions

The things people ask first.

What does a custom chatbot cost to build?
Most of ours land between $3,000 and $12,000 depending on how much of your content needs preparing and how many systems it has to reach. The build is rarely the expensive part; getting documentation into a state a model can answer from usually is.
And what does it cost to run?
Language model usage is an ongoing cost that scales with conversations. We size it against your real volumes before you commit, not after your first bill. For most of our clients it is tens of dollars a month, not hundreds.
How do you stop it inventing answers?
It answers only from your material and cites what it used, and the boundary is explicit: outside that material it says so and hands over to a person. We test that with deliberately awkward questions before launch, including the ones customers ask when they are already annoyed.
Can it hand over to a human?
Yes, and a bot without that is a customer service problem wearing a technology costume. A conversation it cannot finish reaches a person with the transcript attached, in whatever your team already uses, so nobody is asked to repeat themselves.
Where can it live?
On the website, inside WhatsApp, in Slack or Teams for internal use, or behind your own login. We have built all of these. The channel changes the interface and the constraints more than it changes the underlying work.
How will we know whether it worked?
Because we agree the target before building: deflect a share of repetitive tickets, qualify enquiries out of hours, answer product questions before checkout. Then we measure that and show you the conversations it failed at, which are the useful ones.

Also in AI & automation

Tell us what you are trying to ship.

Start a conversation ↗