AI chatbots that know when to hand over.
We build customer service chatbots and internal assistants that answer from your own content, refuse what they should not touch, and pass difficult conversations to a person with the context attached. Our AI-native team gets a working first version in front of users in weeks.
Customer chatbots and internal assistants are different projects
Both run on the same underlying models, but the audience changes almost every design decision. We scope them separately, even when a client wants both.
Customer service chatbot
Answers product, order and policy questions on your website, app or messaging channels. Tone, refusals and handover matter more than breadth, and anonymous visitors never see account details without verification.
Internal assistant
Helps staff find policies, draft replies or query internal systems from Slack or Teams. It can go deeper than a public bot, but must respect who is allowed to see what.
Helpdesk and CRM handover
Escalations land in the tools your agents already use, with the transcript, detected issue and customer record attached, so nobody asks the customer to start again.
Analytics that tell the truth
Resolution, escalation and repeat-contact rates by topic, so you can see where the bot genuinely helps and where it only delays a human.
Built by an AI-native team
Integration code for your helpdesk, CRM and order systems is largely drafted by AI coding agents and reviewed by senior engineers, so budget shifts towards conversation quality.
What a good AI chatbot refuses to do
The quality of a chatbot shows most in the conversations it declines. We agree these boundaries with you before a single prompt is written.
Invent a policy, price or promise
If the answer is not in approved content, it says so and offers a route to someone who knows. Grounding on your knowledge base keeps this in check.
Discuss an account before verifying the user
Order status and personal data sit behind login or a verification step, and lookups use read-only access scoped to that one customer.
Obey instructions typed into the chat
People will try "ignore your previous instructions". Limited tool permissions and output checks stop a clever message turning into a data leak.
Keep a frustrated customer in a loop
Complaints, repeated rephrasing, legal threats and signs of a vulnerable customer trigger handover rather than another scripted apology.
Give regulated advice
Medical, financial or legal questions get a pre-approved response and a pointer to a qualified person.
Wander off topic
A bot that writes poems on request is amusing once and costs tokens every time. Scope stays tight to your business.
Scripted bot, helpdesk AI add-on or custom AI chatbot?
Plenty of businesses do not need a custom build. Here is how the realistic options compare.
| Scripted rule-based bot | AI add-on in your helpdesk | Custom AI chatbot | |
|---|---|---|---|
| Copes with unexpected wording | Poorly | Well | Well |
| Reaches your systems and data | Through fixed flows | Within that vendor ecosystem | Any system with an API |
| Control over answers and refusals | Total, but rigid | Limited settings | Full, and testable |
| Running cost | Flat licence | Per seat or per resolution | Model usage plus hosting |
| Best for | A few fixed tasks | Standard support on one platform | Complex products, several systems, internal use |
If the AI features already in your helpdesk cover your needs, we will recommend switching them on and help you configure and test them properly.
How we build and launch an AI chatbot
Mine real conversations
Week 1We analyse past tickets, chat logs and emails to find the questions that matter by volume and value, and turn a few hundred of them into a test set.
Build the first version
Weeks 2 to 4Grounding on your content, account lookups where needed, handover into your helpdesk, and an admin view for reviewing transcripts.
Replay the test set
Weeks 4 to 5Whenever prompts or models change, historical questions are replayed and scored for accuracy, tone and correct escalation, with people reviewing borderline cases.
Shadow, then soft launch
Weeks 5 to 8The bot first drafts answers for your agents to approve, then goes live to a slice of traffic while we watch escalations and cost per conversation.
What an AI chatbot really costs to run
There are two bills. The build is a one-off project. Running costs depend on the number of conversations, how long they last and how much context goes into each reply. A support bot that pulls several help articles into every answer can use far more tokens than the visible text suggests.
We design for cost from the start: a small, inexpensive model for routing and simple questions, a stronger one only when needed, caching for the fixed parts of each prompt, and trimmed conversation history. Our AI chatbot cost calculator gives a monthly estimate before you commit.
Be wary of judging success on "deflection" alone. A bot can deflect a conversation simply by making a person hard to reach. We track whether customers stopped needing help, not just whether they stopped asking.
Questions we often hear
How much does it cost to build an AI chatbot?
A focused chatbot on one channel, grounded on your content with helpdesk handover, typically fits a Product Build of a few weeks, while assistants spanning several systems take longer. Running costs depend on conversation volume and model choice. Our AI chatbot cost calculator estimates the monthly figure before you build.
How do you stop an AI chatbot from making things up?
We ground answers in your approved content, instruct the bot to say when it does not know, and test it against hundreds of real historical questions before launch. For anything involving money, policy exceptions or personal data, it hands over to a person rather than improvising.
Can an AI chatbot hand conversations over to a human agent?
Yes, and it should. We integrate with common helpdesk and CRM tools so the agent receives the full transcript, the detected issue and customer details. Handover triggers include low confidence, frustration, sensitive topics and a direct request for a person.
Should we just use ChatGPT instead of a custom chatbot?
For staff drafting and research, an enterprise AI assistant with a clear usage policy is often the right answer. A general assistant cannot see your order system, follow your refund rules or escalate to your support team, which is where a custom chatbot earns its cost.
Do you use AI to build the chatbot itself?
Yes. AI writes much of the routine code and role-plays awkward, impatient or hostile customers so we can stress-test the bot before launch. What stays human: deciding what the bot may say, reviewing every change that ships, and signing off the escalation rules.
Related reading and tools
Tell us who the chatbot will talk to.
A few lines on the channel, the questions it should handle and the systems behind them is enough to start. You will hear back within 24 hours on whether a custom build, an off-the-shelf add-on or neither makes sense.
Working with companies globally · Response within 24 hours