EU AI Act high risk obligations are now enforceable. Check your exposure
Insights About us Careers
Contact us
Conversational AI

AI Chatbot Development

A chat assistant on your site or in your product that answers from your own content, admits when it does not know, and hands over quickly rather than looping a customer through variations of the same non-answer.

6 to 10 weeks
Typical build
Fixed scope
Commercial model
Grounded
In your content

The reason people distrust chatbots is not that the technology is bad. It is that most deployments were scored on how many conversations were kept away from a human, and every design decision followed from that number. Build for resolution instead and the same technology produces something people actually use twice.

In one paragraph

AI chatbot development is the engineering of a conversational interface on your website, app or product that understands what a visitor is asking, answers from your own approved content with citations where useful, performs a limited set of permitted actions, and escalates to a person when the question falls outside what it should handle.

What separates a useful chatbot from an irritating one

  • It is grounded. Answers come from your documentation and product data, not from a model's general impression of your industry.
  • It refuses. When retrieval finds nothing relevant it says so and offers a route onward, rather than producing a confident paragraph that turns into a complaint.
  • It knows its scope. A small, well-covered set of topics answered properly beats broad coverage answered approximately.
  • It escalates fast. On request, on repeated failure and on any sign of frustration, with the conversation carried across so nobody repeats themselves.
  • It is measured honestly. Resolution and satisfaction, not sessions contained, because containment can always be improved by making escape harder.

None of that is difficult to build. It is difficult to agree, because it means accepting that the bot will hand over a meaningful share of conversations and that this is the system working correctly.

How we build a chatbot

Choose scope from real questions

We start from search logs, support tickets and live chat transcripts, and rank the questions by volume and by how answerable they are from content that exists. The first release covers the top of that list. Everything else is explicitly out of scope and routed, which is what makes quality provable in the first month.

Ground the answers

Content is indexed with structure preserved and retrieved per question, with the same retrieval discipline described under RAG system development. Where the same corpus also serves staff internally, it is usually worth building once and serving both, as covered by the enterprise knowledge assistant.

Add only the actions you can control

Order status, appointment slots, a simple update after identity verification. Each is a scoped tool with validation and logging. Where the chatbot needs to resolve cases end to end inside a helpdesk, with refunds and account changes, that is a larger build with different controls, described under AI customer support agents.

Write the failure paths first

What it says when it does not know, when it has misunderstood twice, when the user is angry, when the systems behind it are down. These sentences do more for the perceived quality of a chatbot than any improvement in the model, and they are usually written last, badly, by whoever is available.

Instrument from the first day

Resolution rate, escalation reasons, unanswered questions, satisfaction and the exact turn where conversations fail. The unanswered questions are the most valuable output: they are a live map of the gaps in your content, routed to whoever owns it.

Worth knowing

Containment is the wrong target

A chatbot can reach a very high containment rate by making it hard to reach a person, and every one of those contained conversations is a customer who gave up. We report containment, resolution and satisfaction together and treat containment rising while satisfaction falls as a regression, not a win.

Where chatbots pay for themselves

SettingWhat the chatbot handles wellWhat it should hand over
E-commerceDelivery timing, returns policy, sizing, stockPayment disputes, damaged goods, complaints
SaaS productHow-to questions, configuration, plan differencesBilling changes, cancellations, outages
Financial servicesProduct explanations, process steps, document listsAnything account-specific or advice-adjacent
Healthcare adminAppointments, preparation, location and accessAnything clinical, without exception
Public servicesEligibility explanations, form guidance, deadlinesCase-specific decisions and appeals

The second column is where the volume is. The third column is where the reputation risk is, and drawing that line clearly is the single most important design decision in the project.

Process

How the engagement runs

Scope comes from real questions, and the bot drafts before it answers anyone directly.

Weeks 1 to 2

Question analysis and scope

Search logs, tickets and transcripts analysed, topics ranked by volume and answerability, scope and escalation boundary agreed.

Weeks 3 to 5

Grounding and answers

Content indexed, retrieval tuned against a real question set, refusal and failure language written and reviewed.

Weeks 6 to 7

Interface and actions

Chat interface built into your site or product, permitted actions added with validation, escalation and handover wired in.

Weeks 8 to 9

Pilot

A limited share of live traffic with daily transcript review, unanswered questions routed to content owners.

Week 10

Rollout and handover

Wider release with dashboards, plus handover of the evaluation set, content gap process and runbooks.

Deliverables

What you receive

A chatbot measured on whether it helped, with the instrumentation to keep improving it.

01

Deployed chatbot

Integrated into your site or product, deployed as infrastructure as code.

02

Scope and escalation map

Topics in scope, topics that always hand over, and the triggers for escalation.

03

Grounded answer layer

Retrieval over your content with refusal behaviour and citation where useful.

04

Conversation copy

Greeting, refusal, misunderstanding, frustration and outage language, reviewed and versioned.

05

Evaluation set

Real questions with agreed answers, scored as a release gate on every change.

06

Quality dashboard

Resolution, escalation reasons, unanswered questions and satisfaction, against a pre-launch baseline.

Fit check

Is this the right engagement?

Worth being direct. AI Chatbot Development is the wrong spend in some situations, and those are listed rather than buried.

Good fit if

  • A small number of question types account for most of your inbound volume.
  • Documentation exists and is broadly accurate, or someone will own fixing it.
  • You will judge the bot on resolution and satisfaction rather than deflection.
  • There is a human channel to escalate into.
  • Someone can review transcripts daily during the pilot.

Choose something else if

  • The content is thin or contradictory. Fix the content first; grounding cannot invent it.
  • Every question is account-specific and needs deep system access. That is support agent territory.
  • The channel is really the phone. See voice bot and IVR modernisation.
  • The objective is to reduce contact volume regardless of whether customers got an answer.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

How is an AI chatbot different from the old rule-based kind?

A rule-based bot matches phrases to scripted replies and fails outside its tree. A grounded AI chatbot understands the question in context, retrieves the relevant passage from your own content and answers in its own words, with the ability to say it does not know. The scripting effort disappears; the content quality becomes the constraint.

What containment rate should we expect?

It depends entirely on your question mix and content quality, which we measure from your own logs in the first fortnight rather than quoting an industry figure. More usefully, we will not raise containment by making it harder to reach a person, so the number we report is one you can act on.

Will it make things up?

It is designed to refuse rather than infer when retrieval returns nothing relevant, and answers are grounded in your content rather than in the model's general knowledge. That reduces the risk substantially without eliminating it, which is why refusal language, scope limits and escalation triggers are part of the build.

Can it do more than answer questions?

Yes, within limits: order status, appointment booking, a simple update after identity verification. Anything that changes money or account state at scale needs the controls described under AI customer support agents, which is a larger engagement with a different risk profile.

How long does AI chatbot development take?

Six to ten weeks to a piloted release. Most of that is scope analysis, content grounding and writing the failure paths properly. The interface itself is the quick part, which is the opposite of how these projects are usually estimated.

Is this the right engagement?

Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.