Chatbots that answer from your content and admit when they cannot
An AI chatbot answers questions in conversation. Built properly it draws answers from your own content rather than the model's general knowledge, declines questions outside that scope, and hands over to a person quickly when it should.
The failure that matters with chatbots is not being unhelpful. It is being confidently wrong, answering a question about your refund policy with something plausible that you do not actually offer, in your brand's voice, to a customer who believes it. A bot that says it does not know is a mild disappointment. One that invents an answer is a commitment somebody may hold you to.
Bots deployed on general model knowledge answer everything, including things they have no basis for. Without grounding in your content, the response reflects the model's training rather than your policies. Without a scope boundary, it will happily discuss subjects with nothing to do with your business. Without an easy escalation path, a customer who is not getting anywhere becomes a customer who is angry. And without conversation logging, nobody knows what it has been telling people.
When this work makes sense
Questions worth asking any supplier
Are answers grounded in your content, or generated from general knowledge?
What happens outside scope. Does it decline, or attempt an answer anyway?
How quickly can someone reach a person? Buried escalation makes things worse, not better.
Is the conversation logged? You need to see what it has been telling customers.
Who updates the knowledge base? Content drifts, and a bot answering from stale content confidently is worse than no bot.
Deliverables
Retrieval over your own documentation, policies and product content
A defined scope with refusal behaviour outside it
Escalation to a person, visible and reachable within one step
Channel integration, website, WhatsApp, or your helpdesk
Conversation logging with review tooling
An evaluation set of real questions with approved answers
A process for keeping the knowledge base current
Handover context so the person receiving an escalation sees the conversation
Built with
Process
The knowledge base is assembled before any bot is built, and it is usually the largest part of the work. Most businesses discover their documentation is incomplete, contradictory in places, or written for staff rather than customers. That has to be resolved first, because a retrieval system built over inconsistent content produces inconsistent answers, and the bot gets blamed for a content problem.
The standard nine stages, discovery, strategy, UX/UI, development, content and SEO, testing, launch, measurement and continuous improvement, apply to every project. The paragraph above is what differs for this one.
What moves the number
Pricing depends on scope, functionality, integrations and content requirements. These are the factors that change it most:
Rule-based bot, retrieval chatbot, or an off-the-shelf tool
Rule-based / decision tree
Fits: Narrow, stable processes where every path can be scripted
Limits: Rigid; anything unanticipated becomes a dead end
Retrieval-grounded AI chatbot
Fits: Businesses with real documentation and varied questions
Limits: Needs that documentation to be in good order first
Off-the-shelf chat tool
Fits: Common support scenarios with standard workflows
Limits: Less control over tone, accuracy and data handling
Case studies
No published chatbot case study. Deflection rate figures are easy to quote and rarely defined consistently.
MBGDesk operates from 4 offices in India, United States, United Kingdom, United Arab Emirates. We hold no industry certifications and do not imply otherwise.
Frequently asked questions
Will it make things up?
That is the central design problem rather than an edge case. Grounding answers in your own content, restricting scope, and building refusal behaviour reduce it substantially. No approach eliminates it entirely, which is why logging and review matter and why anything high-stakes should route to a person.
Can it replace our support team?
In our experience it changes what they work on rather than removing them. Routine questions get handled; the cases that actually need judgement still need people, and a deployment with no human escalation produces angrier customers rather than fewer.
Can it work on WhatsApp?
Yes, through the WhatsApp Business Platform. In India particularly it is frequently the main channel through which customers make contact, so a bot only on the website misses most of the volume.
How do we keep it accurate as things change?
Someone must own the knowledge base. The bot answers from what it has; if your refund policy changes and the document does not, it will confidently quote the old one. That ownership should be assigned before launch rather than discovered afterwards.
Can it look up a customer's order?
Yes, with integration into your systems and authentication so it only reveals a customer's own data. That is meaningfully more work than answering from documentation, and it is where most of the value usually sits.
What about data privacy?
Depends on the provider and the tier, some train on submitted data by default. We establish what the terms actually say before choosing, which matters when customers may paste personal information into a chat window.
How do we measure whether it works?
Resolution without escalation, escalation rate, and customer satisfaction on bot conversations specifically. Conversation volume alone is a vanity measure, a bot handling many conversations badly looks busy.
What happens when a customer is angry?
The bot should recognise it and get out of the way. Frustration signals, repeated rephrasing, explicit complaints, requests for a human, are detectable, and the right response is immediate escalation with the conversation passed across, not another attempt at an answer. The worst outcome in customer support is a bot cheerfully continuing while someone becomes progressively more annoyed, and it is entirely avoidable.
Services that connect to this
Talk to us about this
Tell us what you are trying to build and we will come back with scope, approach and an estimate.