Case study · Customer Support
The honest limits of a support bot
A support bot earns its place by answering the same four questions perfectly and knowing exactly when to stop.
SupportGenie
Customer Support · 7–14 days · ₹15,000–₹40,000
Go and use the buildScope it to the business's own answers, not to open conversation
The bot answers from the business's real FAQ content.
Human handover is mandatory and always one tap away
'Talk to a person' is a permanent option, not a fallback the bot offers after failing twice.
The situation
What goes wrong here
Most customer messages a small business receives are the same handful of questions: what does it cost, are you open, where are you, can I book. Answering them by hand consumes hours a week and, worse, happens slowly — the message that arrives at eleven at night is answered at ten the next morning, by which point the customer has asked someone else.
The failure mode of the obvious fix is well documented. A bot given free rein over an open-ended model will, sooner or later, state a price that does not exist or promise something the business does not offer. For a small business that is not an amusing screenshot; it is a commitment a customer will hold them to.
The decisions
Every one of these had a cost
A decision with no trade-off is a feature list. Each of these names what was given up, because that is the part a client needs to agree with before the project starts — not after.
- 01
Scope it to the business's own answers, not to open conversation
The bot answers from the business's real FAQ content. It is not there to be charming about topics outside that set, because every sentence it improvises is a sentence the owner did not approve and may have to honour.
What it cost
It will decline questions a more open system would have attempted. That is the feature: a bot that says 'let me get someone' costs nothing, and a bot that invents a discount costs a customer.
- 02
Human handover is mandatory and always one tap away
'Talk to a person' is a permanent option, not a fallback the bot offers after failing twice. Anyone who wants a human gets one immediately, and the conversation history goes with them so nothing has to be repeated.
What it cost
Some conversations a bot could have closed get escalated. Cheaper than the alternative, every time.
- 03
Capture the lead inside the conversation
When the bot cannot answer, it takes a name and number before ending. Otherwise the most valuable interaction — the one with a question specific enough to need a human — is the one that leaves no trace.
What it cost
One more thing asked of the visitor at a slightly awkward moment, which has to be phrased as a way to get them an answer rather than as a form.
- 04
Decide what is logged before launch, not after
A support bot receives phone numbers, order details and occasionally things people should not have typed. What is stored, for how long, and who can read it is a decision made at the start — it is far harder to unwind once a year of transcripts exists.
What it cost
A conversation the client would rather skip during an exciting project. It is the one that prevents a genuinely bad day later.
In the build
- Answers drawn from the business's own FAQ content
- Website and WhatsApp channels
- One-tap human handover with conversation history
- Lead capture on unanswered questions
- Conversation analytics
Built with
- Claude API
- Node.js
- WhatsApp Business API
You provide
- Common questions with the answers you want given
- Business information and policies
- WhatsApp number
- Who handovers should reach
Nothing starts until the scope, timeline and price are agreed in writing.
Measured
How this build actually performs
Taken from the finished build under throttled conditions, not from a score card. All six concept builds share a foundation, so they share these numbers.
Chromium at 412×915, 4× CPU throttling, 1.6 Mbps with 150 ms latency, cold cache — a mid-range Android phone on a typical Indian mobile network, not a developer's laptop. Includes the analytics the live site loads. Measured 2026-09-08.
- Largest contentful paint
- 1.58–2.16s
- When the main content is actually on screen. Google treats anything under 2.5s as good.
- Cumulative layout shift
- 0
- Nothing moves after it appears — no mis-taps caused by a button jumping.
- JavaScript over the wire
- 429 kB
- Compressed, as actually sent — 259 kB of site code plus 170 kB of Google Analytics.
- Total page weight
- ~613 kB
- Everything: markup, styles, fonts, scripts and analytics.
Honestly
What this is and isn’t
The public demo is deliberately not an AI. It matches keywords against a small set of prepared replies, because a demo left open to the internet with a live model attached is a demo someone will make say something regrettable, and the running cost of an unmetered public model is not sensible for a portfolio piece. What the demo genuinely shows is the interaction design — the scope, the quick replies, the handover, the lead capture — which is the part that decides whether the bot works. A delivered project connects real language understanding behind exactly that structure.