A chatbot your support team will actually defend
Most chatbots get switched off six weeks after launch. The ones we build sit on top of your real help centre, your real tickets, and your real tone — and they pass the hard questions to a human instead of guessing. Web, WhatsApp, Slack, Intercom.
The chatbot most agencies sell vs. the one we build
There’s a chatbot most agencies will sell you. Drop it on the homepage, wire it to ChatGPT, ship in two weeks. It looks great in the demo. Your support team quietly turns it off six weeks later because it kept making up features that don’t exist.
The chatbots we build are different on purpose. They sit on top of your actual help centre, your actual past tickets, your actual product docs. They answer in your tone (yes, we test this on real conversations). And when they aren’t sure — which is honestly the hardest part to get right — they hand off to a human with the full conversation attached. So nobody starts from cold.
What ships on day one
These aren't paid add-ons. They're what we mean when we say "production-grade." If a vendor calls any of these "premium," walk away.
Answers grounded in your stuff
We index your docs, PDFs, Notion, Confluence, Zendesk threads, past tickets — anything that has the real answers. The bot can use those sources and only those sources. So when somebody asks about a feature, you get the version your product team actually shipped. Not the version ChatGPT imagined.
Hands off to humans the right way
Every answer gets a confidence score. Below the threshold (we'll help you set it), the bot stops talking and routes the conversation, with full context, to whoever's on duty. The customer doesn't repeat themselves. The agent isn't reading cold. This is the bit most chatbots get wrong.
Speaks the languages your users do
Visitor types in Spanish, bot replies in Spanish. Same for Arabic, Hindi, French, Mandarin and about 30 others. Detection is actually decent now (it used to be a coin flip). Particularly useful for UAE clients running Arabic-plus-English support and SaaS companies with multilingual user bases.
Private when it has to be
If your customers are in healthcare, fintech, or anything regulated, this stops being optional. We deploy to your AWS or Azure tenancy. Every interaction logged with timestamp and user. Anthropic/OpenAI enterprise endpoints with zero-retention if your security team is comfortable. Self-hosted Llama or Mistral if they're not.
A dashboard you'll actually open
Top intents, how often the bot handled it solo, where it failed and why. Two months in you'll know exactly which help-centre article needs rewriting, because the bot keeps getting that one question wrong and the logs will tell you which one.
Tuned weekly for the first 30 days
For the first month after launch we're in there every week. Looking at real conversations, tightening prompts, fixing the easy stuff. After that it's monthly check-ins on a retainer if you want it. Bots that don't get tuned drift. Tuned ones get better as your product and content evolve.
Where teams actually get value
Three patterns we keep seeing. Yours is probably close to one of these. We'll customise from there once we know which.
Tier-1 support deflection
- Handles roughly half to three-quarters of repeat questions on a good week
- The ones it can't handle, it tags by topic for your human agents — saves real time
- Customer-satisfaction scores on routine queries tend to land slightly above human baseline. Bot never has a bad day.
24/7 product Q&A + lead capture
- Answers the same five product questions your AEs are tired of repeating
- Qualifies leads — ICP fit, budget, urgency — before they ever hit anybody's calendar
- If the visitor's ready, books straight into Calendly or HubSpot. If they're not, captures email and intent tag for nurture.
Internal helpdesk (HR / IT)
- Lives inside your Slack or Teams workspace — no new app to learn
- Pulls from Confluence, Notion, SharePoint, wherever your knowledge actually lives
- Honestly the highest-ROI version we ship. People get answers in three seconds instead of hunting through Notion at 11pm.
Four weeks. No 6-month research phase.
Vendors who say "we need 8 weeks of discovery" are mostly billing for hesitation. We ship a working prototype on your real data by the end of week one.
Week 1 · Scope and ingest
We start with a 60-minute call. By the end of it we know what intents matter, who your angry customers are, where the documentation actually lives. By Friday you're looking at a working prototype trained on your real data.
Week 2 · Voice, guardrails, evals
The bot exists. Now we shape it. Voice (it should sound like your brand, not like a chatbot). Refusal behaviour (when does it admit it doesn't know?). Escalation thresholds. Plus we build an eval set from 50-100 questions you've actually been asked.
Week 3 · Plumbing and quality
Connections to your CRM, helpdesk, Slack. We run the eval set end-to-end and spot the five-to-ten places the bot's still weak. Each one gets fixed. This is the unglamorous week that decides whether launch goes smoothly.
Week 4 · Launch and tune
We never go zero-to-100% on day one. Typically 10% of traffic for two days, 50% for two days, then full. Analytics dashboard live the moment the first conversation happens. Then 30 days of post-launch tuning included at zero extra cost.
The questions every buyer eventually asks
These are the actual questions we get on every discovery call. Honest answers, including the disqualifying ones.
01 OK, what does this actually cost?
Pilots start at $12K. That buys 3-4 weeks of senior engineering, one channel (web widget OR Slack OR WhatsApp — not all three), one source of truth. Most clients move to a monthly retainer at $4-8K once they see what the bot is and isn’t doing. We send a real pricing sheet — not a “depends on requirements” email — before any contract.
02 Why not just bolt ChatGPT onto our site?
You can. People do. The problem is ChatGPT doesn’t know your pricing, your latest changelog, or that one weird flow your customers always get stuck on. Drop it on your website and it’ll confidently invent features you don’t have. Our bots are grounded — they can only answer from sources you’ve approved — and an eval suite catches drift before it goes live. Different category of thing.
03 OpenAI, Claude, or open-source?
All three, depending on what the query needs. Reasoning-heavy stuff routes to Claude (it’s strong at structured answers). High-volume classification — like “is this question about billing or shipping?” — goes to smaller cheaper models. Privacy-locked work runs on self-hosted Llama or Mistral. We pick per intent, not per vendor loyalty.
04 What stops it making stuff up?
Three things, in this order. One: it can only retrieve from sources you’ve approved, so it physically can’t answer from stuff you didn’t feed it. Two: every release runs through an eval suite of 50-100 real questions, and if anything regresses, the deploy stops. Three: when confidence drops below threshold, the bot hands off to a human instead of guessing. It’s not magic — bots still get things wrong sometimes. But the wrong answers don’t survive long.
05 Can we self-host?
Yes. We deploy to your AWS, Azure or GCP if your security team requires it. Or we use Anthropic / OpenAI enterprise endpoints with zero-retention contracts if your security team is comfortable. We don’t have a preference; we have a list of trade-offs we’ll walk you through.
06 What happens after launch — do you ghost us?
No, that’s the bit most agencies get wrong. Every project includes 30 days of post-launch tuning at zero extra cost. We’re in the conversations weekly, fixing the small stuff. After that 30, you can move to a monthly retainer (most clients do) or take it in-house — your call, and the choice doesn’t change how we treat you in those first 30 days.
What teams say after going live with AI chatbots
Tell us what your support inbox looks like today
Form takes two minutes. A senior — not a sales rep — replies within four working hours.