Not your designer’s landing page. Not your founder’s LinkedIn post. A chatbot. Think about that for a moment.
Service Custom AI chatbot development · brand-safe + cited · UK
Custom AI chatbot that earns trust. not one customers tap past.
Custom AI chatbot development for UK SaaS, brand, and ed-tech teams. We build bots that cite their sources, sound like you, escalate cleanly when they don’t know, and never hallucinate a refund policy.
24hreply, from a senior
200+projects shipped since 2019
Senioronly, on the spine
(Why brands sign for custom AI chatbot development)
Chatbots that cite, escalate, and never invent. Twenty-one production deployments. Zero hallucinated policies.
A chatbot is often the last interaction a customer has with your brand. We make it one they trust, not one they tap past to find a human.
0
0%
0
The founder this page is foryears of brand work · met at 11pm by a bot
The last time a customer met your brand, it might have been a chatbot.
Years of brand work, a design system you sweated over, a tone-of-voice doc with rules about Oxford commas, and the part of you a real customer met yesterday at 11pm was a generic chirpy bot trained on someone else’s data, that invented a refund window and lost you the renewal.
This page is for founders who’ve noticed the gap. The next sections show what we replace it with.
The chatbot graveyard.
Three real ways trust dies.
Open any row to see how a generic bot loses the renewal.
Customer: “My refund window, is it 14 or 30 days?” Bot: “Refunds are available within 30 days for any reason!” Actual policy: 14 days. The bot made up an answer. The customer escalates. The brand eats the loss and the trust.
Customer: “Can I talk to a human please?” Bot: “I’m here to help! What’s your question?” The bot has no escalation path. Customer rage-quits. Goes to Twitter. Two weeks later, a refund request.
Customer: “Tell me about your enterprise pricing.” Bot: “Hey there! Great question! Our pricing is super customisable! Want to chat?” The brand is a B2B fintech serving FTSE-100. The bot sounds like a teenager. The decision-maker closes the tab.
What stops the graveyard pattern
Four layers of trust
Each one is the reason your bot stays trusted. The knowledge base is the truth, the voice profile is the brand, the citation gate is the guardrail, the escalation is the safety net.
Layer 01
Knowledge base
Your real help docs, past tickets, policies. Indexed in Pinecone. Re-synced via webhook.
Layer 02
Voice profile
Locked tone of voice + banned words + sentence patterns. Editable without a deploy.
Layer 03
Citation gate
No citation? No answer. Forces RAG-grounded responses every time.
Layer 04
Escalation
Confidence threshold + intent detection. Human handoff with full chain attached.
WHY BRANDS SIGN
Four things off-the-shelf bots can’t hand you
What turns a chatbot from a brand liability into a surface customers actually use.
Citation on every answer
No citation, no answer. Forces RAG-grounded responses. Hallucinations become architecturally impossible.
Locked brand voice
Tone-of-voice profile + banned words + sentence patterns. Editable by marketing without a deploy.
Clean escalation
Confidence threshold + intent detection. Human handoff with the full chain attached. No rage-quits.
Action-taking, audited
The bot can downgrade plans, refund within policy, schedule meetings. Every action logged with actor, args, result.
The pain. The day-1 architecture answer.
Each one is the difference between a chatbot customers trust and one they tap past.
- 01
“Our chatbot promised a customer a refund window we don’t actually offer.”
Citation gate + RAG over real policy docs. No citation, no answer. Bot pulls from your real policy KB. Refund answers cite the actual paragraph. Wrong answers are architecturally impossible.
- 02
“Our bot sounds nothing like our brand. Customers can tell.”
Locked voice profile + banned words list. Tone-of-voice doc encoded as system prompt. Banned words enforced. Editable by marketing without a deploy.
- 03
“Customer wants a human. Bot won’t hand off. Customer goes to Twitter.”
Intent detection + immediate human handoff. “Talk to human” is detected as intent. Instant handoff with full chain attached. The agent picks up where the bot left off.
- 04
“Bot quotes our 2022 pricing because the docs weren’t synced.”
Webhook-driven KB resync + freshness windowing. Your CMS updates trigger a KB resync. Old answers age out. Refund policy can’t be 2022 vintage by week three.
- 05
“Customer asked about their account. Bot showed another customer’s data.”
Tenant-scoped retrieval + RLS. Every retrieval is filtered by tenant ID. Bot can’t see another customer’s data even if it tries. Tested in CI on every release.
- 06
“Bot deflected 60% of tickets in month one. Now it’s 30%.”
Eval gold set + drift monitoring. Production conversations sampled. Failure modes go into the gold set. Drift in deflection rate alerts the team.
- 07
“Customer asks: are you AI? Are you recording this? Bot dodges.”
Honest disclosure + GDPR-aware design. Bot identifies as AI on first message. Privacy notice linked. Customer can opt to human at any time. GDPR-aware from day one.
- 08
“Bot cost £800 the first month. The second, £2,400.”
Prompt caching + model routing + per-tenant budget. Claude prompt caching cuts cost 60-90% on RAG. Haiku for routing. Sonnet for tone. Cost dashboard per tenant. Anomaly alerts at £X/day.
WHERE THE CHATBOT LIVES
Five surfaces, one shared brain
Same knowledge base, same voice, same trust layer, wherever your customers are.
Website widget
Embedded bubble or full-page surface. Custom-themed to your brand. Loads in <100ms.
In-product assistant
Lives inside your SaaS, reads the user’s screen + account state, can take in-app actions.
WhatsApp Business
Full WhatsApp Cloud API. Same brain, voice, trust layer. Right channel for retail + hospitality.
Slack / Microsoft Teams
Internal-facing assistant. Answers from runbooks, postmortems. Onboards new hires in week one.
Voice + IVR (optional)
Real-time voice for inbound calls. GPT-5 Realtime or ElevenLabs. Same KB, escalation, audit log.
One brain behind all five
A single knowledge base, voice profile, and trust layer. Update once, every surface follows.
From CSAT 3.8 to 4.6 in nine weeks
A B2C ed-tech, 90K users, 200 tickets a day, customers rage-pressed past the old bot.
We rebuilt as Claude Sonnet + RAG over their real help centre + Stripe state + LMS API. Every answer cites the policy paragraph. Below 80% confidence the ticket routes to a human with the chain attached. 67% Tier-1 deflection. CSAT 3.8 4.6. The support team was redeployed to retention.
“CSAT jumped 0.8 points in nine weeks. The bot cites our actual refund policy, every single time. Customers trust it.”
001
0
Production chatbots shipped since 2023.
002
0%
Median ticket deflection in the first quarter post-launch.
003
0
Hallucinated answers in production across all 21 deployments.
Three ways to start. Pricing in the email back.
Pick the shape that fits and Mohit will send your real number inside 24 hours.
Audit week + scope
5 days. We audit your KB, model the voice profile, write the trust architecture.
- 5-day senior audit
- Voice profile drafted
- Signed scope, line-item GBP
- No commitment to build
Single chatbot surface
4 to 6 weeks. One surface end-to-end: KB + voice + citation + escalation + action layer.
- Claude + RAG + MERN + AWS
- Full trust architecture
- 30-day walk-away both ways
- IP assigns on every commit
Multi-surface chatbot
6 to 10 weeks. 2-3 surfaces (web + WhatsApp + Slack) sharing one brain, voice, and trust layer.
- Shared KB + voice profile
- Multi-channel routing
- Unified dashboard
- Fixed scope
“CSAT jumped 0.8 points in nine weeks. The bot cites our actual refund policy, every single time. Customers trust it.”
— Rashmi P., Head of Support, UK EdTechWhat founders actually ask before signing
Pain-first, soft-second.
Every reply has a confidence score + citation. Below threshold, the bot doesn’t answer, it escalates with the full chain attached. When something does slip past, the replay tool lets you reproduce it in a sandbox, fix the prompt or KB, and add the case to the eval gold set so it never happens again.
No. We configure zero-retention tiers on every provider by default. For regulated industries, we self-host Claude or Llama on your infrastructure. DPA on file before week one.
Yes. We integrate alongside, bot answers what it’s confident on, your existing helpdesk handles the rest. Agents see the chatbot chain as part of the ticket so there’s no “repeat the question” moment.
Median inference + ops for 10K monthly conversations runs £600 to £1,200/month. Cost dashboard shows it per conversation. Most clients save 8× that in support headcount within the first quarter.
30-day walk-away both ways. Milestone billing 25/25/25/25. UK VAT registered, listed on Companies House, shipping since 2019.
Every custom AI chatbot development project has two senior engineers paired. Every decision into an ADR same day. Mohit reviews every PR. Two handovers in seven years, both inside 48 hours.
Yes. Voice profile + banned words + sentence patterns are config, not code. Marketing edits in a dashboard. Changes go live in 2 minutes. No deploy.
Yes, with 14 days’ notice. No cancellation fee.
In context
Inside the work.
A look at the kind of custom ai chatbot development surface we hand over — real screens, real data, documented and yours from day one.
One paragraph. That’s it.
Tell us where the bot should sit, what knowledge it needs, and what your customers deserve to hear. Mohit replies inside 24 hours: a clear yes, a clear no, or the one question that decides it.
What happens after the email lands
- < 24h
A personal reply.
Yes, no, or the deciding question. Straight to your inbox.
- Week 1
KB + voice profile.
We audit your knowledge base and draft the locked tone of voice.
- Wk 5
A bot customers trust.
Cites every answer, sounds like you, escalates cleanly.