AI Chatbots That Actually Help Your Customers
Forget scripted decision trees. We build chatbots powered by your actual data that handle real conversations - and know when to escalate.
AI chatbot development is the engineering work of building a conversational interface backed by a large language model, your real data, and a clear escalation path to humans. The chatbots Borah Labs ships are not scripted decision trees and not generic ChatGPT widgets — they retrieve answers from your documentation, product database, ticket history, and CRM in real time, cite their sources where it matters, and hand off to a human with full conversation context the moment confidence drops or intent moves outside their scope. A typical AI chatbot project ships in 4 to 8 weeks, starts at $8,500, and includes the parts most off-the-shelf chatbot platforms quietly skip: a real evaluation set, structured-output validation, PII redaction, conversation analytics, escalation routing into your help desk, and multi-channel deployment (website, app, WhatsApp, Slack, Microsoft Teams) from a single brain. We measure success against deflection rate, CSAT, and the dollar value of human time recovered — not vanity metrics like total messages sent.
Trusted by
Who this is for
Built for teams with one of these problems.
If your situation looks like one of these three, you'll feel at home with us.
B2B SaaS, DTC ecommerce, fintech, healthtech
Head of Support at a 50-500 person SaaS or ecommerce company
Sees ticket volume scaling faster than headcount, with 60-70% of incoming tickets being repeat questions answered in the help center. Wants to deflect those without making customers angry, while preserving CSAT and giving agents better context on the tickets that do escalate.
Vertical SaaS, two-sided marketplaces, embedded fintech
Founder of a vertical SaaS or marketplace
Customers repeatedly ask the same product, billing, and onboarding questions on sales calls and in-app chat. Wants a chatbot that doubles as a sales assistant and a support deflection layer, integrated into the existing product rather than a third-party widget that breaks the UI.
Real estate, healthcare admin, legal services, professional services
Operations leader at an SMB with high inquiry volume
Owns the inbox and the phone line, both of which are full of repetitive intake questions. Wants a chatbot on the website and on WhatsApp that captures intent, qualifies the inquiry, books a meeting where appropriate, and escalates to a human only when there's real value at stake.
Sound familiar?
- Your current chatbot frustrates customers with rigid scripted responses that never quite match the real question, and your team spends as much time apologizing for the bot as answering tickets.
- Support volume keeps growing month over month but the hiring market for qualified support reps is broken, slow, and expensive — and another full-time hire just kicks the problem down the road.
- You need genuine 24/7 coverage across time zones for a global customer base but the math on staffing a round-the-clock human support team doesn't pencil out at your stage.
How we solve it
- We ground every chatbot answer in your documentation, product data, knowledge base, and historical tickets so responses are specific and citable — not generic LLM hallucinations dressed up in your brand colors.
- Smart escalation routes complex or low-confidence conversations directly into your help desk (Zendesk, Intercom, HubSpot, Help Scout, Freshdesk) with the full conversation transcript, intent classification, and a suggested next step so your agents never start from zero.
- One chatbot brain, multiple channels. Deploy the same model on your website, in your mobile app, on WhatsApp Business, in Slack or Microsoft Teams for internal use, and in email autoresponders — with consistent tone, consistent escalation rules, and consistent analytics.
What you get
Deliverables, with sample artifacts.
Concrete outputs you can expect from this engagement — and a sample of what each one looks like.
Customer Support Chatbot
An LLM-powered support bot grounded in your documentation, help center, ticket history, and FAQ. Answers come with citations so the customer can verify, the bot signals uncertainty rather than making things up, and complex tickets escalate cleanly into your existing help desk with full transcript and suggested resolution. We tune the persona, the tone, and the escalation thresholds with your team during the engagement.
Sample artifact
A deployed support chatbot on your site (or channel of choice), an analytics dashboard showing deflection rate and CSAT by topic, and a Zendesk/Intercom integration for escalations.
Lead Qualification Bot
A conversational bot that replaces the contact form, qualifies the visitor through a natural conversation (not 14 form fields), captures the data you actually need into your CRM, scores the lead based on your criteria, and books a meeting on the right rep's calendar via SavvyCal, Calendly, or Chili Piper. Designed to convert significantly higher than a static form because it actually feels like talking to a knowledgeable salesperson.
Sample artifact
A live bot on your homepage or pricing page, a HubSpot/Salesforce integration writing qualified leads with intent data, and a per-channel conversion dashboard.
Internal Knowledge Bot
A private chatbot your team queries for company policies, product specs, runbooks, security procedures, historical decisions, or onboarding answers. Indexes Notion, Confluence, Google Drive, Slack history, GitHub wikis — wherever your knowledge actually lives. Drops into Slack or Teams so it's where your team already is, with role-based access so different teams see different sources.
Sample artifact
A Slack or Teams bot, an admin dashboard showing query volume by team and most-asked topics, and a re-indexing job that runs nightly so the bot is never stale.
Multi-Channel Deployment
The same chatbot brain deployed across every channel your customers actually use. Web widget, embedded SDK in your mobile app, WhatsApp Business via Twilio or 360dialog, SMS, Facebook Messenger, Slack, Microsoft Teams, and email autoresponders. Single brain means consistent answers, consistent escalation, and consistent analytics — no more divergent quality between web chat and WhatsApp.
Sample artifact
A single chatbot service backing every channel, a unified inbox view of conversations across channels, and channel-specific compliance configuration (WhatsApp opt-ins, etc.).
Tech stack
The tools we ship with.
Battle-tested, boring where it should be, modern where it earns it.
OpenAI
Anthropic
LangChain
- Pinecone
TypeScript
Next.js
- WebSocket
Redis
Process
Week by week — what shipping looks like.
A typical engagement, end to end. Concrete deliverables every milestone.
Week 1
Scoping, content audit, eval set
- Workshops with the support, sales, or ops team that owns the conversations the bot will handle.
- Audit the content sources we'll ground answers in — docs, FAQ, ticket history, CRM notes — and flag the gaps before we ship.
- Build a 50-200 question evaluation set from real customer questions, with the right answer labeled by your domain experts.
- Decide model, retrieval strategy, escalation policy, and the deflection-rate target we'll measure against.
Weeks 2-3
Build the chatbot core
- Ingest, chunk, and index your knowledge sources with a hybrid keyword + vector retriever and a cross-encoder reranker.
- Build the conversation handler: intent classification, retrieval, grounded generation, citation injection, structured-output validation.
- Implement the escalation path with conversation context, intent label, and suggested next steps written into the help desk handoff.
- Run the eval set on every commit, gating deploys on accuracy regression.
Week 4
Channels, persona, analytics
- Wire the bot up to the channels in scope — web widget, mobile app SDK, WhatsApp via Twilio, Slack, Teams.
- Lock the persona, tone-of-voice, and refusal language with stakeholder review on real conversation transcripts.
- Build the analytics dashboard: deflection rate, CSAT signal, topic distribution, escalation rate per topic, latency, cost.
- Privacy & compliance pass — PII redaction, log retention, consent flows, channel-specific opt-ins.
Weeks 5-6
Soft launch & iteration
- Soft launch behind a feature flag to 5-10% of traffic. Daily review of conversations with the support / sales team.
- Tune retrieval, prompts, and escalation thresholds based on real conversation data — not lab tests.
- Ramp to 100% once deflection rate hits target and CSAT signal stays neutral or positive.
- 30-day support window included; optional retainer covers monthly model upgrades and content reindexing.
Featured plan · One-time
Launch4 weeks
Recommended starting point for ai chatbots that actually help your customers. Projects from $8,500.
- 5 custom pages
- Mobile-first responsive design
- Basic SEO setup
- CMS integration
- 2 revision rounds
- 4 weeks delivery
Choosing a stack
When to pick what — and when to skip it.
The honest version. Real trade-offs, not marketing slideware.
Build vs Intercom Fin / Zendesk Answer Bot
When to use
Off-the-shelf bots like Intercom Fin or Zendesk Answer Bot are the right call when your knowledge base is already in their system, your channel surface is just web chat, and your customization needs are low. They ship in days, not weeks.
When to avoid
When you need real customization (custom retrieval logic, specific escalation rules, channels they don't support, compliance constraints, or grounding in data that lives outside their tool), the off-the-shelf vendor pricing scales painfully fast and you hit the ceiling on what you can change. That's when custom is cheaper and more flexible.
Web widget vs WhatsApp
See Mabbly →When to use
Web widget when your customers reach you via the site, when context-rich answers matter, and when you can deploy real interactive UI inside the chat (buttons, file upload, structured forms). Best for SaaS and ecommerce.
When to avoid
If your audience lives on WhatsApp (consumer brands in the US, almost all of LATAM, parts of Europe, India), the website widget will be ignored. Lead with WhatsApp Business via a verified provider like Twilio or 360dialog, and treat the web widget as a fallback.
RAG vs fine-tuning for chatbots
See Koppa AI →When to use
RAG (retrieval-augmented generation) is the right default for chatbots. Your content changes — RAG re-indexes; fine-tuning would need re-training. RAG also gives you citations and prevents hallucination on facts.
When to avoid
Fine-tuning is appropriate only when you need a very specific tone-of-voice the base model can't match through prompting, and you have hundreds of high-quality labeled conversation examples. Even then, most teams ship RAG first and add fine-tuning only if the gap is real.
Self-host the model vs API
When to use
Hosted APIs (OpenAI, Anthropic, Bedrock, Vertex) are the right default for chatbots — fast inference, predictable cost per conversation, model upgrades you don't have to do yourself. Use providers with zero-data-retention contracts for regulated data.
When to avoid
Self-hosting an open-source model only makes sense at very high sustained volume, with hard data residency rules the cloud APIs can't meet, or for chatbots handling content the public APIs refuse. The infra and ops cost of running your own GPUs is rarely worth it before you cross the millions-of-conversations-a-month threshold.
Real shipped work
Production case studies, not pitch decks.
Two engagements that mirror the kind of work you'd hand us.

Marketing Agency
Mabbly
How we built an AI agent platform that automated case study creation for a Chicago marketing agency - from concept to working product in under two months.
Read the case study →
Brand Protection SaaS
Koppa AI
How we rebuilt a broken PHP platform from scratch into a modern brand protection SaaS - new database, new architecture, new features that were previously impossible.
Read the case study →FAQ
Common questions
Everything you need to know about our ai chatbots that actually help your customers services.
Accuracy depends almost entirely on your data quality. On domain-specific questions backed by good documentation, we typically achieve 85-95% accuracy after the first two weeks of tuning. The bot clearly signals uncertainty when it can't ground an answer — and when confidence is low, it offers to connect with a human rather than guessing. We measure accuracy against an evaluation set of real customer questions, not vanity benchmarks.
Yes — and we build the feedback loop into the product from day one. Customers can rate answers, agents can flag incorrect responses during escalations, and we use the flags to expand the evaluation set, improve retrieval, and refine prompts on the next deploy. The system gets measurably better over time, with each improvement gated by the eval suite so quality never regresses.
The minimum viable input is your help center or documentation. From there, the chatbot improves dramatically with each additional source: FAQs, product descriptions, pricing pages, ticket history, win/loss notes from sales, and structured data from your product database. The more grounded sources you provide, the less the bot has to fall back on the base model's general knowledge — and the higher the accuracy on questions that matter to your business.
All data stays in your infrastructure or approved cloud providers. We implement role-based access controls, audit logs of every conversation, automatic PII redaction in logs, and zero-data-retention contracts with model providers (OpenAI Enterprise, Anthropic ZDR, Azure OpenAI, AWS Bedrock) for any sensitive data flow. No training data leaves your environment, and we're happy to sign a BAA or DPA before kickoff for HIPAA, SOC 2, or GDPR-bound work.
We monitor chatbot performance daily — deflection rate, CSAT signal, escalation rate per topic, latency, and cost — update the knowledge base when your documentation changes, and tune responses based on real user feedback and agent flags. Maintenance plans start at $750/month and scale up to a full-retainer model where new features ship monthly alongside performance work. Most clients move to a retainer after the initial engagement.
Yes — that's the line between a chatbot and an AI agent. We can wire the bot up to your APIs to look up an order status, change a subscription, schedule a meeting, create a support ticket, or update CRM data. Anything with a write effect goes through explicit confirmation prompts and gets logged for audit. If you need heavy multi-step automation, see our AI agent development service for that specifically.
A focused single-channel chatbot ships in 4-6 weeks. A multi-channel deployment (web + WhatsApp + Slack, for example) with custom escalation and CRM writes runs 6-8 weeks. The two-week AI Strategy & Audit is a lighter starting point if you're not yet sure which use case to ship first. We commit to dates during scoping and ship to them.
Per-conversation cost depends on conversation length and the model. A typical support chatbot runs $0.01–$0.10 per conversation in API costs after caching and prompt optimization. At 10,000 conversations a month that's $100–$1,000 in model spend. We design every system with token budgets, caching, and fallback to cheaper models for routine questions so the cost is predictable and visible on a dashboard.
Almost never the goal we recommend. The math that works is: deflect 30-70% of repetitive volume so your existing team has more time for the conversations that actually need a human. The bot escalates with full context so agents start ahead of where they would on a fresh ticket. Customers who do reach a human are happier because the bot didn't waste their time, and your team grows into a higher-leverage role rather than a queue-clearing one.
Ready to ship ai chatbots that actually help your customers?
Tell us what you need. We will scope it, price it, and give you a timeline - all before you commit to anything.
No commitment. No sales pitch. Just a clear plan.