AI chatbot development services for customer service, support, and ecommerce. We ship RAG-grounded chatbots on Claude Sonnet 4.6, Haiku 4.5, and GPT-5 mini, deployed to your web widget, WhatsApp, voice, or Slack/Teams. Eval-gated, guardrailed, token-optimized. First chatbot live in 30 days, behind a feature flag.
A production AI chatbot is a conversational interface grounded in proprietary data and operated under measured confidence thresholds, not a stand-alone LLM prompt. Retrieval-augmented generation (RAG) pulls answers from internal docs, help centers, and ticket history. A confidence gate routes high-certainty answers autonomously, drafts replies for human review in a middle band, and refuses below threshold. Unlike a ChatGPT-style assistant bolted onto a website, a production chatbot cites a source document for every answer. Unlike rules-based bots (Intercom Resolution Bot, Drift), it escalates on a confidence threshold rather than fixed intent rules. Accuracy is regression-tested on a frozen golden set with faithfulness and answer-relevance scoring (Ragas, Langfuse). Common stacks: Claude Sonnet 4.6 or GPT-4o-mini for synthesis, pgvector or Pinecone for retrieval.
Every customer service chatbot, ecommerce chatbot, and internal Slack chatbot below has been shipped from this exact playbook. Each one comes with an eval suite, audit logging, confidence gating, and a per-turn cost target. Not a polished demo.
The crown-jewel use case. RAG-grounded chatbot over your help center + ticket history that handles password resets, order status, refund eligibility, and policy lookups. Confidence gate at 0.7; anything below escalates to a human with the AI's draft attached. We've shipped these to 30–45% tier-1 deflection on Zendesk and Intercom queues.
Conversational AI grounded in your Shopify / WooCommerce catalog. Product recommendations from natural-language criteria, order status ("where's my package?"), return initiation, size and fit Q+A. Function calls into your OMS + 3PL APIs so the bot can act, not just describe.
Chatbots inside Slack or Microsoft Teams that retrieve from Notion, Confluence, Drive, or your code repo. Onboarding Q+A, policy lookups, on-call runbooks. Built on Sonnet 4.6 + tool use for the 6+ step internal workflows where a one-shot answer isn't enough.
Where the user actually is. A whatsapp ai chatbot ships via the Meta Cloud API for international support and lead capture; template-message flows for outbound, free-form replies within the 24-hour service window. Voice chatbots on Twilio / Vapi over the OpenAI Realtime API or Deepgram + Sonnet 4.6 for sub-second voice. Latency profiles differ; we'll tell you which channel makes sense before we build.
When the answer lives in a 10,000-document Drive folder, a contract library, or a research archive. Retrieval-augmented chatbot with pgvector or Pinecone, top-k 5 retrieval re-ranked by bge-reranker-v2, eval-tested on your real question set before launch. We measure groundedness, not just BLEU.
Website chatbot that runs a structured intake (BANT, MEDDIC, or your custom rubric), drops qualified leads into HubSpot / Salesforce with a transcript, and books a meeting via Cal.com or Chili Piper. Less marketing-deck, more pipeline. Built for revenue teams that actually want call shows, not vanity completion rates.
Six stages every production chatbot turn moves through, from user message to logged outcome. Skip any one and you get the demo competitors ship instead of a chatbot that deflects tier-1 traffic. Each stage carries its own latency budget, model pick, and failure mode. Hover any stage for the operator detail.
Latencies and token counts are typical production traces from shipped chatbots. Your eval set sets the real budgets.
A real customer-service replay. Left side, a chat surface streaming. Right side, the same six stages from the anatomy lighting up as each turn fires. This is what 'shipped' looks like, not what the demo deck shows.
Illustrative per-decision economics — typical engagement-band figures, not literal client numbers.
Most chatbot vendors will quietly say yes to any channel. We won't. Pick a channel to see the deployment surface, latency profile, the actual stack, and the part competitors hide: where it fails. Channel mix is decided in the audit, not in the sales call: web widget, WhatsApp, voice, Slack + Teams, Discord, Telegram, Instagram + Messenger, and SMS + iMessage.
Embedded floating widget on your marketing site, in-product help center, or post-login dashboard. We ship a Preact-based widget that loads under 35KB gzip and streams responses token-by-token. Same widget surface across desktop and mobile.
<800ms first token · streamed · p50 1.4s end-to-end
Web widget bounces hit ~70% of sessions in B2C. If your customers aren't already on your website (e.g. they're in WhatsApp or your mobile app), the web widget is the wrong channel. Pick the channel your buyers already live in.
The chatbot reply is half the story. When a turn fires a tool, an automation graph kicks in: classify, lookup, decide, act. Here's a real WhatsApp refund flow playing out node-by-node, with a branching decision that escalates to a human when policy says it should.
Built on n8n, LangGraph, or custom, depending on your stack. Cost chips are illustrative per-decision economics.
The naming has drifted. Every vendor calls everything an “AI agent” now. The honest distinction: chatbots are scoped and short-turn; agents are multi-step and long-horizon. Most teams asking for an agent need a chatbot first. Per-dimension honest comparison below.
Generalizations from shipped client work. Specifics vary per workload; we benchmark on your eval before recommending.
A conversational AI platform (Kore.ai, Cognigy, Yellow.ai, Boost.ai) is a hosted multi-channel runtime (intent designer, NLU layer, dialog manager, channel adapters, analytics) sold as a SaaS subscription. You configure intents in their console, deploy to web + WhatsApp + voice from one place, and the platform owns the runtime.
What we ship is the opposite shape: production conversational AI built in your repo, against your data, on Claude Sonnet 4.6 + GPT-5 mini routed per turn, deployed to the same channels but with the model layer, retrieval layer, and eval suite owned by you. The trade-off is honest. If you want a no-code console and accept a vendor-shaped runtime, buy a conversational AI platform. If you need RAG over a 10,000-document corpus, custom tool use into Salesforce + Zendesk + your OMS, sub-second voice on the Realtime API, and an eval suite you can extend, build conversational AI as software. Most of our clients hit the platform's ceiling on retrieval depth or tool-call complexity by month 6 and migrate.
The conversational AI company you pick should be honest about which shape fits. A conversational AI platform vendor will tell you their console covers everything; we'll tell you when it does (generic FAQ surfaces, single-product Q&A, low retrieval depth) and when it doesn't (RAG over private corpora, multi-system tool calls, custom eval).
A production chatbot is not one model. It's a routed pipeline. Cheap classify in front, grounded generate in the middle, cheap log + eval at the back. Here's the default chatbot stack we ship; we'll re-pick per workflow if your eval data demands it.
Five tactics stacked, in order of impact for chatbots. Most chatbot pilots see effective per-turn cost drop to 6–10% of the naive baseline at the same eval-suite quality. This optimization pass is included in every chatbot pilot, post-cutover.
Four stages, milestone-billed, with a walk-away point at the retrieval baseline. Most chatbot failures happen because the team skipped the eval set or skipped retrieval tuning. Both are in week 1 and week 2 here, not bolted on at the end.
We harvest 50–200 real questions from your ticket archive (or run a structured user interview if you're greenfield) and build the eval set the chatbot will be measured against. Scope locked: channels, knowledge sources, tool surface, escalation rules.
Ingest your docs into pgvector or Pinecone, run chunking experiments (semantic vs fixed-size, header-aware vs not), tune top-k and re-ranker, and score retrieval against the eval set. Most chatbot quality issues are retrieval issues, fixed here.
Wire the full anatomy: classifier → retrieval → tool use → generate → guardrail → log. Behind a feature flag, in your repo (or ours, your call). PII scrub, refusal rules, confidence gate, audit-log every turn. Channel-specific UI shipped in parallel.
Shadow mode against your existing channel for 2 weeks. Score on the eval set, score on real traffic, score on cost. Roll out at 10% → 50% → 100% if numbers hold. Run the token-optimization pass. Most chatbots see 60–85% cost reduction at the same quality.
The same chatbot turn (classify → retrieve → tool → generate → guardrail → log) across Sonnet 4.6, Haiku 4.5, and GPT-5 mini. Pick a model on the left; the model= line swaps and the per-turn cost stat updates. This is how we choose: run your eval, then look at the bill.
from anthropic import Anthropic
client = Anthropic()
def chat_turn(user_msg: str, history: list[dict]) -> dict:
# 1. Intent classify with Haiku 4.5 (~$0.0001 / turn)
intent = classify_intent(user_msg)
# 2. RAG retrieve from pgvector (top-k 5, re-ranked)
docs = retrieve(query=user_msg, k=5, rerank=True)
# 3. Tool-aware generate: switch the reply model
response = client.messages.create(
model="claude-sonnet-4.6",
max_tokens=600,
system=SYSTEM_PROMPT + format_docs(docs),
tools=[zendesk_create_ticket, order_status_lookup],
messages=history + [{"role": "user", "content": user_msg}],
)
# 4. Guardrail: confidence + PII + policy
verdict = guardrail.check(response, intent=intent)
if verdict.action == "escalate":
return handoff_to_human(response, verdict)
# 5. Log to Langfuse for nightly eval
langfuse.log(trace_id, response, verdict, tokens=response.usage)
return response
Same pricing as our other engagements. Most clients begin with the audit to scope channels + scope retrieval, run a 4–6 week pilot on the highest-ROI channel, then move to monthly for the next 2–3.
Find the chatbot workflow worth shipping before you commit a budget.
One chatbot shipped end-to-end on your highest-ROI channel, with eval data, not a demo.
Embedded squad shipping the next chatbot channel + tuning the live one.
We didn't want a demo. We wanted something on-call could ship Friday. The deflection numbers held at week 8. The eval suite caught the one regression we'd have shipped otherwise.
Three anonymized chatbot capability patterns drawn from real engagements. Named references shared under NDA once we know what you're building.
Inbound Zendesk queue averaging 6-hour first-response time; tier-1 reps spending 60% of time on password resets, billing questions, and feature-availability lookups.
Web-widget chatbot with RAG over the help center + ticket history. Haiku 4.5 classifier, Sonnet 4.6 generate, function calls into Zendesk for ticket creation. Confidence gate at 0.7; sub-threshold escalates with a drafted reply attached for the agent. Voice-channel sibling: see the published openai realtime voice agent case study for the same deflection pattern on inbound voice at $0.10/call.
International D2C brand getting product Q+A and 'where's my order' inquiries via WhatsApp from 14 countries. Manual reply queue 18 hours long at peak.
WhatsApp Cloud API chatbot grounded in Shopify catalog + 3PL tracking data. Multilingual via Sonnet 4.6 native multilingual; function calls into Shopify Admin API + Aftership for live order status. Refund initiation gated to human review.
On-call rotation drowning in repeated questions about runbook locations, alert ownership, and dashboard URLs. Same 12 questions answered nightly by senior engineers.
Slack chatbot with RAG over Notion runbooks + the team's PagerDuty service catalog. Tool calls into PagerDuty for on-call lookup and Grafana for dashboard linking. Escalates to senior on-call if confidence drops or alert is sev-1.
Book a free AI chatbot audit. We'll review your existing chatbot or support queue, recommend channels (web · WhatsApp · voice · Slack/Teams), pick models per stage (Sonnet 4.6 / Haiku 4.5 / GPT-5 mini), project token cost vs your current spend, and give you a 90-day chatbot roadmap, whether you're starting from scratch (build ai chatbot from zero) or refitting a conversational AI platform you already pay for. No deck, no obligation to build.
The long-form pieces our delivery team points new clients to when they're scoping a chatbot build.
The category definition — what counts as a platform, what's just a chatbot SDK.
Which channels (web, WhatsApp, voice, in-app) actually pay back for support deflection.
End-to-end build: product catalog ingestion, RAG, intent routing, conversion eval.
The architecture pattern we deploy: retrieval, rerank, citation, eval, guardrails.
WhatsApp Cloud API + Claude Sonnet — the 6-day shipped-to-staging recipe.
Comparative analysis of chatbots actually running in production at named brands.
Building a chatbot often connects to a sibling AI service. These pages go deeper on the adjacent decisions.
Anthropic Claude integration: the default reply model for our chatbots.
GPT-5-mini as classifier · Realtime API for sub-second voice.
When you need multi-step planning and long-horizon tool use, not single-turn replies. Sub-2s latency? Stay on chatbot. 10s-10min plans? Agent.
Plug your chatbot into Salesforce, Zendesk, HubSpot, NetSuite.
Ecommerce chatbots: WISMO, product Q&A, abandoned-cart recovery.
HIPAA-scoped triage and patient-intake chatbots on EHR-integrated stacks.
Pre-build discovery audit when you're not sure if a chatbot is even the right primitive for your workflow.
When the KB is internal, not customer-facing. The sibling pillar for agent-assist + employee Q&A patterns.
Umbrella pillar for when the engagement spans chatbot + agent + RAG + app shell, not just chat surfaces. End-to-end AI development services across Claude and GPT.
When the front door is a conversation and the back end is the workflow. Chatbot pilots that route, ticket, refund, and update CRM.
Cross-pillar index. All 12 AI service pillars, methodology, benchmarks, and the open-source eval harness in one map.
How we keep chatbot quality from drifting in production — frozen eval set, weekly replay, no-promote on regression.