AI chatbots that answer from your content and know when to hand over
Support and sales assistants grounded in your knowledge base, with citations, clarifying questions, clean escalation to people and honest measurement of what was resolved.
A good bot is not one that answers everything. It is one that answers what the content supports, asks when the question is unclear, and hands over, with context, when a person is needed. That routing decision is the product.
The deflection and escalation flow. The gold check is the control point: it is tuned on your eval set, and its thresholds are what you adjust when the bot is too eager or too timid.
Who this is for
Support teams, sales teams and product owners
Customer support
Deflect repetitive questions, resolve account tasks through tools, and hand the hard ones to agents with a summary already written. See support copilot for the agent-facing version.
Sales and pre-sales
Answer product and pricing questions from approved content, qualify the visitor and book a call. Useful on retail and SaaS sites; see retail and e-commerce.
In-product assistants
Help inside your own app, aware of what the user is looking at and what plan they are on. This is how the assistant inside our Eventify platform works.
Everything between the widget and the knowledge base
Content audit and retrieval index
We find gaps and contradictions in your help content, then build the retrieval layer with our AI knowledge base practice.
Account tools, scoped
Order status, bookings, plan details and simple changes, through least-privilege tools with limits and approvals for anything that costs money.
Escalation that agents thank you for
Hand-offs arrive with the transcript, a two-line summary, the sources the bot used and a suggested reply, inside the helpdesk you already run.
Safety and injection defences
Topic limits, PII handling, and resistance to instructions hidden in user messages or retrieved pages, tested by our prompt injection defence team.
Measurement you can trust
Containment rate, escalation quality, customer rating, cost per conversation and the top unanswered questions, reviewed monthly to drive content fixes.
Process
Four to eight weeks from content audit to launch
Audit and evals
Content review, top-100 real questions collected, eval set with expected answers and expected escalations.
Ground and tune
Retrieval index, answer style, citation format and the confidence thresholds tuned against the eval set.
Shadow mode beside human agents, then a cohort, then everyone. Thresholds adjusted on real data.
What goes wrong
The classic failures are a bot that answers confidently from stale articles, a bot that escalates everything because thresholds were never tuned, and a bot whose hand-offs arrive as a bare "customer needs help" with no transcript. Each is cheap to prevent in build and expensive to discover in reviews.
Three ways together: every answer is grounded in retrieved content from your knowledge base, the bot must cite what it used, and a confidence check routes anything weakly supported to a clarifying question or a human. The glossary explains hallucination and RAG.
What content does the bot need?
Help articles, policies, product documentation, past resolved tickets and, where appropriate, live data such as order status through a tool. Content quality decides bot quality; we audit it in week one and tell you what is missing or contradictory.
How long does it take to launch?
A grounded support bot on an existing help centre typically takes four to eight weeks to reach production, including evals, escalation wiring and a staged launch. Adding account-specific actions such as order lookups or refunds extends that.
Which channels can it run on?
Website widget, in-app chat, messaging platforms, internal chat tools and, through our voice AI service, the phone. The same grounded core serves all of them; what changes is identity, context and hand-off.
Have you built one yourselves?
Yes. Our own event platform, Eventify, includes an AI support and knowledge bot used by event organisers and attendees. The case study covers what we learned about escalation and content upkeep.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.