We build assistants grounded in your own documentation, catalogue and policies — so answers are specific, sourced and correct, and anything outside their remit goes to a person instead of being invented.
Why most chatbots fail
Scripted bots break the moment somebody phrases a question differently. Ungrounded AI bots do the opposite — they answer everything, confidently, including the things they have no basis for. Both end the same way: customers hammering "talk to a human" and the team trusting the bot less than the queue it was meant to shorten.
We build the middle path. The assistant retrieves from your approved content before it answers, shows where each answer came from, refuses gracefully when it has no source, and hands over to a person with the full conversation attached. What it cannot answer becomes a report of what your documentation is missing.
Answers are retrieved from your help centre, product data and policies — not from whatever the model happens to remember.
When a person is needed, the transcript, customer record and intent travel with the conversation instead of starting over.
Topic boundaries, refusal behaviour, escalation triggers and tone are set by you and testable before launch.
Deflection, containment, satisfaction and unanswered questions reported weekly, so the bot keeps improving.
What we build
Four assistant types that cover the majority of the conversations businesses want handled properly.
A first line that resolves the repetitive 60–70% — order status, delivery windows, returns, account access, opening hours — using your live systems and published policies, and escalates the rest with context intact.
An assistant on your site that answers product and pricing questions, qualifies against the criteria your sales team actually uses, books meetings into real calendars and writes the whole conversation into the CRM.
The bot your own team asks: what the policy is, where the template lives, which supplier we use, what we told this client last quarter — answered from internal documentation with permissions respected per user.
The same assistant, one knowledge base, across every channel your customers use — website widget, in-app panel, WhatsApp, Messenger, email autoresponse and voice — with consistent answers on all of them.
Operating standards
A chatbot is only an asset while people trust its answers. These are the standards we hold every build to.
Every response traced to the document or record it came from.
Streaming responses, so the first words arrive immediately.
Escalation is always one step away, never buried in a menu.
Unanswered questions reviewed weekly and turned into content.
Engagement models
Prove it on one channel, or run the whole conversational layer as a managed service.
One channel, one knowledge base, one clearly bounded set of questions — live and measurable.
from $540 /month
The smallest version that can prove the value.
A bot wired into your live systems, so it can look things up and act rather than only explain.
from $980 /month
The lane most support and sales teams end up in.
We run the assistant as a service — content, tuning, monitoring and new channels as you add them.
from $1,690 /month
Best for teams where chat is now a primary channel.
Send us a month of real transcripts. We will tell you what proportion is genuinely automatable, what should stay with a person, and whether the numbers justify building anything at all.
How we work
We start from the questions you already receive, so the first version answers the things people actually ask.
We read your real tickets, chats and calls, cluster them by intent and volume, and agree which clusters the bot should own on day one.
Content is gathered, cleaned, chunked and indexed, with gaps flagged where no approved answer exists yet.
Topic boundaries, refusal behaviour, escalation triggers and voice agreed and written into the system prompt and evaluation set.
The assistant is built and connected to the systems it needs to read — orders, accounts, CRM — with every tool call logged.
Scored against a held-out set of real questions with the correct answers known, before a single customer sees it.
Staged rollout behind a percentage of traffic, then weekly review of unanswered questions and escalation reasons.
Outcomes
The measurable difference once the repetitive half of your inbox answers itself.
No queue at 2am, no weekend backlog on Monday, no waiting for the one person who knows the answer.
Routine questions resolved in the chat, so your team spends its day on the cases that need judgement.
Enquiries arrive with the requirement, budget signal and timeline already captured in the transcript.
One knowledge base behind every channel, so the answer does not change with the door the customer walked through.
Unanswered questions become a ranked list of the documentation your customers needed and you did not have.
Retrieval-grounded responses with refusal behaviour beat a confident guess every time.
FAQ
Still unsure? Send us a note — we reply personally.
It is built not to. Answers are retrieved from your approved content before the model writes anything, each response carries the source it used, and when retrieval finds nothing relevant the assistant says so and offers a person instead of guessing. We also score it against a held-out set of real questions with known answers before launch.
Whatever you would give a new hire: help centre articles, product and pricing information, policies, FAQs, and past ticket resolutions. If the content is thin we will tell you during the audit — a bot cannot answer what nobody has written down, and the gap list is usually valuable on its own.
Always, and in one step. Handover triggers on explicit request, on repeated failure to answer, on detected frustration and on any topic you mark as human-only. The transcript, customer record and detected intent go with the conversation so nobody has to repeat themselves.
Website widget, in-app panel, WhatsApp, Facebook Messenger, SMS, Slack and Teams for internal use, plus email triage. One knowledge base sits behind all of them, so the answer does not change depending on the channel.
Yes. Connected to your systems it can look up an order, check availability, book a meeting, create or update a CRM record and raise a ticket. Anything that changes data runs behind explicit guardrails and is logged with the conversation that triggered it.
The assistant detects the customer language and replies in it. Where your source content only exists in one language we either translate the knowledge base or restrict the bot to the languages you can support properly — an accurate answer in English beats a fluent guess in Spanish.
No. We use enterprise API tiers with training disabled, keep transcripts inside your infrastructure, apply retention rules you set and can run open-weight models in your own environment where policy requires it.
Send us a month of transcripts. We will tell you what proportion an assistant should handle, what it must never touch, and what it costs to build.
