A bot that answers, or no bot at all.
An agent on your site that knows your prices, your stock and your policies, cites where each answer came from, and hands to a person the moment it should — rather than a widget that deflects until the visitor gives up.
Almost everyone has met the bad version. It greets you, offers three buttons that do not match your question, answers a fourth question you did not ask, and eventually offers a contact form — which is where you were headed before it interrupted. It exists to reduce tickets, not to help, and customers can tell.
The difference is grounding. An agent that has been given your actual content — prices, stock, policies, delivery times, the answers your team types out twenty times a week — can answer rather than deflect. One running on a general model with your logo on it is guessing, and it will guess confidently about your refund policy in front of a customer.
The second difference is being able to do something. Answering "do you deliver to Sharjah" is useful; checking the postcode, quoting the fee and booking the slot is a different product. Where you have a system worth connecting, the agent should reach into it.
What we build
Grounded in your content
Your pages, catalogue, policies and documents — indexed and kept current, so answers come from what you actually published rather than from what a model half-remembers about your industry.
Citations on every answer
Each reply carries the page it came from. The customer can check it, and you can tell instantly whether a wrong answer is a bad model or a page that needs fixing.
It says when it does not know
The single most important behaviour and the one most bots lack. Outside what it has been given, it says so and offers a person, rather than inventing a plausible sentence about your warranty.
Actions, not just answers
Book the appointment, check the stock, quote the delivery, raise the ticket, take the details. Connected to whatever you already run — and left as answers-only where connecting it would be more risk than it is worth.
A clean handover
When it escalates, the person receiving it gets the whole conversation and what the customer already told it, so nobody is asked their order number twice.
Guardrails and a transcript
Topics it will not touch, claims it cannot make, and every conversation logged and reviewable. You find out what customers are actually asking, which is usually worth more than the deflection rate.
How we work
Read what you already answer
We start from your real inbox and your real pages, because the questions worth handling are the ones your team is already typing out by hand.
Narrow first
One area, answered properly, before it is allowed near everything. A bot that is excellent about delivery and honest about the rest beats one that is mediocre everywhere.
Watch the transcripts
The log is the product feedback. What it could not answer becomes next month's content, next month's integration, or a page on your site that should have existed already.
What separates one from the other
The same question, asked of the widget most sites have and of an agent that has been given something to work with.
| Asked | A typical site bot | A grounded agent |
|---|---|---|
| Do you deliver to Sharjah? | Here are some helpful links! | Yes — AED 25, next day if ordered before 4pm. Checks the postcode and offers the slot. |
| Is this in stock in blue? | Please contact our team. | Reads the catalogue, answers, and offers to hold one. |
| What is your refund policy? | Invents something plausible. | Quotes the policy and links the page it came from. |
| Something outside its knowledge | Answers anyway, confidently. | Says it does not know and puts a person in the conversation. |
| A complaint | Loops through the same three buttons. | Escalates immediately, with the transcript attached. |
Deflection rate is the metric agencies report because it always looks good. We report what it answered correctly, what it escalated, and what it was asked that nobody had written an answer for — that last column is usually the most valuable thing the agent produces in its first month.
What you get
- Agent deployed on your site, in your brand
- Content index with a refresh schedule
- Citations, escalation path and handover to your team
- Integrations where they earn their place
- Written guardrails and refusal behaviour
- Transcript log with the questions you are actually being asked
Before you ask
- Will it make things up?
- It is built not to, and that is a design decision rather than a hope. It answers from an index of your own content, cites what it used, and is instructed to refuse outside it. Nothing removes the risk entirely, which is why the guardrails name the topics it must never answer on its own and why every conversation is logged.
- Which model do you use?
- Whichever fits the job, and it is replaceable. Most of the value is in the grounding, the guardrails and the integrations — the model is the cheapest part to swap and the part that improves without us doing anything.
- Can it book, quote or check stock?
- Where there is a system to connect, yes. Where connecting it would put a customer-facing bot inside something fragile, we will say so and keep it to answers — a bot that cannot book is fine, a bot that books wrongly is not.
- What does it cost to run?
- A build fee, then a monthly cost driven mostly by conversation volume and how often the index refreshes. We size it against your real traffic before you commit, the same way we size an ad budget.
- What if we already have one?
- We will audit it before proposing a replacement. Often the widget is fine and the problem is that nothing was ever given to it — no index, no policies, no integrations — in which case fixing that is cheaper than starting again.
The rest of the stack
Tell us what you're building.
Send the brief — or the problem, if the brief doesn't exist yet. We come back within one business day with an honest view of how we can help.
Start a project