Demand
Most enquiries are not lost to a competitor. They are lost to the eleven hours between the question and the answer. An agent grounded on your own catalogue closes that gap — and, configured properly, disqualifies the people you should not be selling to.
Enquiries arrive at eleven at night and are answered at nine the next morning.
Your team answers the same four questions all day and calls it sales.
Every enquiry gets the same reply, whether it is a serious buyer or a student doing research.
You bought a chat widget and it invents answers, so nobody trusts it and it has been quietly switched off.
Nobody can say which enquiries were worth answering, because none of them were scored.
- 01
The knowledge base is the product. The model is not
What decides whether this works is whether your catalogue, prices, prerequisites, exclusions and edge cases exist in a form something can retrieve — and in most companies they exist only in a salesperson's head. Building that base is the majority of the engagement and it is the part that keeps working when the model underneath is replaced next year. If your product cannot be written down, this is not the service you need first.
- 02
Grounded answers, and a refusal that works
The agent answers from your material or it does not answer. Where retrieval returns nothing, it says so and hands to a person rather than producing a fluent guess — which is the single failure that destroys trust in these systems and the reason most of them get switched off. Refusal is designed, tested adversarially and measured, not left to the model's disposition.
- 03
Transparency as configuration, not tone of voice
Price stated early. What the product does not do, said plainly. Enquirers it is wrong for told they are wrong for it, rather than advanced into a meeting. This is the opposite of the default configuration, which optimises for booked calls — and it is why the funnels we build look worse in the middle and better at the end. Filtering early does not improve the numbers on paper. It improves the only number that is revenue.
- 04
Scoring, qualification, priority
Every conversation produces a qualification state, not just a transcript: who this is, whether they meet the prerequisites, how specific the questions were, how fast they came back, which offer they entered on. The score sets the path — booked, nurtured, or told honestly that this is not for them. A queue treats a serious buyer and a curious student identically; a scored list does not.
- 05
Escalation that does not restart the conversation
When a person takes over they arrive with the history, the qualification state and the open question — not a fresh chat window. The handover is the moment most deployments lose the sale, because the customer has to explain themselves twice and concludes they are talking to a machine after all.
- 06
In the channel the customer chose
WhatsApp through the official Business API, Telegram, Instagram, web chat, email — on your platform, with the number and the history owned by the company rather than by an employee's handset. We have built this on Respond.io and we will build it on what you already run: the configuration is the work, and the same platform set up the ordinary way produces the ordinary result.
- 07
Measured, so it can be argued with
What was asked, what was answered from source, what was refused, what escalated, what converted, and what it cost per qualified conversation. A deployment nobody measures is a deployment nobody can improve — and the interesting number is usually the refusal rate, because that is where the missing half of your documentation shows up.
- 01
A retrieval knowledge base built from your catalogue, pricing, prerequisites and exclusions
- 02
A configured agent in your channels, with grounded answering and a designed refusal path
- 03
A lead scoring and qualification model, with the routing rules that follow from it
- 04
Escalation to your team carrying full history and qualification state
- 05
An adversarial test set — the questions designed to make it invent an answer — with results
- 06
Reporting on answered, refused, escalated, qualified and converted
Invariant
An agent that will not say "I do not know" is not a sales asset. It is a liability with a fast response time.
- Will our customers know they are talking to software?
- Ours are configured never to deny it and to say so on request. That is a commercial position before it is an ethical one: a sale that depends on the buyer being wrong about who they spoke to is a refund and a review waiting to happen, and in a regulated or professional category it is worse than that. What we optimise for is that the question stops being interesting — in one engagement the operator's read of the logs was that roughly 95% of contacts never raised it, which is an assessment rather than a study, and we report it as one.
- How is this different from the chatbot we already tried?
- A widget answers from a model and a prompt. This answers from your material, and refuses when your material does not cover the question. Most of the engagements we are called into begin with switching off something that was confidently wrong — and the fix is almost never a better model. It is that nobody built the knowledge base.
- Will it make things up?
- That is the risk and it is the thing we test for rather than promise about. Answers are grounded in retrieved source, refusal is a designed path, and we build an adversarial set specifically to provoke invention — the questions your material does not cover, the near-misses, the ones with a plausible wrong answer. You get the results, including the failures.
- We do not have our product written down anywhere. Can you still do this?
- Then that is the first half of the engagement and we will say so in the estimate. Extracting what the product actually is from the people who hold it in their heads is work we do anyway — it is the same extraction the meaning and naming practice runs, and it produces something useful whether or not an agent is ever deployed.
- Does this replace our sales team?
- No, and we would not sell it that way. It replaces the eleven-hour delay and the fourth repetition of the same answer. The people are still what closes a four-figure or six-figure sale — they just arrive at the conversation already knowing who they are talking to and whether it is worth their afternoon.
- Which platform do you build on?
- The one you run, where it can carry the job. We have deployed on Respond.io and we work with the official WhatsApp Business API. We take no vendor commissions, so when a platform genuinely cannot do what you need, the recommendation to move is worth what it costs you to read it.
- How long does it take?
- Four to eight weeks for a first deployment in one channel, and the variable is never the software. It is how much of your product exists in writing on the day we start.
- Have you run this on your own money?
- On a client's, under a signed scope, and the record is published: a cold-traffic test where every enquiry was handled by an agent we configured, leads were scored rather than queued, and two thirds of the people who reached a session paid. The figures, the confidence interval and the limits are in the case record.
The hard ones — the price question, the edge case, the thing your product does not do. We will run them against whatever is answering your enquiries today and show you the transcript, then tell you honestly whether the constraint is the software or the fact that nobody has written your product down. That conversation is free and it is usually the whole diagnosis.