TL;DR: the "Chat with my AI" page on mathieuhaye.fr lets a visitor describe their project to an AI agent that replies instantly, at any hour, then steers the conversation toward a call or an email. The case for that agent rests on two figures published by Harvard Business Review: a company that contacts a lead within an hour is nearly 7 times more likely to qualify it, and 23% of companies never respond at all.
My website hosts a public AI agent that handles the first conversation with any visitor who wants one. This article covers what it actually does, what it is forbidden to do, and the cases where I would advise against installing one.
Key takeaways
- According to a Harvard Business Review study of 2,241 US companies, 23% never respond to an inbound lead, and the average response time among the rest is 42 hours.
- A company that contacts a lead within an hour is nearly 7 times more likely to qualify it than one that waits 24 hours or more, per the same study.
- Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
- The public agent on mathieuhaye.fr works because its scope is narrow: it qualifies, it routes, and it promises nothing; every commitment goes through a human.
The problem: an inbound lead ages fast
The finding is old and has not moved. In March 2011, Harvard Business Review published "The Short Life of Online Sales Leads", a study by James Oldroyd, Kristina McElheran and David Elkington that audited 2,241 US companies by sending each a test lead through its own website. The results: 37% responded within an hour, 24% took more than 24 hours, and 23% never responded at all. The average response time, among companies that eventually replied, was 42 hours.
The number that matters to a business owner is this one: firms contacting the lead within an hour were nearly 7 times more likely to qualify it than those that waited, and over 60 times more likely than those that let 24 hours pass. An inbound lead does not wait; it cools down, then it shops elsewhere.
My situation is the one shared by thousands of independents and small firms: enquiries arrive whenever they please. A business owner comparing vendors often does it in the evening or over the weekend, precisely when nobody answers. A contact form records the enquiry; it does not move it forward. That is the gap the agent fills.
What exactly does the agent on mathieuhaye.fr do?
The Chat with my AI page invites the visitor to talk directly to an AI version of me, powered by Claude, Anthropic's model. The visitor describes their project in a few sentences: a freelance assignment, a CRM or automation brief, a partnership idea, or plain curiosity. The agent replies instantly, free, with no sign-up.
Its job fits in three verbs. It listens: it asks the questions a first call would ask, about context, team, and the process that hurts. It positions: it knows my published projects and services, and can say honestly whether the enquiry matches what I do, or not. It routes: when the conversation warrants it, it offers the logical next step, a 30-minute call or an email. The page says so plainly: if the project fits, the discussion continues off the page.
And behind the agent there is a human: I read every contact request myself. The agent holds the first conversation in my place at 11pm on a Saturday; it does not replace me in the relationship. That nuance is decisive: the agent cuts the first-response delay from hours to seconds, but final qualification and any commitment remain human.
The guardrails: what a public agent is not allowed to do
An agent that speaks in your name on a public page is as much a risk as an asset. Gartner estimates that over 40% of agentic AI projects will be canceled by the end of 2027, and points at three causes: escalating costs, business value that was never defined, and inadequate risk controls. The firm also describes "agent washing": chatbots and assistants rebranded as "agents" without real capabilities; out of thousands of vendors claiming the label, Gartner counted only about 130 credible ones in 2025.
The remedy is not technical, it is contractual: you define what the agent is forbidden to do before defining what it can do. Mine follows simple rules. It quotes no firm prices; terms are discussed with me. It makes no commitments on behalf of the business. It asks for no sensitive data and performs no irreversible action: no emails sent in your name, no writes into third-party systems. When it does not know, it says so and hands over. Every conversation leaves a log I can review.
These rules mirror what I set up for clients: a visible door to a human is not a concession, it is the condition for the agent to earn trust. An agent that promises everything scares off good prospects and attracts the wrong ones.
When a public AI agent is a bad idea
The honest answer: in most cases, start without one. If your site gets a handful of enquiries per month, a short form and an alert on your phone beat any agent: the speed problem is solved by discipline, not software. If your visitors always ask the same three questions, a well-written FAQ answers them at zero cost and, as a bonus, gets cited by generative engines. And if nobody on your team will ever reread the conversations, the agent becomes one more blind spot.
The table below sums up when each option is enough.
| Option | What it does well | Where it fails | Enough when |
|---|---|---|---|
| Contact form | Reliable capture, zero cost | No answer until a human reads it | Few enquiries, responsive team |
| Static FAQ | Answers recurring questions, good SEO | Does not qualify, does not converse | Enquiries all look alike |
| Direct booking link | Zero friction to the call | Unqualified calls in your calendar | Your call time is not saturated |
| Public AI agent | Instant 24/7 reply, qualification, routing | Real risk without guardrails and review | Steady volume, varied enquiries, speed matters |
An agent earns its place when three conditions stack up: enough volume that off-hours cost you leads, enquiries varied enough that a FAQ falls short, and someone to review the conversations and improve the agent. It is the same discernment as for a CRM or an automation: custom software is justified by the need, not by the trend.
Frequently asked questions
Can an AI agent replace a website's contact form?
No, it complements it. The form remains the most reliable channel for capturing an enquiry, and some visitors prefer writing without a dialogue. The agent adds an instant reply and qualification for those who want to move now, which matters given that the average company response time to a lead is 42 hours according to Harvard Business Review.
What should a public AI agent be forbidden to do?
Four things at minimum: quoting firm prices, making commitments on behalf of the business, collecting sensitive data, and performing irreversible actions such as sending emails or writing into third-party systems. Every conversation should leave a reviewable log, and the agent should offer a visible human handover at all times.
How much does a lead-qualification AI agent cost for an SME?
It depends on the scope: data sources to connect, channels, guardrails. On my side, a first narrow scope is handled as a sprint from €400 excl. VAT; a full agent is quoted after a 30-minute scoping call. The best starting point remains a single measurable task, not an "assistant that does everything".
The simplest test requires no tooling: send an enquiry to your own company through your website on a Saturday at 9pm, and time the first useful reply. If the result bothers you, the topic is worth 30 minutes. I build this kind of AI agent for SMEs, with the same guardrails as the one you can try on this site; we can discuss it on a free scoping call.
The takeaway: a public AI agent is not worth what it can say, but what it is forbidden to do. Mine qualifies and routes; it does not sell, promise or sign. That is exactly why it works.