An AI Sales Manager That Finds Leads and Builds Demo Sites: How It Works and Where It Breaks
An AI sales agent can find local businesses with weak websites, build each one a sample site, and prepare a briefing so someone can reach out with something concrete. It works when targeting is narrow and every demo is checked. It breaks on consent rules for small businesses, deliverability, factual errors in demos, and unreviewed sending.
What the setup in the video does
The video behind this topic comes from creator Tommy Yip, as day four of a build series sponsored by Hostinger. According to the video's own description, he runs Hermes Agent, an open-source agent from Nous Research, on a virtual private server, then turns it into an "AI sales manager" named Kush. The agent researches businesses with weak websites, builds a WordPress demo page for each one on the server (one Docker project per site), and drops a briefing document into Discord. The description says the creator then does the outreach himself.
That last detail matters more than the title. The agent prepares, and a person decides who gets contacted and what they receive. That division of labour is the safest version of this workflow, and it is the one this guide builds on.
The workflow, step by step
Whatever agent framework you use, a demo-site prospecting loop has the same parts.
1. Define the target
Pick one trade and one area, such as dentists or landscapers in one city, and write down what "weak website" means for you: no mobile layout, no online booking, no HTTPS, broken pages, or no site at all. Vague criteria produce vague leads.
2. Discover businesses
The agent pulls candidates from public business listings and maps data, then removes chains, franchises, and anyone already in your CRM or on your suppression list.
3. Audit and score
For each site, the agent records specific, checkable problems and scores fit against your criteria. Every claim it will later make in an email should come from this step and be verifiable by a person in under a minute.
4. Build the demo
The agent fills a template with the business name, services, and area taken from public sources, using generic or licensed imagery. One template per trade keeps quality predictable.
5. Brief a person
The briefing lists what is wrong with the current site, what the demo fixes, the contact route found, and the region the business operates in, so the reviewer can judge both quality and whether contact is allowed.
6. Approve, send, and log
A person checks the demo, edits the draft, and releases the message. Replies, opt-outs, and outcomes are logged to the lead so the next batch improves.
Why it works when it works
A demo turns a pitch into evidence. Instead of claiming you can improve someone's website, you show them their own business on a better one. For owners who have never thought hard about their site, that is easier to respond to than a paragraph of benefits. It also forces the seller to do research, which tends to improve targeting. The cost is that it is spec work: every demo consumes compute, hosting, and review time, so the economics depend on how many demos turn into held meetings, not on how many the agent can produce.
Where it breaks
The failure points are predictable, and most of them get worse as volume rises.
Consent and the law
In the US, the FTC's CAN-SPAM guidance applies to business email too: no misleading headers or subject lines, a clear identification that the message is an advertisement, a valid postal address, and an opt-out honoured within 10 business days. In the UK, the ICO says you may email corporate bodies, but sole traders and some partnerships are treated as individuals and need specific consent or a previous customer relationship. A single-owner local business is often exactly that. In Canada, the CRTC's guidance on CASL allows implied consent only in narrow cases, such as a published business address where the message relates to the person's role. Record the region and business type for every lead. This is not legal advice.
Quality
Models fill gaps confidently. A demo can show the wrong opening hours, services the business does not offer, invented reviews, or placeholder text, and an email can assert a problem with their site that is not real. One wrong detail tells the owner the outreach was automated and careless.
Their brand on your server
Copying a logo, photos, or written copy into a page the business did not commission can raise copyright and trademark questions, and a public, indexable page in their name can look like impersonation. Keep demos unlisted, block search indexing, label them as concepts, and delete them on request or after a fixed period.
Deliverability
Cold email from a new domain, sent at volume, linking to freshly created pages, is a pattern spam filters watch for. Use a separate sending domain with authentication set up, warm it gradually, keep daily volume low, and host demos on one stable domain with HTTPS rather than raw server addresses.
An always-on agent with deploy rights
An agent that browses arbitrary websites and can deploy code on a server is reading untrusted content with real permissions. Treat scraped pages as untrusted input, run the agent in an isolated environment with only the permissions it needs, and do not give it unattended access to your mailbox.
Where human approval belongs
Put people at the decisions that are expensive to reverse: approving the target criteria, checking each demo before its link goes out, releasing first-touch messages, and handling every reply. Let the agent own the repetitive parts: discovery, audits, builds, briefings, and reminders. When drafts and demos pass review consistently for several weeks, raise the agent's autonomy one step, with a daily cap, rather than switching it to fully automatic.
What to measure
Track demos built, demos that pass review on the first try, messages sent, replies, positive replies, meetings held, and opt-outs and complaints. Divide total cost, including the time people spend reviewing, by held meetings. If the review pass rate is low, fix the templates and audit criteria before sending more.
Running it inside a CRM
Keeping the agent's leads, drafts, and sends in the same system as everything else makes the approval step real rather than a promise. In Koryo, an agent is an operator with its own identity, permissions, and daily cap; new agents default to send with approval and 100 messages a day. Every message, from a person or an agent, passes one sending gate that checks permissions, permanent suppression, consent and region, required content, and capacity, and the product's default rule asks for recorded opt-in or an existing-customer basis before cold mail to the EU and EEA, UK, Switzerland, and Canada. Those are product settings, stricter in places than the law requires.
Frequently asked questions
- Is it legal to send a local business an unsolicited demo website?
- It depends on where the recipient is and what kind of business it is. In the US, CAN-SPAM allows commercial email without prior consent if it is honest, carries a postal address and an opt-out that is honoured. In the UK, the ICO treats sole traders and some partnerships as individuals who need consent. Check each region before sending. This is not legal advice.
- Should the AI agent send the outreach email itself?
- Not at the start. Let the agent research, build the demo, and draft the message, then have a person check the demo for errors and approve the first send. Once the drafts are consistently accurate, an agent can send within a daily cap, but every message should still pass the same suppression, consent, and content checks.
- Can a demo site use the business's logo and photos?
- Copying a business's logo, photos, or written copy into a site it did not commission can raise copyright and trademark questions, and it can look like impersonation if the page is public. Safer practice is to use the business name and public facts, generic or licensed imagery, a clear concept label, and to delete the page whenever asked.
- How many demo sites should an agent build each day?
- No more than a person can review properly and no more than your mailboxes can send safely. Every demo needs a factual check, and a new sending domain should ramp slowly. A small batch that is accurate and well targeted will usually produce more held meetings than a large batch with errors, and it protects the domain.