By KYN AI Advisory Team — AI implementation specialists, Singapore
Our lead-research agent reads a public source about a company, writes one fact it can cite, and drafts a first line from that fact. It costs about S$0.02 per email on our live outreach system. It never sends. A named person approves every message that goes out. This article covers the build, the unit cost and what the number leaves out. For monthly cost at the system level, see AI sales agent cost per month for a small business.
What does the agent do?
Four steps, in order.
- Take an account. The input is a company name and a website from a list you already hold, or from a registry search you ran.
- Read a source. The agent opens the company's own site, a public registry entry or a news page, and extracts what is stated there. It stores the URL and the exact passage.
- Write one checkable fact. For example: "Your site lists a Jakarta showroom and a Surabaya warehouse." The fact must point at the stored passage. If no passage supports it, the agent writes nothing.
- Draft the first line and the email. The first line states the fact. The body states your offer, using wording you approved. The draft goes to a review queue.
The rule that matters is step 3. A fact with no source is dropped. This is the difference between personalisation a reader can check and flattery a reader can smell.
Why only one fact?
Because every extra claim is another chance to be wrong, and a wrong claim in the first line ends the conversation. One claim also fits the review: the reviewer clicks the source link, sees the passage, and approves in seconds.
We also keep the first line free of inference. "You are expanding" is an inference. "Your site lists a second warehouse" is a fact. The agent writes the second form only.
What does it cost per lead?
On our live outreach system, the cost is about S$0.02 per email. That figure covers the model calls to read the source, write the fact and draft the email. It depends on:
- Page length. A long page costs more to read. We trim pages to the relevant sections before reading.
- Model choice. A smaller model reads and extracts; the same model drafts. Using a larger model for every step raises the cost several times with little gain on this task.
- Retries. A failed fetch or a draft that fails the source check is retried once. A second failure drops the lead into a manual list.
- Language. Drafts in two languages cost more than one.
Here is how the S$0.02 sits against the other costs of a lead.
| Cost | Typical nature | Order of size | |---|---|---| | Model calls for research and draft | Per email | About S$0.02 | | Data source (registry, enrichment) | Per record or subscription | Set by the provider; check their price page | | Sending infrastructure | Per mailbox, per month | Set by the provider | | Reviewer time | Per approved batch | Minutes per batch | | Build and retainer | Fixed | Quoted after the review |
The model cost is the smallest line. The cost that grows with volume is reviewer time and the data source. Do not judge an agent by the S$0.02 alone. Judge it by cost per approved, sent message, and later by cost per reply and per meeting.
How do we keep it from being wrong?
- Source or silence. No stored passage, no fact.
- Freshness check. The agent records when it read the page. Facts older than a set age go back for re-reading before use.
- Duplicate check. A company already in the CRM, or already contacted in a set period, is skipped.
- Suppression list. Any address that unsubscribed or bounced is excluded before drafting, so we do not pay to write to it.
- Wrong-company check. Names collide. The agent checks that the site's registered entity or address matches the account record, and flags a mismatch.
A labelled test set, built from real accounts, scores these steps before launch. See how to evaluate AI output.
What does the agent not do?
It does not send. Outbound messages wait for a named person, and the review screen shows the draft, the fact and its source side by side.
It does not set price or promise terms. The offer text comes from a template you approved.
It does not scrape what it is not allowed to read. The sources are public pages and registries you may use, and the agent honours the rules of each site. Where you hold a legal duty on marketing messages, such as unsubscribe handling, the rule is built in as a step. For the sending side, see is it safe to let an AI agent send emails to customers.
It does not read hostile instructions as commands. A page can contain text aimed at an agent. The agent treats fetched text as data, not as instructions, and has narrow permissions so a trick cannot reach anything valuable. See prompt injection in business AI agents.
How does it connect to the rest of sales?
The output lands in the CRM with the actor, time, source and draft. Replies are read by a second agent that sorts them: interested, not now, wrong person, unsubscribe, question. That step costs about S$0.01 per reply on our live system. A person handles anything that needs judgement. The design of the whole flow is in our sales systems page.
For an insurance brokerage, we built exactly this: outreach with one checkable first line, a reply desk, and a group benefits comparison that turns into a client PDF at about S$0.03 per comparison.
What are the trade-offs?
- Quality over volume. One checked fact per email limits how fast you can send. That is deliberate. A shorter list of accurate messages usually beats a long list of generic ones, and it protects your sending reputation.
- Source coverage. Companies with thin websites yield no fact. Those go to a manual list or a different channel.
- Review time. Approval is quick per message, but it is a daily task someone owns.
Where KYN fits
KYN builds the research agent inside a sales system in your own cloud account. The agent has its own restricted account, every action is logged with the actor, and outbound messages wait for a named person. The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. Pricing is a one-off build fee and a monthly retainer of at least twelve months, quoted after the review. See pricing.
Frequently asked questions
How much does it cost to research one lead with AI?
On our live outreach system, about S$0.02 per email for the model calls that read a source, write a checkable fact and draft the message. Data subscriptions, sending tools and reviewer time are separate.
Can the agent send the emails itself?
It can be built to, but we do not. Outbound messages carry your name, so a named person approves each batch after seeing the draft and its source.
What if the agent cannot find a fact?
It writes nothing and moves the account to a manual list. A message without a checkable first line is the generic outreach the agent exists to avoid.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
