By KYN AI Advisory Team — AI implementation specialists, Singapore
To research a lead with an agent, give it a company name and website, a fixed list of sources, and one output: a short record with one checkable fact, a source link for each claim, and a flag for anything it could not confirm. Run in a single pass, this takes about 90 seconds per lead. The agent researches. A person decides whether the lead is worth contacting and approves any message.
The 90 seconds is a design target, not a promise. It depends on the number of sources, page size and model speed. Test it on your own leads.
What should the agent produce?
Decide the output before the method. A good lead record fits on one screen:
- Company: legal name, registration number, country, size band if public.
- What they do: one sentence, in the company's own words from its site.
- One checkable fact: something specific and recent, with the URL. A new office, a hiring page, a tender, a product launch.
- Fit: match or mismatch against your ideal customer rules, with the rule that fired.
- Contact candidates: role and name where the company publishes it.
- Unconfirmed: claims the agent could not verify.
The "checkable fact" is the design choice that matters. It forces the agent to cite. A salesperson clicks the link in five seconds and knows whether the line is true. This is what separates an agent from a template with a name merged in.
What are the five steps?
- Resolve the entity (10 seconds). Match the name to the registry record, so you research the right company. In Singapore, that is the ACRA record. A name like "Apex Trading" matches dozens of entities; the registration number removes the doubt. See company data sources for Southeast Asia.
- Read the website (20 seconds). Home, about, products, careers, news. The agent extracts what they sell, who they sell to, and the locations.
- Look for a trigger (30 seconds). Hiring posts, a new branch, a tender, a funding or leadership announcement from the company's own site or an official source. See buying signals for SMEs.
- Score fit (10 seconds). Apply your rules to the facts. Rules are written by you: country, size band, industry, systems in use.
- Write the record (20 seconds). Fill the template, attach the sources, list the unconfirmed items.
Total: about 90 seconds. Each step is limited so the whole run is bounded. Set a hard stop; if a site is slow, the agent writes "not confirmed" and moves on.
Which tools can do this?
Several approaches work, and the choice is about control, cost and data.
- A research table product. Clay offers Claygent, an AI research agent that runs web research inside a table of leads. Fast to start, and the data and steps live in the vendor's tool.
- A custom research agent. A small program that calls a model and a search or fetch tool, runs in your cloud account and writes to your CRM. More set-up, full control over the sources and rules. See Clay vs a custom research agent.
- A person with a checklist. Slower, 10 to 15 minutes per lead, and the benchmark to test against.
Whichever you pick, the quality rules below apply.
What keeps the agent honest?
An agent that researches will fill gaps with plausible text unless you forbid it. Four controls:
- Every claim carries a URL. No URL, no claim. The record rejects any sentence without a source.
- An "unconfirmed" field is mandatory. If the agent finds nothing, it says so. An empty trigger is a valid result.
- A freshness rule. A fact older than your limit, such as 12 months, is labelled as old, not recent.
- A spot check. A person reads 20 records a week and marks each claim true or false. Track the false share. If it rises, stop and fix the prompt or the sources.
For the first week, check every record. After that, sample.
What should the agent not do?
- Send anything. Research ends at the record. Outbound messages wait for a named person.
- Guess an email address. Use published contacts or a vendor with a documented source. See how to find business emails legally in Singapore.
- Collect personal data beyond what the task needs. A contact's role and public business details are enough.
- Treat a social profile as permission to message. Public is not the same as consented.
Singapore's PDPC PDPA overview sets the rules on personal data. Hong Kong, Indonesia and Malaysia have their own. Check the country of the person you are researching.
How do you measure whether it works?
Compare the agent with a person on 50 leads.
| Measure | How | |---|---| | Time per lead | Seconds from start to saved record | | Cost per lead | Model and tool cost per run | | Fact accuracy | Share of claims a reviewer marks true | | Usable trigger | Share of records with a confirmed, dated fact | | Reviewer time | Seconds to read and approve a record |
If accuracy is below what you would accept from a new hire, do not scale it. Fix the source list first. Most errors come from the wrong entity or an old page.
What does a worked example look like?
Take an illustrative case: a 40-person logistics firm in Singapore that sells to freight forwarders. The numbers below are illustrative, not a client result.
| Step | Output | Time | |---|---|---| | Resolve entity | Registry record, UEN, status live | 8 seconds | | Read website | Offers cold-chain trucking, two depots | 22 seconds | | Trigger | Careers page lists a fleet coordinator role, posted this month | 28 seconds | | Fit | Matches 4 of 5 rules; size band unconfirmed | 9 seconds | | Record | One-screen summary, 3 source links, 1 unconfirmed item | 21 seconds |
The salesperson reads the record, clicks the careers link, sees the post is real, and decides to write. The first line of the message would be: "Saw your fleet coordinator role; are you adding routes?" It cites something the reader can check.
Now the failure case. The agent finds no trigger. The record says "No trigger found" and fit is a match. That lead goes to a nurture list, not to a rushed message with an invented reason.
Edge cases to write into the rules:
- Two registry matches: stop and flag, do not pick one.
- Website in a language the agent reads poorly: mark the fields unconfirmed.
- Site blocks automated access: respect it and record "not read".
Where KYN fits
KYN builds research and outreach agents that run in the client's own cloud account. For an insurance brokerage, outreach carries one checkable first line at about S$0.02 per email, taken from a cited source. Agents use their own restricted accounts, 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. See sales systems and marketing systems.
Frequently asked questions
How accurate is AI lead research?
It depends on the sources and the controls. With a mandatory URL per claim and a spot check by a person, errors are visible and fixable. Without those, plausible but false statements appear. Measure the false share on a sample before you scale.
Is it legal to use an agent to research leads in Singapore?
Researching a company from public pages is routine. Collecting and using personal data is governed by the PDPA, and marketing messages are governed by the Spam Control Act and the Do Not Call rules. Keep the record to business details and check the rules before any message goes out.
How many leads can an agent research in a day?
Volume is limited by cost and by the sites it reads, not by the agent. At about 90 seconds a lead, one worker finishes roughly 40 an hour. Parallel runs raise that, but a person must still review the records.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
