Poldermarketing
How can AI automate lead research and outreach for a startup?
AI can build the list, read public pages, draft from one checked fact, and log the CRM. Do not let it invent an email, choose the buyer, or send alone. Define the ideal customer and the disqualifiers, verify addresses in a data source, and approve early messages. Measure replies and meetings. Raise volume only after one segment works.
Automate the file, not the conversation
AI can filter accounts, read public pages into fixed fields, draft from one checked fact, and log the CRM. It should not choose the buyer, invent a contact, or hold the conversation. Automate a step only when a wrong answer is obvious, the step repeats, and someone can stop it. See automating repetitive marketing work and when marketing automation makes sense.
Freeze one buyer, then one source
Write the buyer first: industry, size band, country, role, the problem, and at least five disqualifiers. A person signs the filters before the search. Have the agent show its derived filters, and correct them before the search. Flag a borderline company instead of silently keeping or dropping it. A second source may be companies like ones you already closed. Prefer a dated public trigger over a static export. Pull companies before people.
Take the address from a database, or a waterfall that stops at the first verified hit, and name the provider. Hunter’s 2026 test of a dozen models found 76% of the addresses invalid. Verify again if the list sat for more than a few weeks. Skip catch-all domains in the first batch. After that, keep dropping them, or send them only from a separate mailbox. Google says not to buy addresses.
Can AI research the company before outreach?
Yes. AI can automatically research a company before sales outreach only when it quotes a page you named, or leaves the field blank.
| Field | Pass | Fail |
|---|---|---|
| What they sell | One sentence from the site, with the URL | A guessed category |
| Who they sell to | Businesses, consumers, or both, from the site | A guess from the logo |
| Buyer | One role that matches the persona | Every director |
| Timing | A dated public event with a link, or a blank | Funding with no source |
| Open role | The posting names the problem you solve | Any job ad |
| Public post or talk | Linked by URL, or a blank | Paraphrased with no link |
| Stack or size | Named provider, or a blank | A model estimate |
| Verified, provider named | Written by the model |
If the fact is not on the linked page, or not from the named provider, nothing sends. Research does not enroll a sequence.
Can AI research accounts for founder-led outbound?
Yes. AI can automate account research for founder-led outbound.
Six lines, then stop: fit against the buyer you signed, one dated public fact with a URL or a blank, the role, a hypothesis marked as a hypothesis, the evidence link, and one sentence only the founder would say. Personalize outbound without manual research uses that fact, not a first name. The founder signs the buyer, approves early notes, and takes the call.
How do you automate outreach without an automation team?
You can automate personalized outreach without building an internal automation team. One owner writes the buyer, the source, the verification rule, the prompt, who reads replies, and the stops.
A person approves early batches. LinkedIn forbids scrapers and bots that add contacts or send messages, and it can restrict or shut down the account. If you will not do the weekly review, hire one person for a bounded build or a few days a week.
How can AI enrich a new inbound lead?
AI can automatically enrich a new inbound lead by filling blanks only. Attach a known domain to the existing account, and do not overwrite a value a person set.
| Check | Pass | Fail |
|---|---|---|
| Address they typed | Kept | Replaced by a model |
| Company | Domain match, or blank | A guessed employer |
| Fit | Match, no match, or uncertain | A score with no closed deals behind it |
| Route | A person, a review queue, or no mail | A cold sequence |
A page view is not a match. Country rules and suppression still apply before any marketing mail.
Keep fit and timing on two lines
Fit is the buyer you signed. Timing is a dated public reason to talk now. Do not merge them into one score. Fit with no dated reason waits. A dated reason on a row that fails a disqualifier is a drop. A score with no written reason does not enroll.
One checked fact, then a person
The opener is one checked detail, one problem, and one question. Stop on a reply, an opt-out, a hard bounce, or a meeting. Log each of those in the CRM and the email tool, and check suppression before every import.
Judge a tool by the job it fails
Judge a tool by the job it fails: inventing an address, sending with no review, or treating a page view as fit. Hunter splits the market into contact data, research, and outreach, or all three bundled. Add the CRM. Test each job alone, even inside a bundle.
As a worked example, HubSpot’s prospecting agent watches signals you set, finds up to three contacts per company, and drafts from CRM data. Generating the email uses credits. Leave autonomous sending off, and leave enrollment on a page view off, unless you signed them. The agent still counts page views, form fills, calls, meetings, notes, and email opens from the past year as engagement. That history is not a setting or a buying signal you approved. Only one external data source connects at a time. If you will not own the weekly stop, see lead generation and the system under it.
A lawful email can still miss the inbox
For personal Gmail accounts, Google requires SPF or DKIM, matching forward and reverse DNS, and TLS. Above 5,000 messages a day to Gmail accounts, also use SPF and DKIM together, DMARC aligned to the From domain, and one-click unsubscribe on marketing and subscribed mail. Keep the Postmaster Tools spam rate below 0.10%, and do not reach 0.30%. Google does not track open rates and cannot verify open rates other tools report. Do not fake a Re: or Fwd: subject.
Split the list by the recipient
The rule follows the recipient, so split rows by country before you enroll them. A country not covered here needs its own rule first. Canada, for one, has its own consent law, CASL, so a GDPR legitimate interest does not carry over.
In the United States, the FTC CAN-SPAM guide covers every commercial message, business-to-business included, at any volume: honest headers, a subject that matches the body, a clear ad disclosure, and a valid postal address. The opt-out must work for 30 days, be honored within 10 business days, stop all marketing, and ask no fee, no extra data, and no step beyond a reply or one page. Never sell or transfer an opted-out address. The company promoted and the sender are both responsible.
If you offer goods or services to a person in the EU, the GDPR applies even when you are not established there. They may object to direct marketing at any time, and then you stop. That ground does not authorize the email. Directive 2002/58/EC requires prior consent for electronic mail to a subscriber who is a natural person. The exception: your own customer’s address from a sale, your own similar products, and a free easy objection at collection and in every message, unless refused at the start. Member states protect other subscribers, so one EU list is not one rule.
The Autoriteit Persoonsgegevens requires advance consent before you email a company, unless that company consciously published the address for this purpose. If the company is in a country outside the EEA, follow that country’s rules instead. Give an easy free objection when you collect the data and in every message, and stop if they object.
UWG section 7 requires prior express consent for advertising by electronic mail in Germany, including a business address. The exception: an address from a sale, the seller’s own similar goods or services, no objection, and a clear notice at collection and in every message that they may object at any time without a fee from you.
In the UK, the ICO treats an employee address at a company, an LLP, or another corporate body as the organisation. That mail needs no PECR consent. Identify yourself, give a valid opt-out address, and honor a corporate opt-out even though PECR does not require it. If the address identifies a person, the right to object is absolute and you stop. Sole traders and some partnerships need consent, or the soft opt-in: the customer rule above, which here also counts a negotiation of a sale. If unsure, treat the row as an individual. Social direct messages count as electronic mail. The guidance is under review because of the Data (Use and Access) Act.
Four weeks, with a stop at each gate
Four weeks, and stop at the first failed gate. An inbound lead does not join this batch. Turn the experiment into a workflow only after real replies.
- Week one: five disqualifiers, one source, about one hundred companies. Stop if a field is unverified or model-filled.
- Week two: sample the research. Stop if the fact is not on the linked page.
- Week three: a small reviewed batch, off the customer domain, with SPF or DKIM, a postal address, and a working opt-out already on the message. Stop on a wrong fact, a bounce, or a spam complaint.
- Week four: keep real replies. Raise volume only after the loop has run a week on its own.
How I would set this up
I would not send until the buyer, the source, the blank-field rule, and the reply owner are written down. The first batch stays small and stops at the first bad gate.
I work alone as Poldermarketing, an AI-native growth marketer, remote, in Dutch and English, for a bounded build or a few days a week. For a freshly funded startup anywhere, the offer is an all-in GTM solution: one freelance marketer for marketing, AI, and automation, fully remote, from the first message to the first customers, without a separate specialist for every part. I am strong in AI, content, automation, and workflows, and I build as well as plan. Google Ads and Meta Ads are newer for me: I set them up and review them, and I do not scale a large program on either.
Start with the free growth scan, then how I work and positioning and messaging.
Questions people ask
Can AI replace a salesperson at a startup?
No. Use it to filter accounts, read public pages, draft from a checked fact, time a follow-up, and log the CRM. A person still signs the ideal customer, drops a bad-fit account, edits what a prospect will read in the early batches, and answers a reply that is a question rather than a yes. A calendar link fired at an objection ends the thread. The output you want is a conversation, not a larger send report.
Should the model look up the email address?
No. Take the address from a database or a waterfall that stops at the first verified hit, and store the provider. A model asked to fill a blank will invent one. Hunter's 2026 test of a dozen leading models found 76% of the addresses they returned were invalid. The same write-up says one model returned a working address on 19.5% of asks alone, and on 78.1% with Hunter's own data tools attached. Drop any address with no source.
Is cold email to a work address allowed?
It depends on the recipient, not on the tool. In the United States, commercial email includes business mail and must meet the CAN-SPAM rules on identity, a postal address, and a working opt-out. In the EU, a GDPR legitimate interest does not by itself allow the send, and a company address follows national law. In the UK, a company subscriber and a sole trader are different tests. Split the list by country. One sequence cannot carry every regime.
What should you measure instead of opens?
Qualified replies, meetings held, and customers won, split by list source and by message. Google does not track open rates and says it cannot verify open rates reported by other tools, so an open is a weak success signal. Stop a sequence on a reply, an opt-out, a hard bounce, or a booked meeting. If a segment produces no real replies, end it before you raise volume.
When should a startup leave this manual?
Leave it manual when the offer still changes every week, when nobody will read replies this week, or when you cannot name a bad fit. A new list only sends the unclear sentence faster. The same stop applies to an inbound score you have not checked on closed deals.