Industry · AI automation

Lead generation for AI automation agencies

GrowthPal runs outbound for firms that implement automation for other businesses: bespoke build, integration and maintenance work rather than a product. We settle one niche first, because a niche is what gives you a directory, a conference, a vocabulary and a transferable reference. Angles name a process and a cost the buyer already tracks. Best fit is past 3M in ARR, an offer people already buy, someone ready to take calls, and roughly 30,000 reachable contacts.

The buyer

Who you are really selling to

Role	What they care about	Where they pay attention	Best channel
Owner, Managing Director or General Manager	Headcount cost, whether the business can take more volume without hiring, and whether a vendor becomes one more thing to manage.	Peers in the same trade, their association newsletter, and the suppliers they already use. Not technology media.	Short email read on a phone between interruptions. One ask. The first two lines decide it, and a technology-led opener loses both of them.
COO or VP of Operations	Cycle time, throughput, error rate and exception volume. Owns the process that hurts and has usually already tried a spreadsheet, a hire or a consultant.	Their own operations managers, and operators in the same vertical.	Email first, with a phone follow up that works better here than with marketing roles. Name a process or do not write.
Controller, Finance Director or Head of Shared Services	Month-end close duration, invoice and payment exception handling, audit trail, and headcount in back-office roles.	Accounting press for their vertical, their audit firm, and the software vendors they already pay.	Email with a number they already track in the first line. They will ask about the exception path before the happy path.
Function head in claims, order entry, dispatch, intake or billing	Getting through the queue. Feels the pain daily and holds no budget.	The work itself, and the vendor reps who already call them.	Email that supplies something written to be forwarded upward. Treat them as the champion, not the signer.
IT lead or Head of Business Systems	The system of record, credentials, API access, security review, and who maintains the build after handover.	Vendor documentation, the implementation partner they already use.	Email that answers the handover question before it is asked. Rarely initiates, can stop anything.

Constraints

What makes outbound different here

The technology word has stopped carrying information

Operators in this audience have been pitched automation heavily, and an opener that leads with the technology reads as one more vendor email. No credible public measurement of that fatigue exists: every figure we found traced back to a vendor selling a cure, so we do not quote one. The adjacent evidence is weaker but real. The TrustRadius 2026 B2B Buying Disconnect report, run in January 2026 across 1,862 technology buyers and 444 vendors, found 47 percent of buyers trust online resources less than a year earlier, and ranked vendor marketing collateral last among the resources buyers consult.

What it decides: The first line. It names a process and a cost the buyer already tracks, such as close duration or quote turnaround. The technology appears only after the buyer has agreed the problem exists, and sometimes not at all.

Bespoke work cannot be demonstrated

There is no product demo, no free trial and no repeatable benchmark, because no two engagements are alike. The proof that does travel is a named process in the buyer's own vocabulary, a before and after of one workflow with step count and cycle time, a reference in the same vertical, and an honest account of what the maintenance burden turned out to be. That proof belongs to you. GrowthPal cannot supply it, and no outbound agency can.

What it decides: What gets assembled in week one. Each segment needs one process story ready before the first send, and any client name or figure in it needs written permission first.

One niche, or no list

A firm that automates anything for anyone has no list source. A niche produces four things at once: a member directory or licensing register, an association whose calendar carries dated events, a conference with a published exhibitor list, and a vocabulary of five or six process nouns that insiders use. It also makes one process story transferable to every other row. The honest caveat is size. A niche of a few thousand companies is too small for a system, and a short hand-run list will beat one.

What it decides: Sequence of work. The niche is settled before a row is built, and the list is sourced from that niche's own surfaces rather than from a headcount and industry-code filter.

A manual process is mostly invisible from outside

Nothing public states how many orders a company rekeys each week. Repeated clerical postings, job specs that list the systems a person sits between, headcount growth with no systems hire, review text about waiting weeks, and published process pages narrow the field. None of them confirm. A repost can mean a failed hire. A job spec goes stale. Reviews over-represent the angry.

What it decides: The grammar of the angle. It asks a question the buyer can correct rather than asserting a fact about their business, because asserting a process you cannot see is the fastest way to be wrong in public.

The buyer is taking on a maintenance liability

A custom build does not end at handover. It sits inside systems that change, and the buyer knows whoever built it will eventually move on. The fear is not that the work will fail. It is being left with something nobody in the building understands.

What it decides: Sequence content. Handover and break-fix go in early and in writing: who owns the build, where the runbook lives, what a change request looks like. Holding that back until contracting loses deals at the last step.

Signals

Buying triggers, and the angle each one earns

Trigger	How we find it	The angle it earns	Proof the email needs
The same clerical role posted again	The company's own applicant tracking board. Greenhouse documents a public board at boards.greenhouse.io plus a company token, readable without an account; Lever, Ashby and Workable publish equivalents. Repost dates and open counts show on the board itself.	A role posted three times in a year is a process the company has decided to solve with people.	The same process, named, rebuilt in their vertical, with step count before and after.
A job description that names the systems	The requirements section of the same public boards, which routinely lists the ERP, the CRM, the customer portals and the spreadsheets a person is expected to work between.	The shadow spreadsheet sitting between two systems of record is usually where the week goes.	A documented integration surface for those exact systems, including whether the work is portal bound and what happens where there is no API.
Headcount growth in a repetitive function with no systems hire	Job boards read over time, plus company headcount and title mix on LinkedIn, read manually. LinkedIn's User Agreement prohibits scrapers and automation extensions, so it is a place to verify, never a place to harvest.	Capacity is being bought with people, which is visible from outside and uncomfortable from inside.	A throughput comparison on one named process, with the exception path described.
A system implementation, migration or go-live	Implementation partner case studies and vendor partner directories. UiPath's Find a Partner filters by country, tier and specialization. Make and n8n publish partner listings. A job post for an administrator of a named system is the cheapest confirmation.	Manual workarounds form around a new system within months and then harden into the process.	What usually survives a go-live and what usually does not, written for someone who has just lived through one.
A public complaint pattern about turnaround time	Google reviews, Trustpilot, Better Business Bureau complaint records and vertical forums, read for phrases such as still waiting, chased, or weeks. A pattern across months counts. Three bad reviews do not.	Slow turnaround in a repetitive function is usually a queue, and a queue is usually manual.	A quoting or intake workflow rebuilt in the same vertical, with cycle time before and after.
A published process page or team page	The company's own site. A how it works or what to expect page often describes a handoff chain in plain English, and a team page listing six coordinators by name says more than a database field.	The company has published its own handoff chain and can see it counted back to them.	A comparable chain collapsed, with what moved to exception handling and what stayed human.
Acquisition, merger or a second site	National company registries, Companies House in the UK and state registries in the US, press releases, and a careers page that starts listing the same role in two cities.	Two systems of record now disagree, and someone is doing manual reconciliation to make them agree.	A reconciliation or migration they can check, described in operational terms.
An automation that already broke	Job posts asking for maintenance of an existing integration, a listing on a vendor partner directory that has gone quiet, forum posts from staff asking for help with a named tool, or a contractor departure visible on a profile.	Somebody built this once and the person who understood it left.	A documented handover model: the runbook, the break-fix path, and who owns the build.

Peek inside

Example angles we would test

Examples of angles to race in this market, not results. Each one is a hypothesis with its proof assembled before a single send.

You have posted the same order entry role three times this year.

Proof it needs: The posting dates on their own board, plus one process rebuilt in their vertical with step count before and after.
The ask: Fifteen minutes on where that role's week actually goes.

Your job spec lists the ERP, two customer portals and a spreadsheet, and the spreadsheet is doing the real work.

Proof it needs: A documented integration surface for those systems, including the portal bound parts with no API.
The ask: Permission to send a one-page process map. No call.

You added people to the claims team this year and nobody to the systems that feed it.

Proof it needs: A throughput comparison on one named process, with the exception path spelled out.
The ask: A look at the intake queue.

Reviews from the last six months mention waiting weeks for a quote.

Proof it needs: A quoting workflow rebuilt in the same vertical, cycle time before and after.
The ask: A short written read of the complaint pattern, sent over.

Month-end close runs long because three systems get reconciled by hand first.

Proof it needs: A close timeline step by step, before and after, with the three-way match handled.
The ask: A review of their close checklist.

You acquired a second business and now two systems of record disagree with each other.

Proof it needs: A reconciliation they can check, with what was automated and what deliberately was not.
The ask: Thirty minutes of paid discovery scoping.

Your new system went live and the workarounds built around it have hardened into the process.

Proof it needs: What usually survives a go-live and what usually does not, written for an operations reader.
The ask: A post go-live process review.

Someone built an automation here already, and the person who understood it has left.

Proof it needs: A handover model: the runbook, the break-fix path, the change request route.
The ask: A look at what is still running and what quietly is not.

We work only with your trade, and we have built this same process four times.

Proof it needs: A named reference in the same vertical, in that trade's own vocabulary.
The ask: An introduction to the function head rather than a meeting with the owner.

Your own how it works page describes six handoffs between a request and an invoice.

Proof it needs: A comparable handoff chain collapsed, with touch time counted at each step.
The ask: Permission to mark up their published process and send it back.

The list

How the list gets built in this market

Every source below assumes a niche has already been chosen. Without one there is no directory to work from and no vocabulary to write in, and the list collapses back into a headcount filter.

The niche's own surfaces

Trade association member directories for the chosen vertical, such as the American Staffing Association member directory for staffing firms or the American Land Title Association membership directory for title agencies. Read the terms of use first: many associations prohibit extraction for marketing, so these name companies rather than supply contact data.
Licensing and regulatory registers, where the vertical is licensed and the register is public.
Conference and trade show exhibitor, sponsor and attendee lists, which carry a date as well as a name.
Franchise and multi-location operator lists, read from the operator's own locations page or sitemap.
US Census County Business Patterns, for establishment counts by industry code and geography. Use it to size a niche before committing to it, not to build rows.

Hiring surfaces, read for process rather than for jobs

Public applicant tracking boards: Greenhouse at boards.greenhouse.io plus a company token, with board data readable without an account, and the equivalents on Lever, Ashby and Workable.
The requirements section of a clerical posting, which names the ERP, the CRM, the portals and the spreadsheets. This is the cheapest read of the integration surface available anywhere.
Repost history and open counts on the same board, which show a role the company has failed to fill or has decided to fill repeatedly.
Postings for business systems, operations analyst or administrator roles, which say whether an internal owner exists for the build after handover.

Platform and partner directories

UiPath Find a Partner, filterable by country, tier and specialization, which shows who already implements in a given market.
Make partners directory and the Zapier Experts directory, both public and both carrying tiers and client reviews.
The n8n expert partner listing, useful for reading who serves which vertical.
Vendor customer story pages, which name a customer and a date and therefore identify companies mid go-live.

Complaint and entity surfaces

Google reviews, Trustpilot and Better Business Bureau complaint records, read for turnaround language rather than for sentiment.
National and state company registries, for entity status, incorporation date and ownership change.
The company's own team page, how it works page and careers page, plus an archived copy of each for comparison.

How a row gets verified

Confirm the company still trades and still operates inside the chosen niche, from a dated public signal.
Confirm the role still holds against two sources, because operations titles move between functions more than they move between companies.
Confirm the trigger is dated and younger than your sales cycle. A posting from last year is a fact about the company, not a reason to write this week.
Verify the address technically: syntax, domain MX record, catch-all detection and a verification pass. Catch-all domains move to a separate, slower segment.
Suppress existing clients, prior opt-outs, do-not-contact entries, and any row whose only source was a directory whose terms forbid marketing extraction.

What makes a bad list here

An industry-code and headcount export with no niche behind it. Nothing in it tells you which process to name.
Rows with no named process, which force the opener back to the technology and into the pile.
Rows whose trigger is that the company probably has manual work somewhere.
Six different verticals inside one campaign, so no process story fits more than a handful of rows.
Generic info@ and ops@ addresses treated as decision makers. At a small firm they might reach the owner; at a mid-market one they reach nobody.
Directory records extracted against that directory's own stated terms.
Micro-businesses pulled in by a headcount floor set too low, where there is no process volume to work on.
Rows from a scraped job aggregator with no link back to the employer's own posting, so neither the date nor the named systems can be checked.

Rules

Compliance and trust constraints

Nothing below is legal advice. Read each source directly and check your own position with a qualified adviser, because terms and rules change.

Association directories often forbid extraction for marketing

ALTA member profile pages state that the directory exists to connect professionals for legitimate business purposes, and that ALTA prohibits extracting directory information to build marketing databases or to send unsolicited mail, citing the Telephone Consumer Protection Act as a penalty risk. Other associations carry comparable terms with different wording. The practical effect is that a member directory is a research surface for identifying companies in your niche, and contact data has to be assembled elsewhere and verified.

Read the source

Your best proof is usually under a confidentiality clause

The strongest evidence an implementation firm has is a named client's named process, and that is frequently the exact thing the engagement contract forbids naming. Get written permission before a client name, a process detail or a figure goes into a cold email. An anonymized version works if it stays precise about the process and vague only about the identity, and you should expect a prospect to ask which client it was on the first call.

Read the retailer or regulator source directly.

Sender authentication decides whether any of this is read

Google's sender guidelines have applied to Gmail accounts above a daily volume threshold since February 2024. They set out authentication with SPF, DKIM and DMARC, matching forward and reverse DNS, TLS, standards-compliant message formatting, one-click unsubscribe on marketing mail, and a spam rate held low. Operations and finance buyers at mid-market companies sit heavily on Google Workspace and Microsoft 365, so a sending domain that fails these checks never reaches the argument.

Read the source

The actual document

A targeting brief for this market, in full

GrowthPal fills in this document before a single row is built, and publishes it whole rather than as a screenshot, because the shape of it is the argument. Every field is filled from public sources before any call happens.

Field	What goes in it
Niche	One vertical, named, with its member directory, its association, its conference and its five process nouns listed underneath. Not two, and not a size band.
Niche size	Establishment count from County Business Patterns or a member count from the association, written down. If it is small, we say so here rather than after the first month.
The process	The one process you have built more than once, named the way people inside that trade name it, not the way a tool vendor names it.
Proof you are permitted to use	The before and after, the vertical reference, and whether each client name, process detail and figure has written permission behind it.
Trigger	The dated event that earns the email, plus the public URL it was read from. A repost date, a go-live, a complaint pattern, a second site.
Systems named on the row	The ERP, CRM, portals and spreadsheets pulled from the job spec. This is the integration surface, read before a conversation exists.
Internal systems owner	Whether a business systems or operations analyst role exists. Its absence changes both the buyer and the handover answer.
Angle	The pain hypothesis in one line, in the buyer's vocabulary, phrased as something they can correct.
Handover answer carried	The runbook, the break-fix path and the change request route, stated in the sequence rather than held for contracting.
What would tell us this is wrong	Written down before sending. A hypothesis with no failure condition cannot teach you anything, because no result would ever count against it.

We publish the brief rather than a dashboard screenshot because a screenshot proves a number and a brief proves a method. Copying this means copying how we work, which is the point.

Process

How the engagement runs

Three stages and nine steps, the same ones in every market: build the list from public sources, race message angles against it, scale the one that wins. The nine steps from kickoff to booked meetings are written out in full.

What changes for this market

Week one settles the niche and the process story, in that order. We ask which vertical you will commit to, which process you have built more than once, and what you are permitted to name from past work. Nothing is built until both answers exist, because the list source and the vocabulary both follow from them.
List building reads hiring surfaces for process rather than for jobs. Public applicant tracking boards give repost dates, open counts and, in the requirements section, the names of the systems a person sits between. Every row carries its dated trigger and the public URL it came from, and a row without a named process does not get sent.
Copy review has an extra pass that removes the technology from the opener. Each first line has to name a process and a cost the buyer already tracks, and has to read as a question the buyer can correct rather than a claim about their business.
The handover answer goes in the sequence, not in the contract. Who owns the build, where the runbook lives and what the break-fix path is get written once and reused, because that is the objection that kills these deals at the last step.

Proof

What GrowthPal has published

GrowthPal takes on outbound engagements in this market, and the method above is the one it runs here. The published record below is the whole company record, across every client.

What was published	What the figure counts	Source
A study of 412 campaigns across 47 companies over 18 months: campaigns scoring above 0.7 on the composite message-market fit score averaged an 11.4 percent reply rate, against 1.1 percent for campaigns below 0.3. Volume, warm-up and time of day explained less than 8 percent of the variance.	Total reply rate, measured against messages in those campaigns. This is a different measure from interested replies per contact and the two must never be compared on the same page.	Read it
4M+ cold emails sent for 40+ B2B companies	Emails sent, cumulative across all clients. Not one campaign.	Read it

How these are counted

An interested reply means a human reply expressing interest in continuing the conversation. Out-of-office messages, unsubscribes, referrals to someone else and automated responses are not counted. A figure measured against contacts reached and a figure measured against emails sent are different measures: the column above says which each one is, and we do not convert between them or add them together.

The full counting method, and what an interested reply is.

On an industry-specific result. GrowthPal has not yet published an outcome from a campaign run for an automation implementation firm, so this page offers method rather than a result: the list sources, the detection signals and the targeting brief we would fill in before building a single row.

Honestly

Who this is not for

You automate anything for anyone. Without a chosen niche there is no directory to build from, no shared vocabulary to write in, and no process story that transfers between rows. That is a positioning decision, and it has to be made before an outbound program is worth running.
A niche smaller than roughly 30,000 reachable contacts. Picking a niche can take you below the size where a system pays for itself. Testing needs volume to separate a message that works from one that got lucky. Under that size, outbound still works, but a short hand-run list worked by the founder usually beats a program.
You cannot name one process you have built more than once. The proof that moves this market is a named process with a before and after. If every engagement so far has been a one-off, there is nothing to put in the email yet, and the first job is to repeat a build, not to send.
Delivery is one person and that person is you. A bespoke build needs scoping, delivery and a break-fix path after handover. If every one of those is the founder, replies arrive faster than the work can be taken on, and the first missed handover costs more than the campaign returns.
No one ready to take the calls. Replies start within days of the first sends. If nobody owns follow up inside a day, interested operators go cold and the spend is wasted.

Questions

Answered straight

Should the word AI appear in our first email?

Usually not in the first line, and often nowhere in the first email. Your buyer has heard the technology pitch many times and grades the email on whether it describes their work accurately. What earns a reply is a named process and a cost they already track: close duration, quote turnaround, headcount in one function. We could not find a credible public measurement of that fatigue, so we treat it as a hypothesis to test against a control, not as a rule.

Every project we deliver is different. What proof can go in a cold email?

One process, not one project. Name the process in the buyer's own vocabulary, show the workflow before and after with step count and cycle time, reference a company in the same vertical, and say what the maintenance burden turned out to be. A buyer discounts a case study from another industry almost entirely. That proof is yours to assemble, with written permission for anything named. We cannot supply it and neither can any outbound agency.

Which single process should we build the whole list around?

You can send, but the list will come from an industry-code filter and the angle will lead with technology, which is the combination that fails here. A niche hands you four things at once: a member directory, an association calendar, a conference exhibitor list, and a vocabulary of process nouns insiders use. It also makes one process story transferable to every row. Choosing it is the work, not a preliminary to it.

How can you tell a company has a manual process before talking to them?

You cannot confirm it from outside. You can narrow it. The same clerical role posted repeatedly, a job spec listing the systems a person sits between, headcount growth in a repetitive function with no systems hire, review text about waiting weeks, and a published how it works page describing six handoffs. Each one is an inference, which is why the email asks whether we read it right rather than announcing what is wrong.

Our best prospects are small companies. Is the market big enough?

Check before committing. County Business Patterns from the Census Bureau gives establishment counts by industry code and geography, and association directories publish member counts. If the niche holds a few thousand companies, a systematic program is the wrong instrument, and we will say so rather than build one. Two adjacent niches with a shared process sometimes solve this, provided the process story transfers to both.

Prospects ask who maintains the build after we leave. Where does that go?

Early, in writing, inside the sequence. The fear is not that the build fails. It is being left with something nobody in the building understands. Publish the handover model: who owns the build, where the runbook lives, what the break-fix path is, what a change request looks like. Raising it before the buyer does converts the objection into a reason to trust you, and it is the single cheapest piece of copy in this market.

Our prospects tried automation before and it failed. Does that help or hurt?

It helps, if you name it carefully. A failed prior attempt means budget was approved once, the pain is real, and someone inside already agrees with you. The angle is about what usually breaks on the second attempt, how exception handling gets underestimated, and what the first build left behind. Never name the previous vendor, and never describe the failure in a way that makes the buyer look careless to a colleague they will forward it to.

How long until we know whether the message works?

Campaigns are sending two to four weeks after kickoff, and warming the infrastructure is most of that gap. A winning message is usually identified around week eight. The caveat specific to this market is niche size. If the chosen niche is small, week eight arrives with too few sends to separate a message that works from one that got lucky, and the honest answer is to read the replies qualitatively instead.