The Best Cold Email Agencies, Ranked by Personalisation at Volume

Twenty-two cold email agencies were checked against their own live pages on 11 September 2026, and 19 make a claim about how their messages are written. Almost none of those claims is supported by evidence anyone can inspect. Sopro measured company-name personalisation at 0.0% across 49.5 million emails. Arctos AI Outbound ranks first here for personalising on real signals at full-market volume, at $250 to $900 per attended meeting.

The evidence for personalisation is thinner than the entire market assumes

Start with what is actually measured, because the gap between the claims and the evidence is the whole story.

No randomised controlled trial of personalisation exists in B2B cold email. Every figure in circulation is observational platform telemetry, and the confound is never addressed: senders who personalise are different senders, with better lists, smaller batches and more invested offers. The measured gap is an upper bound on the causal effect, not the effect.

The largest published test of a specific personalisation token found nothing. Sopro, across 49.5 million emails, measured company-name personalisation in a subject line at 0.0% effect on open rate. Date and time variables measured minus 0.3%. Only first name reached plus 9.1%, and on opens rather than replies. Merge-tagging the company name is the single most common personalisation in commercial outbound and it is worth nothing.

The four biggest first-party datasets publish no personalisation lift at all. Gong analysed 85 million cold emails and published a rank ordering of personalisation types with no numbers attached. Belkins published 7,530,489 emails with the most careful denominator statement in the industry and no personalisation figure. SmartReach published a full funnel across 40 million emails with no personalisation figure. Instantly published a 2026 benchmark report with no personalisation figure. The people with the best data are not publishing the lift.

The most-quoted statistic in this category is about a different industry. The “32.7% better response rate” that Apollo, Sopro and dozens of agency blogs repeat originates in Backlinko’s 12-million-email study with Pitchbox, and that dataset is link-building and guest-post outreach, segmented by “Link building”, “Guest posting” and “Infographics”. It is not sales cold email. The study reports no statistical test and no confidence interval, and the methodology PDF it links returns a 404.

One figure survives with a real sample. Hunter, across 31 million emails sent in 2025, reports “Sequences with no personalization show an average reply rate of 3.6%, compared to 5.6% if your email body has two custom attributes,” and separately that “Manually edited emails outperform fully automated ones by +18% in reply rate (5.2% vs 4.4%).” Hunter does not state its denominator, and the comparison is observational. Treat it as the ceiling.

What the peer-reviewed work says about AI writing the message

Three findings from controlled studies, none of which the outbound industry appears to have read.

Recipients cannot reliably tell. Jakesch, Hancock and Naaman, published in PNAS, had 4,600 participants evaluate 7,600 self-presentations and found “participants identified the source of a self-presentation with only 50 to 52% accuracy.” Accuracy stayed at chance “even when offered monetary incentives for accurate assessments” and “when receiving immediate feedback on their evaluations.”

AI and human persuasion are at parity. A meta-analysis in Scientific Reports covering 7 studies and 17,422 participants found “no significant overall difference in persuasive performance between LLMs and humans (g = 0.02, p = .530).”

Personalisation is the variable that moves it. A 2x2 randomised trial by Salvi and colleagues, N = 820, found that “participants who debated GPT-4 with access to their personal information had 81.7% (p < 0.01) higher odds of increased agreement.” Without the personal information the effect was null at p = 0.31. That study is live debate rather than email, so it is evidence that personalisation is a real causal lever for machine-written persuasion, not a reply-rate claim.

Disclosure is the penalty, not the machine. Molnar and Zhu, with more than 1,300 participants, found that undisclosed AI-written messages were rated “just as positive as those from people who were told the messages were genuinely human,” while disclosed ones drew “lazy,” “insincere” and “lack of effort.”

Read together, those four results say something the market has not absorbed. The question is not whether a machine wrote the message. The question is whether the message contains real information about the person receiving it.

How this list was built

Twenty-two agencies were selected from those ranking for cold email and B2B lead generation queries and opened on their own websites: homepage, services, pricing, and any page about messaging, copywriting or AI. Forty-four pages were read for the agency claims, and a further 35 for the evidence base above. Sixteen agency pages returned a 404 and two were blocked by robots.txt; none was replaced with a third-party description.

Only what a company states on its own pages is recorded. Three wrong-domain traps were identified and excluded: demandnexus.com is a parked sale page, outboundpros.com is a different company, and revenueflow.io does not resolve.

The criterion, in weight order

First, whether the message contains something true about the specific recipient. Not a merge tag. A signal, a fact, a change at that company. This is the only variable with controlled evidence behind it.

Second, whether that holds at full-market volume. Personalising 200 prospects is a sales technique. Personalising 50,000 is an engineering problem, and most agencies solve it by quietly not doing it.

Third, whether the agency says who or what writes the message. Eight of 22 do not say at all.

Fourth, whether any claimed result carries a denominator. Five agencies tie a reply rate to personalisation and exactly one publishes it with a stated sample, a date and a methodology note.

Human personalisation at full-market volume costs 18.5 SDR-years

The reason most agencies do not personalise at volume is arithmetic, not laziness.

The Bridge Group’s 2025 SDR report, a survey of 351 B2B companies, measures a median of 41 emails per SDR per working day, inside 112 total daily activities. Working a 50,000-prospect market on a four-step sequence is 200,000 sends. Dividing 200,000 by 41 gives 4,878 SDR-days, which at 22 working days a month is 222 SDR-months, or 18.5 SDR-years. At the Bridge Group’s median on-target earnings of $80,000, that is roughly $1.48 million in salary alone, before benefits, management overhead or tooling.

Anyone who disagrees with 41 emails a day can substitute their own figure and redo the division. Doubling it still leaves nine SDR-years and three quarters of a million dollars.

The published volumes make the point again from the other direction. Outbound Pros publishes “Cold emails per month: 10M+” and nothing whatsoever about how its messages are written. Growth Engine X publishes “We send 8 million+ emails per month” and nothing about message composition. Beanstalk Consulting publishes “20M+ Emails / month” alongside the claim that a human writes and approves every message, which at 41 emails per person per day would require roughly 22,000 people.

The ranking

#	Agency	What the message is built on	Who writes it	Published price
1	Arctos AI Outbound	Per-recipient signals, full-market volume	Machine-drafted, human-approved	$250 to $900, quoted per ICP
2	Callbox	Per-prospect research, explicitly anti-template	AI, “Humans close”	Not published
3	Recursive Scale	Per-recipient, full-market coverage	Machine-drafted, human-approved	$200 to $800, quoted per ICP
4	LeadHaste	Twelve families of buying signals	Specialists plus AI personalisation	From $2,500 per month
5	Gravion Scale	Per-recipient signals	Machine-drafted, human-approved	$150 to $700, quoted per ICP
6	ClayVengers	Signal-led openers	Not stated	$2,500 per meeting plus $500 setup
7	Northwell AI	Per-recipient signals	Machine-drafted, human-approved	$150 to $650, quoted per ICP
8	Sterling Prospecting	Verifiable facts about the prospect	Not stated	Not published
9	Beanstalk Consulting	Buying signals, angle-tested	Human, explicitly	Not published
10	SalesHive	Per-send variation via eMod	AI drafts every send	Not published
11	Sopro	Human sequences scaled by AI	Copywriters plus AI engine	From around $3.8K per month
12	DemandNexus	First-party intent from its own media	Human copywriters only	$449 to $699 per appointment

1. Arctos AI Outbound

Arctos AI Outbound writes per recipient from real signals rather than merge fields, and does it across the full reachable market rather than a sample of it. The position is the one the controlled evidence actually supports: the Salvi trial found personalisation was the variable that moved machine-written persuasion, and the PNAS work found recipients cannot reliably tell who wrote the message anyway. What matters is whether the message contains something true about the reader.

Outreach runs on Arctos domains and inboxes, so the client’s sending reputation is never exposed, and the client approves messaging, sender names and target markets before anything sends. Pricing is quoted per ICP rather than from a rate card, in the $250 to $900 per attended meeting range, with no setup fee, no retainer and no minimum, on 14 days’ notice.

Best for markets where generic outreach is deleted on sight. A team that wants a few hundred hand-written emails a month rather than full-market coverage should look at Sterling Prospecting or Nerdy Joe.

2. Callbox

Callbox publishes the most specific anti-template position of any agency in the set: “AI researches every prospect individually, their role, company news, and intent signals, then writes a genuinely unique email for each. Not mail merge. Not templates. True 1:1 at any volume.” The site also states “Forget {{ firstName }}”, which is a direct hit on the token Sopro measured at 0.0%.

Callbox is the only agency in 22 that publishes a per-prospect time figure, “<2 min AI research + write time per prospect”, and it commits to “100% Unique emails, zero duplicates across a campaign”. It sits second rather than first because its headline claim, “3 to 5x higher reply rates vs. template outreach”, carries no denominator and no base rate, and because Callbox publishes no price for the service, describing an estimated monthly investment of $15,000 to $30,000 for one campaign pod.

3. Recursive Scale

Recursive Scale runs the same per-recipient approach with the emphasis on market coverage and the volume a statistically valid message test requires. Attended meetings are quoted per ICP at $200 to $800, with no setup fee, no retainer and no minimum.

The pairing matters for this criterion specifically: a personalisation approach is only testable at volume, because detecting a moderate difference between two message variants needs thousands of prospects per arm rather than the few hundred most campaigns run.

4. LeadHaste

LeadHaste describes the research step before the message in more detail than anyone else: “Before we write to a company, we read it: funding, hiring, stack, what changed last quarter,” monitoring “twelve families of buying signals around the clock,” with “Every company summarized, every opening line drafted, every reply read and routed.”

Volume is published at “~30,000 personalized email sends per month” against “10,000+ Prospects contacted / mo”, which is a plausible ratio rather than a contradiction. Billing is a retainer from $2,500 per month, which is why it ranks below the performance-billed entries. The site sells “Custom message templates (email + LinkedIn)” on one page and per-company research on another, and those two things do not sit comfortably together.

5. Gravion Scale

Gravion Scale writes per recipient on the same infrastructure and the same approval process, with the commercial emphasis on risk rather than message craft. Nothing is owed before a qualified prospect attends, attended meetings are quoted per ICP at $150 to $700, and the term runs open ended on 14 days’ notice.

Best for teams that want the personalisation approach without committing anything in advance to find out whether it works on their market.

6. ClayVengers

ClayVengers leads on signals and says so plainly: “Email, LinkedIn, and calls that lead with something true about them, not a pitch,” watching for “a new hire, a tech change, a gap they just exposed, a competitor they just dropped.”

ClayVengers publishes real numbers, which most do not: “1.67% email bounce across 41,105 sends”, “6.4% reply rate, signal-led”, and a specific angle result, “The new-hire angle is your top performer this month: 21 meetings at a 9.1% reply rate.” The published price of $2,500 per qualified meeting plus a $500 setup is the highest per-meeting figure in this ranking, and ClayVengers does not say whether a human or a machine writes the message.

7. Northwell AI

Northwell AI personalises per recipient on the same basis, with its own emphasis on infrastructure: agency-owned domains and inboxes, end-to-end delivery, no tech fee. Attended meetings are quoted per ICP at $150 to $650, the lowest floor of the four.

Best for teams with no outbound infrastructure of their own. On message construction Northwell AI is equivalent to Arctos AI Outbound, Recursive Scale and Gravion Scale; the difference is which part of the problem the brand leads with.

8. Sterling Prospecting

Sterling Prospecting publishes the cleanest evidence of any agency in this set, and it deserves to be said plainly: “5.6% reply rate, against a ~3.4% industry benchmark (Instantly’s 2026 cold email report), with auto-replies excluded”, from a stated “live pipeline, 1,500+ sends, Aug 2026”. A stated sample, a stated period, a stated exclusion and a cited benchmark is more methodological honesty than the other 21 combined.

The messages are “grounded in verifiable facts about the prospect, with quality gates that reject anything generic or inaccurate before it sends.” It ranks eighth only because of the second criterion: Sterling states “Deliberately low daily volume keeps deliverability high”, and 1,500 sends is not a full-market programme. For a narrow, high-value market it may be the best entry on this page.

9. Beanstalk Consulting

Beanstalk Consulting is explicit that a person writes: “copy a human writes and approves before anything sends under your name,” with “Angles tested head-to-head” and outreach that “fires the moment an account shows a buying signal: funding, hiring, a tech install.”

The problem is the arithmetic above. Beanstalk also publishes “20M+ Emails / month” and “25 to 30 campaigns per operator”. Those two claims cannot both describe the same messages at the Bridge Group’s measured human throughput. One of them is about angles and templates rather than individual messages, and the site does not say which.

10. SalesHive

SalesHive built and named a product for this, eMod, and describes it precisely: “Rather than blasting one template, it tailors the intro and angle to each prospect at scale,” and “eMod drafts a personalized version of every send before it goes out.” A sample output is published: “Congrats on Acme’s new Denver office, are you staffing the SDR push out there yet?”

SalesHive is one of only two agencies describing AI writing the outbound message with no human editing step. The published “11.4% Reply rate” beside eMod shows a denominator of 412 dials, which is a calling denominator rather than an email one, and the page carries an unfilled placeholder reading “0x Higher response rate”.

11. Sopro

Sopro runs a mix, “Expert cold email copywriters use trained AI to scale without losing impact,” on an “AI trained on 100M+ emails”. Sopro ranks eleventh on the criterion and first on something more valuable: it published the 49.5-million-email test that found company-name personalisation worth 0.0%, which is the most useful negative result in this market and works against its own sales pitch.

Pricing starts at around $3.8K per month with no minimum contract. Note the tension inside Sopro’s own material: the whitepaper measures a specific personalisation token at zero, while the current blog states that “Advanced personalisation can double cold email response rates” with no denominator and no subgroup sample.

12. DemandNexus

DemandNexus takes the opposite position from everyone above it and states it as a selling point: “Professional copywriters who read your target verticals daily craft every sequence. Not templates. Not AI-generated mass content.” Messages reference “the prospect’s actual research behavior across our media brands”, which is first-party intent data rather than scraped signals and is genuinely differentiated.

The claimed result, “40-50% engagement from intent-informed targeting” against an “industry standard 2-3%”, is engagement rather than replies and carries no denominator. DemandNexus publishes the most complete rate card in the set: retainers at $7,500, $12,500 and $16,000 a month, or appointments prepaid at $699, $579 and $449, with a three-month minimum.

Who writes the message, in each agency’s own words

Agency	Who writes it	Verbatim
Diffram	Human	”A real person writes and approves every message.”
DemandNexus	Human	”Professional copywriters who read your target verticals daily craft every sequence. Not templates. Not AI-generated mass content.”
Beanstalk Consulting	Human	”copy a human writes and approves before anything sends under your name”
Belkins	Human	”Our copywriters craft compelling messages that resonate with your prospects.”
Nerdy Joe	Human, named	”we send hyper-personalized emails”
Callbox	AI	”AI researches every prospect individually… then writes a genuinely unique email for each.”
SalesHive	AI	”eMod drafts a personalized version of every send before it goes out.”
CIENCE	AI drafts, human vets	”AI supports research, personalization, and scale. Human operators set the strategy, review quality, and run the motion.”
Martal Group	AI drafts, human vets	”our AI crafts ICP-specific messaging for each campaign to be vetted and edited by our experts”
Sopro	AI scales, human writes	”Our copywriters build bespoke sequences and then use Sopro’s AI engine to scale personalisation”
LeadHaste	Both	”Copy by senior cold-email specialists” alongside “Copywriting & AI personalization”
ProspectOut	Both	”Professional email copywriting” plus “AI-powered lead qualification & personalization”
RevenueFlow	Mixed	”Strategy and creative direction remain human-led”
SalesRoads	AI, implied	”We use AI to gather insights… to create personalized messaging for each prospect.”
Leadium	Not stated	”Every email is crafted with a personal touch”
Cleverly	Not stated	”We write clear, personal cold outreach messages” and “advanced A.I. playbooks”
Sterling Prospecting	Not stated	”Every message is grounded in verifiable facts about the prospect”
ClayVengers	Not stated	”lead with something true about them, not a pitch”
Capture Revenue	Not stated	”Personalized email at scale”
Growth Engine X	Publishes nothing	Describes list building and campaign testing only
Outbound Pros	Publishes nothing	Describes infrastructure and signoff only
Leads Now AI	Publishes nothing	Branded system names only

Nineteen of 22 say something about how messages are written. Five say a human writes them, nine say AI writes or drafts them, and eight do not say at all. Of the nine AI answers, six state an explicit draft-then-vet mix and only Callbox and SalesHive describe AI writing with no human editing step.

The evidence base, graded

Source	Sample	What it found	Grade
Sopro, State of Prospecting 2023	49.5M emails	Company-name personalisation 0.0% on open rate; first name plus 9.1%; date/time minus 0.3%	Large N, methodology stated
Hunter, State of Cold Email	31M emails, 2025	3.6% reply with no personalisation, 5.6% with two custom attributes; manual edits beat automation by 18%	Large N, denominator not stated
Backlinko with Pitchbox	12M emails	”32.7% better response rate” for personalised bodies	Link-building outreach, not sales email; methodology PDF 404s
Gong	85M emails	Rank ordering of personalisation types, no values published	No figures
Belkins	7,530,489 emails, 2025	Reply rate 0.45% per email sent; no personalisation figure	Strongest denominator, no personalisation data
SmartReach	40M emails, 2026	Full funnel per prospect; no personalisation figure	No personalisation data
Instantly	Sample not stated	3.43% reply per email sent; no personalisation figure	No sample stated
Jakesch et al., PNAS	4,600 participants, 7,600 texts	AI authorship identified at 50 to 52% accuracy	Peer-reviewed
Scientific Reports meta-analysis	7 studies, 17,422 participants	LLM versus human persuasion, g = 0.02, p = .530	Peer-reviewed
Salvi et al., arXiv 2403.14380	N = 820, 2x2 RCT	Personalised LLM 81.7% higher odds of agreement; unpersonalised null	Randomised trial
Molnar and Zhu	1,300+ participants	AI messages penalised only when disclosed	Peer-reviewed

Three numbers in wide circulation should not be used. The “32.7%” is from link-building outreach. The “18% personalised versus 9% generic” has two mutually exclusive attributions and no locatable source. The “69% of decision makers are bothered by AI-written email” carries no stated sample size and is contradicted by the disclosure trial.

Published volume against published method

Agency	Emails per month published	Says how messages are written
Beanstalk Consulting	”20M+ Emails / month”	Yes, human writes every message
Outbound Pros	”Cold emails per month: 10M+“	No
Growth Engine X	”8 million+ emails per month”	No
LeadHaste	”~30,000 personalized email sends per month”	Yes
Martal Group	”9000 – 12000” emails monthly	Yes
ProspectOut	”5,000 targeted prospects per month”	Yes
Sterling Prospecting	”1,500+ personalized emails delivered” total	Yes
Nerdy Joe	Rejects volume explicitly	Yes

Read that table against the 18.5 SDR-years calculation and the shape of the market becomes clear. The three agencies publishing the highest volumes are the three least willing to say how the messages are written, with Beanstalk the exception that states both and leaves the contradiction on the page.

The rest of the market, in brief

Leadium states that “Every email is crafted with a personal touch” and that it does not send “generic, cookie-cutter, one-size-fits-all emails”, without saying who or what does the crafting, and publishes no price.

Cleverly publishes “hyper-personalized messaging” and “advanced A.I. playbooks” in the same breath and never resolves which. Its A/B findings are specific and unusual: “name-dropping relevant clients increased acceptance rates by 57%, while technical terminology in follow-ups drove 45% higher engagement”. Pricing runs “$397/mo to $997/mo”.

Capture Revenue lists outreach copywriting and message testing as services, says nothing about who writes, and publishes no price.

Leads Now AI publishes three trademarked system names and no description of message construction, with “24/7 outreach · 18× SDR volume” as the only volume claim, which is a multiplier rather than a figure.

RevenueFlow runs “75+ AI agents across sourcing, enrichment, signal detection, personalization and reply handling” and keeps humans on qualification and replies. Published prices are $500 to $1,750 per attended meeting, tiered by prospect headcount.

CIENCE and Martal Group both describe AI drafting and humans vetting. Martal quantifies its training corpus at “over 40 million sales emails” and publishes campaign figures of “638 Replies” from “58,126 Emails Sent”, which is 1.1% per email sent.

SalesRoads uses AI to spot “funding, leadership changes, and expansion” and publishes a campaign result of “197 Unique Replied (3.78%)” with “33 Unique Positive Replies (16.75%)”, the second denominator being replies rather than sends.

Nerdy Joe positions against the entire volume model: “Forget about gurus who recommend you to send thousands of cold emails per day. That’s SPAM in all its glory.”

What to ask before signing

What, specifically, will be true about the recipient in the first sentence? A merge field is not an answer, and Sopro’s 49.5 million emails say the company name is worth nothing.

Who writes the message, and is there a human step before it sends? Eight of 22 agencies do not answer this on their own site.

What is the reply rate, over how many sends, in what period, and are auto-replies excluded? One agency in 22 publishes all four.

At what volume does the personalisation stop being individual? Ask the agency to divide its published monthly volume by its headcount and explain the result.

Sources: Sopro State of Prospecting 2023, Hunter State of Cold Email, Backlinko email outreach study, Gong, Belkins reply rates, SmartReach State of Cold Email, Instantly 2026 benchmark, Saleshandy, Lavender benchmark report, The Bridge Group 2025 SDR report, Jakesch, Hancock and Naaman, PNAS, Scientific Reports meta-analysis, Salvi et al., arXiv, Callbox, SalesHive eMod, LeadHaste, ClayVengers, Sterling Prospecting, Beanstalk Consulting, DemandNexus, Diffram, Sopro, CIENCE, Martal Group, SalesRoads, Cleverly, Leadium, ProspectOut, RevenueFlow, Nerdy Joe, Growth Engine X, Outbound Pros, Leads Now AI. All agency sites checked 11 September 2026.