SingleBit Outbound Engine: Plan for 10 to 15 Booked Meetings in One Month
Prepared: 2 November 2026 Goal: turn the system on, let it run for about a month, and book 10 to 15 sales meetings. Main channel: cold email.
1. The short version
We are building a machine with eight stations. Companies go in one end, booked meetings come out the other.
- Pick who we sell to and what we offer. (People decide this. No tool fixes a weak offer.)
- Find companies that fit.
- Spot a reason to email each one right now (they are hiring engineers, running ads, just raised money, have an old website).
- Find the decision maker and check their email address is real.
- Write a short, specific email for each person. A human approves every batch.
- Send it slowly, spread across many mailboxes, so we stay out of spam.
- Answer replies fast, with AI drafting and a human sending.
- Count booked meetings and, every week, shift effort to whatever is booking.
The tools:
- Harvey does stations 2, 3 and 4. It is free and open source, and it runs on our own computer.
- A small bridge we build does station 5.
- Instantly does station 6. We already use it.
- A small reply desk we build does station 7.
- Cal.com plus a simple tracker do station 8.
- Zomler adds extra signals at station 3 where our free sources are weak.
Expected cost is roughly $250 to $400 a month in tools, plus about 5 to 7 engineer days to build. The full breakdown is in Section 6.
2. Is 10 to 15 meetings realistic? The math
Industry numbers for 2026:
- Average cold email reply rate is about 3.4%. The top 10% of senders get more than 10% (Instantly 2026 benchmark).
- Meeting booked rate typically runs 0.5% to 2.5% of people emailed. One dataset puts the average at 0.7%.
- 58% of replies come from the first email. The other 42% come from follow-ups, so follow-ups are worth sending.
Our capacity:
| Item | Number | Why |
|---|---|---|
| Mailboxes we already have | 9 (on 3 domains) | From our earlier setup |
| Safe cold emails per mailbox per day | 25 | Most 2026 guides put the safe range at 20 to 30 for warmed Google Workspace inboxes |
| Sending days in the window | about 20 | 9 Nov to 11 Dec, skipping US Thanksgiving (26 and 27 Nov) |
| Emails per month (9 mailboxes) | about 4,500 | 9 × 25 × 20 |
| Emails per person | about 2.5 | A 3-email sequence, but repliers stop early |
| People reached (9 mailboxes) | about 1,800 | |
| People reached if we add 6 mailboxes now | about 2,300 | New ones join late, after warming |
What that gives us (planning estimates, not promises):
| Meeting rate per person reached | 1,800 people | 2,300 people |
|---|---|---|
| 0.5% (below average) | 9 | 11 |
| 0.7% (average) | 13 | 16 |
| 1.0% (good, signal-based targeting) | 18 | 23 |
What this means: with only the current 9 mailboxes, we hit the goal only if we perform at or above average. Adding 6 mailboxes gives us room for a slow first week. The thing that moves us from 9 meetings to 20 is who we email and why (stations 1 and 3), not more volume.
Timing warning: people emailed in the last week of the run will still be getting follow-ups and replying into mid December. Count meetings booked up to about 18 December, not just up to day 30. US business slows down after about 18 December, so the window is tight. Start this week.
3. The eight stations, explained simply
Each station below covers: what it means in plain words, what we use, what we build or set up, the cost, and how we know it is working.
Station 1: Who we sell to, and what we offer
In plain words: we decide which kinds of businesses to email, and the one thing we offer them. Everything later depends on this.
Our likely ICPs (ideal customer profiles) from earlier work. Please correct these if they have changed.
- Founders (Tier 1 MVP builds, Tier 3 retainers): startups that need product built.
- CPA firms (Tier 2 digital presence): US accounting firms with weak websites and no online booking.
- NGOs (Tier 2): nonprofits with dated sites, especially ones running ads.
- CBD brands (Tier 2 / Tier 3): e-commerce brands running Meta ads with weak sites or SEO.
Recommendation for month 1: run two ICPs, not four. Splitting 1,800 to 2,300 people four ways leaves about 500 per ICP. That is too few to learn what works. I suggest:
- Founders. Hiring and funding are the strongest, cleanest signals we can detect.
- CPA firms. November and December is a good moment to reach them, because the January to April tax season makes them hard to reach later.
NGOs and CBD join in month 2, reusing the same machine.
The offer. Each ICP needs one concrete, low-commitment thing to say yes to. For example: "a 20-minute teardown of your booking flow" for CPAs, or "a fixed-price MVP scoping call" for founders. People write this, not AI.
What we set up:
- A one-page brief per ICP: who exactly, what pain, what offer, and what proof we have (case studies).
- A do-not-contact list of every current and past client and partner domain. This includes Medicileaf, Arista Botanicals, Guiding Elephant, Spiral Studios, and anyone else we have worked with. This matters most for the CBD ICP, so we never cold-email our own clients by accident. Also decide whether pitching direct competitors of Medicileaf and Arista is a conflict you want to avoid.
Done when: two briefs are written and the do-not-contact list is loaded.
Station 2: Find companies
In plain words: make a long list of businesses that might fit.
What we use: Harvey's "Discover" step. It searches the web and business listings and shows the cost before it spends anything.
- Founders: Harvey has built-in searches for YC-backed companies, accelerator-backed companies, recently funded companies, and companies hiring for roles we name. These run on a web search service called Serper, which has 2,500 free searches and then about $0.30 to $1.00 per 1,000.
- CPA firms: DataForSEO Business Listings at about $0.37 per 1,000 businesses. It is good for local firms and can even find firms with no website.
- Backup source: Apollo. If Harvey's free sources do not find enough founders, we export a filtered company list from Apollo and feed it into Harvey. This needs a small import adapter (see Section 5), because Harvey cannot import a CSV today.
Cost: under $20 a month at our volume.
Done when: we have about 2,500 or more companies across the two ICPs. Harvey finds up to 2 people per company, and some emails will not verify, so we need more companies than our target of people.
Station 3: Spot the reason to email right now (signals)
In plain words: for each company, find a true, specific, current fact that makes our email relevant today. "You're hiring a founding engineer" beats "I saw your website" every time. This is the station that decides our meeting rate.
What we use: Harvey's "Profile" step. It is free. Harvey reads each company's homepage and a few other pages, with no AI involved, and records facts ("signals") with the page they came from. You approve which of its 29 signals to collect. Useful ones for us:
| Signal | What it means | Best for |
|---|---|---|
HIRING_ROLE | They are hiring for a role we name (founding engineer, full-stack, mobile). Harvey reads their Greenhouse, Lever, Ashby or Workable job board directly. | Founders |
FUNDING_ROUND | They announced a raise on their own site in the last 18 months | Founders |
YC_BACKED / ACCELERATOR | They say they went through YC or an accelerator | Founders |
NO_ONLINE_BOOKING | Clients have to phone in to book | CPA firms |
BLOG_STALE | They started content and abandoned it | CPA, NGO, CBD |
INCUMBENT_AGENCY | The footer names the agency that built their site | All Tier 2 |
RUNNING_META_ADS | Careful: this really means "Meta Pixel is installed", not "ads are live" | CBD, NGO |
SERP_RANK | Where they rank on Google for their own service (paid, cheap) | CPA |
Where Zomler helps. Zomler fills gaps that Harvey's free signals cannot.
- Confirming real ads: Zomler's Facebook Ad Library lookup shows whether a page is actually running ads right now, at about 4.3 credits per page. We use it only on a shortlist, to fix the pixel-versus-real-ads problem above.
- LinkedIn job posts: about 1.7 credits per job found. I checked: 50 "founding engineer" postings from the last 7 days would cost about 87 credits. This works as a top-up, not the main source, because Harvey reads job boards for free.
- LinkedIn posts with buying intent: founders posting things like "looking for a dev agency". About 1.7 credits per call. Low volume, but these are the hottest leads we will ever find. These go to a human, not the machine (see Station 7).
- Not used for email addresses at scale. Zomler's contact reveal costs about 178 credits per person. The wallet has about 232 credits. Harvey plus a verifier does the same job for cents.
What we build: a Zomler adapter so Zomler's findings land in Harvey as signals, next to Harvey's own. Harvey's code has a clean plug-in point for this.
Then we build "cohorts". A cohort is a saved filter, for example "Founders: hiring a full-stack engineer AND raised in the last 18 months". Each cohort becomes one campaign with one angle. This is how we will later know which reason to email actually books meetings.
Done when: each ICP has 2 or 3 cohorts with at least 300 companies each.
Station 4: Find the right person and a real email address
In plain words: find who decides (owner, founder, CEO), work out their email address, and check it is real before we ever send. Bad addresses bounce, and bounces push all our mail into spam.
What we use:
-
Harvey "Enrich" (free). It reads team and about pages and pulls out owners and founders by name. At most 2 people per company, with owners first.
-
Harvey "Verify". Harvey learns each company's email format from any address published on their site (for example
jsmith@). It builds one best guess and checks it with a verifier. Every address gets one of these labels:- verified: safe to send.
- risky: the domain accepts everything, so we cannot be sure.
- guess: no verifier has checked it yet.
- invalid: do not send.
Only verified addresses (and optionally risky ones, in small amounts) are sent.
-
Verifier: Reoon, at roughly $1.20 to $3 per 1,000 checks. ZeroBounce is more thorough but around $16 per 1,000. Reoon is fine at our volume.
Cost: about $5 to $15 a month.
Done when: we have about 2,300 or more verified contacts queued, and bounce rate in sending stays under 2%.
Station 5: Write the emails (with a human approving)
In plain words: AI writes a short email for each person that mentions the specific fact we found. A person reads and approves every batch before anything goes out.
What we use:
- Claude (Anthropic API, Sonnet 5.5), priced at $2 per million input and $10 per million output tokens. Writing one person's email costs well under 2 cents.
- Harvey's writing rules. Harvey ships with plain-text rule files: banned AI phrases, short-email rules, follow-up frameworks. We reuse them and add our own
email_rules.mdwith our voice rules:- No em dashes.
- No invented numbers.
- "One team, under one roof".
- Never mention freelancers.
- Say what SingleBit does in one line.
What we build: the Writer Bridge, about 2 days of work. This is the main piece of custom code.
- Reads Harvey's database: the person, the company, and the signals with their source links.
- Asks Claude to write email 1 for that person (under 90 words: the observation, one line on what we do, one easy question). It also writes two short follow-ups.
- Runs safety checks: no broken name fields, no links in email 1, no banned phrases, under the word limit, nobody on do-not-contact.
- Puts every draft into a Google Sheet review queue. A person approves, edits or rejects each row.
- Sends approved rows to Instantly, with the personalized text stored on each lead.
Why we build this instead of using Harvey's own Instantly connection: I read Harvey's current code (version 0.2.1, last updated 1 Nov 2026). Its Instantly connection has these problems:
- Harvey itself labels it "legacy".
- It does not attach our mailboxes to campaigns.
- It only sends one shared template per campaign, not a personal email per person.
- With approval turned on, it does not send through Instantly at all.
Harvey's own Gmail sending is personal, but it uses one mailbox per install and cannot rotate across our 9.
Email shape rules (from the data):
- Short, plain text, no images.
- No link in email 1.
- Open and click tracking off, because tracking links hurt deliverability.
- End with a specific, time-based ask, for example "Worth 15 minutes Thursday?". One analysis found time-based asks book meetings at about 3.4x the rate of vague problem-based asks.
Running AI through the API, not a Claude subscription: Harvey is built to run on a Claude Pro/Max subscription. Anthropic's rules for automated use of subscriptions changed several times in 2026 and are still unsettled. A sales machine that makes money should not depend on that. We give Harvey and the bridge an API key and pay per use, roughly $50 to $150 a month at our volume. On day 1, check that the calls show up in the Anthropic Console.
Done when: for 3 batches in a row, the reviewer approves at least 80% of drafts without edits.
Station 6: Send (slowly, across many mailboxes)
In plain words: send the emails at a human pace, spread across all our mailboxes, at normal working hours in the recipient's time zone, and stop the moment anyone at that company replies.
What we use: Instantly, which we already use. It handles:
- Rotating across all mailboxes.
- Warming them up.
- Sending schedules and per-mailbox daily caps.
- Stopping the sequence for a whole company when one person replies.
Plan upgrade needed: our volume (about 4,500 or more emails a month) is over the Growth plan's 5,000 limit once follow-ups and new mailboxes are added. We also need webhooks (instant alerts when someone replies) and full reply-from-inbox. Webhooks start at Hypergrowth: $97 a month, or $77.60 a month billed yearly.
Settings:
- 25 cold emails per mailbox per day for warmed mailboxes. New mailboxes start at 5 to 10 a day and ramp up by about 5 a week.
- Warmup stays on, always.
- Send window: 8am to 5pm in the recipient's time zone, Tuesday to Thursday weighted. Wednesday has the highest reply rate in Instantly's data.
- Sequence: email 1, then a follow-up on day 3, then a final one on day 7.
- Pause sending on 25 to 27 November (US Thanksgiving).
Add 6 mailboxes this week: 2 new domains with 3 Google Workspace mailboxes each. Each mailbox costs roughly $7 to $8 a month (check the current Workspace price), and each domain about $10 to $15 a year. They need 2 to 4 weeks of warmup, so buying them now means they join around 30 November.
Health check first: before sending, confirm on all 9 existing mailboxes and 3 domains:
- SPF, DKIM and DMARC records are set (these prove to Gmail that we are who we say we are).
- No domain is on a blocklist.
- Instantly's warmup health is good.
Harvey's harvey warmup command checks the DNS records and blocklists for free.
Legal basics (US, CAN-SPAM). Not legal advice.
- Honest subject lines and sender name.
- A working opt-out line in every email ("Reply 'no thanks' and I won't email again").
- Opt-outs honored right away.
- A real physical postal address in the footer. Non-US senders often use a US virtual mailbox for this.
- Keep spam complaints under 0.3%, which is Google's sender rule.
Done when:
- Bounce rate is under 2%.
- No mailbox gets flagged.
- Instantly shows inbox placement as healthy.
Station 7: Answer replies fast
In plain words: when someone replies, we answer quickly and push toward a booked call. Speed matters a lot here.
The time zone problem: US business hours (9am to 5pm Eastern) are 7:30pm to 3:30am in India. Most positive replies will arrive during our evening and night. This station has to be designed around that.
What we build: the Reply Desk, about 1 to 1.5 days of work.
- Instantly sends an instant alert (webhook) the moment a reply arrives.
- A small web function (hosted on our existing Vercel) receives it and asks Claude to:
- Sort the reply: interested, question, objection, not now, wrong person, unsubscribe, out of office.
- Draft an answer.
- Positive replies and questions ping the team's phone (Slack or WhatsApp) with the reply and the draft.
- A person sends the answer from Instantly's inbox, editing if needed.
- Unsubscribes are honored automatically. Out-of-office and "not now" replies are scheduled for later.
One decision for you: do we allow one kind of reply to go out automatically? I mean only "yes, interested, when can we talk?", answered with the booking link and two suggested times. Everything else waits for a human. This is the only way to answer a 2pm Eastern reply in minutes rather than 10 hours later. The alternative is an evening rotation (someone on duty until about 1am IST on weekdays).
What we do NOT automate: objections, pricing questions, anything legal, and anything from a big-name account. Those always get a human.
Done when: the median time to answer a positive reply is under 1 hour during US business hours.
Station 8: Book and count meetings, then adjust weekly
In plain words: make booking easy, count every booked meeting, and record why that person was emailed in the first place. Then each week, put more effort into what works and stop what does not.
What we use:
- Cal.com booking page (free). One event type per ICP, with 2 or 3 short questions.
- Cal.com webhook. When a meeting is booked, our Reply Desk function receives it, marks the lead as "meeting booked" in Instantly, and adds a row to a Meetings sheet.
- Meetings agreed by email ("Tuesday 3pm works") get marked by hand in Instantly. That also fires a "meeting booked" alert to the same sheet.
Every lead carries tags: ICP, cohort, signal used, copy version. So the Meetings sheet answers the question that matters: which reason-to-email books meetings?
Weekly 30-minute review (every Friday). Check these numbers:
| Number | Healthy | If not |
|---|---|---|
| Bounce rate | under 2% | Stop. Fix verification before sending more |
| Reply rate per campaign | 3% or more | Under 1% after 300 people: rewrite or kill the campaign |
| Positive replies | about 1% of people reached | Check the offer, not just the copy |
| Meetings booked | about 3 or more per week from week 2 | Move volume to the best cohort |
| Spam complaints | under 0.3% | Slow down, check targeting |
Done when: we can say, for every booked meeting, which ICP, cohort and signal produced it.
4. How it fits together
[Station 1: ICP + offer briefs] (people)
|
v
[Harvey: Discover] -> Serper / DataForSEO / Apollo import
|
v
[Harvey: Profile] -> free signals from each company's own site
+ Zomler adapter -> real Meta ads, LinkedIn jobs, intent posts
|
v
[Harvey: Enrich + Verify] -> decision maker + Reoon-verified email
|
v
[Writer Bridge] -> Claude API writes per-person emails
| -> safety checks -> Google Sheet review queue
| -> approved rows pushed to Instantly
v
[Instantly] -> 9 to 15 mailboxes, warmup, schedule, stop-on-reply
|
v (webhook on every reply)
[Reply Desk on Vercel] -> Claude sorts + drafts -> phone alert -> human sends
|
v
[Cal.com booking] -> webhook -> Instantly "meeting booked" + Meetings sheet
|
v
[Friday review] -> shift volume to the cohorts that book
Where things run:
- Harvey: a small always-on Linux server, for example a DigitalOcean droplet, since we already use DigitalOcean for Zomler. Harvey's dashboard is private and opened through a secure tunnel.
- Reply Desk and webhooks: a small function on our existing Vercel.
- Everything else: hosted tools (Instantly, Cal.com, Google Sheets).
Safety design: Harvey itself never sends. Harvey's own sending stays in approval mode, pointed at one test mailbox, with nothing approved, so it is impossible for it to email anyone. All real sending goes through Instantly, after a human approves the batch in the Sheet.
5. What we build vs. what we just set up
| Piece | Build or set up | Effort | Notes |
|---|---|---|---|
| ICP and offer briefs, do-not-contact list | Set up (people) | 0.5 day | Founders decide offers |
| Buy 2 domains, 6 mailboxes, DNS, warmup | Set up | 0.5 day | Do on day 1, since warmup takes weeks |
| Instantly upgrade to Hypergrowth | Set up | 10 min | Needed for webhooks |
| Harvey install on server, train on singlebit.xyz, confirm signals, cohorts | Set up | 1 day | One Harvey install per ICP, since Harvey holds one product and ICP at a time |
| Switch Harvey to an Anthropic API key | Set up | 1 hour | Verify usage appears in Console |
| Apollo / domain-list import adapter for Harvey | Build | 0.5 day | Harvey has no CSV import today |
| Zomler adapter for Harvey | Build | 1 day | Ad Library check, LinkedIn jobs, intent posts |
| Writer Bridge (Claude, checks, review Sheet, push to Instantly) | Build | 2 days | The core custom piece |
| Reply Desk (webhook, sort, draft, alert) | Build | 1 to 1.5 days | Reuse our existing Instantly reply-alert work |
| Cal.com event types, webhook, Meetings sheet | Set up + small build | 0.5 day | |
| Total | about 5 to 7 engineer days | Can run in parallel with mailbox warmup |
6. Budget (monthly, approximate, USD)
| Item | Cost | Notes |
|---|---|---|
| Instantly Hypergrowth | $97 (or $77.60 billed yearly) | Webhooks and the volume we need |
| 6 extra Google Workspace mailboxes | about $45 to $50 | 9 existing mailboxes already paid |
| 2 extra domains | about $25 to $30 per year | One-time-ish |
| Email verification (Reoon) | $5 to $15 | Roughly 3,000 to 5,000 checks |
| Discovery (Serper, DataForSEO) | $0 to $20 | Serper's free tier covers a lot |
| Claude API (Harvey + bridge + reply desk) | $50 to $150 | Estimate. Measure in week 1 |
| Server for Harvey | about $12 to $24 | Small droplet |
| Cal.com | $0 | Free plan covers this |
| Zomler | internal credits | Top up the wallet if we use Ad Library checks heavily |
| Apollo | only if needed | Use our existing plan if we still have one |
| Total | about $250 to $400 / month | Plus build time |
If this books even one Tier 1 or Tier 3 client, it pays for itself many times over. The real cost is team time: building it, approving batches, answering replies in the evenings, and running the Friday review.
7. Timeline
| When | What happens |
|---|---|
| Week 0 (2 to 8 Nov) | Decide the 2 ICPs and offers. Buy domains and mailboxes (start warmup). Upgrade Instantly. Health-check existing mailboxes. Install Harvey, train it, confirm signals, run discovery. Build the import adapter and Writer Bridge. |
| Week 1 (9 to 13 Nov) | First sends from the 9 warmed mailboxes at about half volume (10 to 15 a day each). Every draft reviewed. Build the Reply Desk and Cal.com tracking. Build the Zomler adapter. |
| Week 2 (16 to 20 Nov) | Full volume (25 a day per mailbox). First Friday review: kill weak cohorts, double good ones. |
| Week 3 (23 to 25 Nov) | Send Mon to Wed only. Pause 26 and 27 Nov (Thanksgiving). |
| Week 4 (30 Nov to 4 Dec) | New mailboxes start joining at low volume. Full volume on the best cohorts. |
| Week 5 (7 to 11 Dec) | Last new sends. Follow-ups finish the following week. |
| Final count (around 18 Dec) | Total meetings booked, broken down by ICP, cohort and signal. Decide month 2: add NGO and CBD. |
8. From the Orphic catalogue: what we use and what we skip
Searched the full catalogue (155 tools), especially the Sales CRM, AI agents and Business ops categories.
| Tool | Verdict | Why |
|---|---|---|
| Harvey | Use (stations 2 to 4) | Real research engine: free site profiling, decision-maker finding, honest email verification, signals database. Its sending and calendar parts are not ready for our setup (calendar booking is still a placeholder in the code), so we do not use those. |
| Google Maps Scraper Kit | Maybe, month 2 | Free local-business lists (names, phones, emails, sites by city and niche). Useful for NGO or local CBD retail later. Not needed for founders or CPAs, where DataForSEO is cheap and cleaner. |
| Agent Reach | Skip | Reads LinkedIn, X and Reddit for free, but some platforms need a logged-in session, which risks our accounts. Zomler does this cleanly. |
| n8n | Skip for now | A visual automation tool. Our pieces are small enough to write directly, and fewer moving parts is safer. |
| Agency Agents, Anthropic sales plugin | Skip | Prompt collections. Harvey's sales rule files already cover this, built for cold email. |
| Browser Use | Skip | Automating LinkedIn breaks LinkedIn's rules and gets accounts restricted. |
| Blueticks, whatsapp-mcp | Skip for US outbound | Cold WhatsApp is not normal in the US B2B market. Could matter later for Indian clients. |
9. Is there something better than email?
For our goal, email is the right main channel. It is the only channel we can run at about 2,000 people a month safely and legally.
Two additions:
- A manual LinkedIn touch for the top accounts. For the best 20 or so leads each week (intent posts from Zomler, or replies that went quiet), a founder views their profile, connects with no note, and comments on a recent post. A human does this, with Claude suggesting what to say. Not automated.
- No AI cold calling. Under US rules (FCC 24-17 / TCPA), AI-generated voice calls count as "artificial voice", with penalties of $500 to $1,500 per call and no business exception. Harvey's own maintainers refuse to build it.
10. Risks, and what we do about them
| Risk | What it looks like | What we do |
|---|---|---|
| Weak offer | Replies say "not now" or nothing | Rewrite the offer, not just the copy. Test a second offer in week 2. |
| Spam folder | Reply rate near zero, Instantly health drops | Lower volume, check DNS, pause the affected mailbox, keep warmup on |
| Bad data | Bounces over 2% | Stop the campaign. Send only "verified" addresses, no "risky" ones |
| Not enough verified contacts | Queue runs dry by week 3 | Apollo import as backup. Add a second cohort per ICP |
| Slow replies at night IST | Positive replies go cold | Evening rotation, or allow auto-send for "interested" replies only |
| AI writes something wrong | Made-up fact in an email | Every email cites a stored fact with its source link. A human approves every batch. Harvey's rule bans invented facts. |
| Subscription policy changes | AI calls stop working | We run on the API, so this does not affect us |
| Holiday slowdown | Week 3 and mid December quiet | Front-load volume in weeks 1, 2 and 4 |
| Emailing a client by mistake | Embarrassment | Do-not-contact list is checked at send time |
11. Decisions needed from you
- Which 2 ICPs for month 1? I suggest founders and CPA firms.
- The one offer per ICP, and which case studies we can name.
- Add 6 mailboxes now? I suggest yes.
- Auto-send for "interested" replies only, or an evening rotation?
- The earlier plan kept the CPA campaign off Instantly. Does that still stand? If yes, CPA runs on 1 or 2 mailboxes through Harvey's own Gmail sending (separate install, approval on), and only founders go through Instantly.
- Who reviews draft batches daily (about 20 to 30 minutes a day), and who runs the Friday review.
- Postal address for the email footer (US virtual mailbox or our Bengaluru office).
12. What I checked and what I could not
Checked directly:
- Harvey's code and docs, cloned 2 Nov 2026 (version 0.2.1, last commit 1 Nov 2026). I read how its Instantly connection, Gmail sending, verification, discovery providers and calendar code actually work.
- Instantly pricing and webhook events from their pricing and developer docs.
- 2026 cold email benchmarks from Instantly, Mailshake, Cleanlist and others.
- Safe sending limits per mailbox from several deliverability guides.
- Verifier prices, Cal.com webhooks, and Claude API pricing.
- Zomler's live prices, using its free capabilities tool and a free cost estimate.
Not verified, so check these:
- The current Google Workspace price per mailbox.
- Whether our 9 mailboxes are still healthy after months of warmup.
- Exactly how many verified contacts Harvey will find per 100 companies. This varies a lot by industry, so we measure it in week 0.
- Real Claude API spend. Measure in week 1.
- The benchmark numbers are industry averages, not our results.
13. Plain-language glossary
- ICP: ideal customer profile. A description of the exact kind of business we want.
- Signal: a true, current fact about a company that gives us a reason to email ("hiring a founding engineer").
- Cohort: a group of companies that share the same signals, emailed with the same angle.
- Warmup: a mailbox quietly exchanging normal-looking emails for weeks so Gmail and Outlook learn to trust it.
- SPF / DKIM / DMARC: three settings on our domain that prove our emails really come from us.
- Bounce: an email sent to an address that does not exist. Too many of these hurt every mailbox.
- Verifier: a service that checks whether an email address is real before we send.
- Webhook: an instant automatic message from one tool to another ("someone just replied").
- API: the way our code talks to a tool (Claude, Instantly) directly.
- Catch-all ("risky"): a domain that accepts mail to any address, so we cannot tell if the person exists.
- CAN-SPAM: the US law for commercial email. Honest headers, an opt-out, and a postal address.
Sources
- Harvey repo: https://github.com/ethanplusai/harvey
- Orphic catalogue (Sales CRM): https://orphic.adarshji.com/explore/category/sales-crm.md
- Instantly 2026 Cold Email Benchmark Report: https://instantly.ai/cold-email-benchmark-report-2026
- Cold email meeting rates: https://www.cleanlist.ai/blog/2026-02-18-cold-email-response-rate-statistics and https://mailshake.com/blog/cold-email-benchmarks-2026/
- Instantly pricing: https://instantly.ai/pricing.md and https://emelia.io/software/instantly/pricing
- Instantly webhook events: https://developer.instantly.ai/guides/webhook-events
- Safe sending limits: https://leadhaste.com/blog/google-workspace-cold-email-limits-2026
- Reoon pricing comparison: https://prospeo.io/s/reoon-alternatives
- Cal.com webhooks: https://cal.com/docs/core-features/webhooks
- Claude Sonnet 5.5 pricing: https://apidog.com/blog/claude-sonnet-5-5-pricing/
- Claude subscription automation changes in 2026: https://www.unite.ai/the-cheap-era-of-always-on-claude-agents-is-quietly-ending/
- Time-based CTA data: https://www.autobound.ai/blog/cold-email-guide-2026