The Outbound Signal Engine: Timing Beats Lists

A list tells you who. A signal tells you when.

Most outbound starts with a spreadsheet of 5,000 names and one template. Instantly, a cold email platform, says the average reply rate across its users is 3.43% in its 2026 benchmark. That is its own data, and it means roughly 96 of every 100 people say nothing.

The list is not wrong about who. It is wrong about when. The same company that ignores you in March is shopping in May, because in May it raised money, hired a new VP, or watched its reviews fall apart.

Greg Isenberg makes timing the core of the third system in his marketing engineer episode: watch for the events, research the account, draft a specific angle, and send it to a person for approval. It is one of the six systems in what a marketing engineer builds, and it is the one owners ask about first.

We hold every message to one test. Call it the Because Test.

Before anything sends, finish this sentence with a fact from the last few weeks: "I am writing because." If the end of that sentence is "you are in our target market," the message does not go. If it is "you opened a second location on Monday," it might.

Buyers pick early, so timing is most of the game

Ninety-four percent. In 6sense's 2025 survey of more than 4,000 B2B buyers, that is the share of buying groups that had already ranked their preferred vendors before they first contacted one. They bought from that early favorite 77% of the time.

The same report found first contact moved earlier, from about 69% of the way through the journey in 2024 to 61% in 2025. 6sense sells intent data, so weigh it as a vendor's study. The direction is still the useful part.

Read together, the numbers say one thing. A message that lands while the buyer is still forming the short list has a chance; the same message a month later is noise.

Timing also applies after the reply. The Harvard Business Review's study of lead response found that interest fades fast once someone reaches out, which is why a reply from this engine should be routed the way speed to lead describes, in minutes, not tomorrow.

Eight signals worth watching, and where the data comes from

A signal is only useful if three things are true. You are allowed to collect it, it is fresh, and it earns a specific first line.

Some of the best are public. A private company raising money files a Form D, a short SEC notice listing the offering size and the date of first sale. Job posts are public too, and Greenhouse's Job Board API returns any company board's published jobs as JSON without a key.

Local signals are easy to miss. Google's Places API returns reviews and ratings, and can mark a place as FUTURE_OPENING with an opening date, which is a clean trigger for anyone who sells to businesses with storefronts.

Other signals are easier to buy than to build. Clay's signal triggers cover job changes, job postings, new hires, promotions, news and fundraising, topic intent and website visits. Common Room's job change docs show the best version of the champion play: filter for people whose previous company was a customer.

One hard line. LinkedIn's User Agreement, section 8.2, bans software, scripts and bots that scrape profiles, and bans automated methods to add contacts or send messages. Read posts there as a person, buy job-change data from a licensed source, and never point an agent at someone else's platform to do what its terms forbid.

The loop: enrich, score, draft, approve

Signals arrive as a feed. The loop turns the feed into a short daily queue, small enough for one person to read every draft.

Enrich. For each account that fires a signal, pull the basics: size, location, what they sell, who holds the buying seat, a verified work email. Keep the source of each fact next to it, so the draft can only use what is on file.

Score. Give each account three marks: fit with your ideal customer, 0 to 3; signal strength, 0 to 3; freshness, 0 to 2. Draft only for 6 and above, and change the cut once you see which scores turn into replies.

Draft. One short email per account, first line from the Because Test, one ask. Instantly's benchmark says its best-performing campaigns keep emails under 80 words; treat that as their claim and test it against your own.

Approve. Every draft lands in a queue with the signal, its source and the score beside it. A person approves, edits or kills each one, and the decision goes to the log.

Give the agent that runs this a written brief: inputs, schedule, filters, output file, what needs approval, and the metric it answers to, in the format from how to write a job spec for an AI agent. Its files live in the outbound folder of your growth repo, next to the ideal customer profile and the banned-language file.

If this sounds like a GTM engineer's job, it is. In our October 16, 2026 pull of 55 US GTM engineer postings, 82% mentioned data enrichment and 73% mentioned outbound or sequencing; Marketing Engineer vs GTM Engineer covers who should own it.

The banned-language file

AI outbound fails in a recognizable way. It sounds like a stranger who read your profile for six seconds and decided you are friends.

The fix is a plain text file the drafting agent reads every time, and a check that rejects any draft that breaks it. Ours would start with four sections.

Phrases that never appear: "I hope this finds you well", "quick question", "just following up", "I noticed you", "loved your post" unless a person actually read it, and any line that pretends to a relationship that does not exist.

Claims that need a file: no result, number, customer name or award unless the proof is in the repo and you are allowed to share it. Pressure that is banned: fake deadlines, fake scarcity, "last chance" to someone who never asked for a first one.

Personal lines that are off limits: family, health, location history, anything from a personal account. A signal about the company is fair; a signal about the person's private life is not.

Every time a person kills a draft, they add the reason to this file. After a month, the file is the most valuable thing in the outbound folder.

The rules before the first send

This section is not optional, and it is not legal advice. Have counsel review your program once before launch.

Start with the FTC's CAN-SPAM compliance guide. The law covers all commercial email and makes no exception for business-to-business messages. Each separate email that breaks it can draw a penalty of up to $53,088, per the guide.

The requirements are short. No false headers and no deceptive subject lines; identify the message as an ad; include a valid physical postal address; give a clear way to opt out of all marketing email; keep the opt-out working for at least 30 days after the send; and honor it within 10 business days.

One more line in the guide matters for anyone using an agency or a tool. You cannot contract away responsibility, so the company being promoted and the company sending can both be liable. The rule itself is at 16 CFR Part 316.

Then the inbox providers. Gmail's sender guidelines require SPF or DKIM for everyone, and SPF, DKIM, DMARC and one-click unsubscribe for senders of 5,000 or more messages a day to Gmail, with spam rates kept below 0.3%. Its FAQ says to aim below 0.1% and that enforcement ramped up from November 2025; Yahoo's Sender Hub sets similar authentication rules, and you read your own rate in Postmaster Tools.

Other countries are stricter. Canada's anti-spam law requires consent before commercial email, and a published address implies consent only if there is no "no solicitations" note and the message fits the person's role. The UK's regulator says emails to companies do not need consent under PECR, but sole traders and some partnerships count as individuals, and UK GDPR still applies to named contacts.

Keep the agent away from phones unless a lawyer has signed off. The FCC ruled in February 2024 that AI-generated voices count as artificial under the Telephone Consumer Protection Act. For claims inside the email, the same truth-in-advertising rules apply as anywhere; The FTC Risk Drivers in Marketing Claims lists the ones that cause most trouble.

The approval step: nothing sends without a person

The agent drafts. A person sends. That line does not move, no matter how good the drafts get.

Anthropic's guide to building effective agents describes agents that pause for human feedback at checkpoints, with stopping conditions to keep control. Outbound is the clearest case for a checkpoint, because every message carries your name to a stranger and cannot be unsent.

Make the approval concrete. One named approver per day. Each draft shows the signal, its source link, the score and the eight gate checks above; the approver clicks approve, edit or kill, and the reason for any kill goes into the banned-language file.

Cap the queue. Start at 20 approved messages a day, read every reply yourself for the first two weeks, and raise the cap only when the spam rate and the reply quality both hold.

Measure qualified replies, not messages sent

Messages sent is an activity count. It goes up every time you do more of the thing, whether or not the thing works.

The number that matters is qualified replies: a human reply from an account that fits your ideal customer, either interested or asking a real question, counted once per person. Out-of-office notes, unsubscribes and "wrong person" do not count.

Track it by signal, not just in total. If funding signals produce qualified replies and review signals produce none after a fair run, stop watching reviews. The engine should get narrower before it gets bigger.

Isenberg's example of a month-one case study reads like this: about 75 targeted messages, nine warm replies, three calls booked. Treat that as the shape of a good report, not a benchmark to promise anyone.

Replies that say "not now" go to a nurture track like the lead-first email sequence, with their own consent and opt-out. Replies that say yes go straight to a booked call.

We run our own outbound as an enrichment and sequencing system, with contacts staged across 12 industries. We do not quote reply rates for it here, because a rate without its definition and its dates is noise. If you want one built inside your company, it starts the way our marketing engineering engagements do: one system, one number, then the next, wired into the same CRM work as our CRM and RevOps service.

"This is just spam with better targeting"

Sometimes it is, and the critics are right to say so. Point an agent at a big list with a clever template and you get spam faster, which is why the average reply rate is a low single digit.

The difference is not the targeting tool. It is fewer messages, a real reason in the first line, an opt-out that works, and a person who reads each one before it goes.

So here is the honest test for your own program. If you cannot finish "I am writing because" with a dated fact, if nobody approves the sends, or if your spam rate is climbing, you are running spam with better targeting. Stop, fix that, and then send again.

What to read next, and why

The Founder Content Engine [system]. Your sharpest outbound angles are sitting in your founder's calls. This is how to get them out.

The Customer Truth System [system]. Tells you which pains buyers are naming this week, which tells you which signals to watch.

Book-a-Call Math [data]. Work backward from revenue to how many qualified replies you actually need. The number is usually smaller than people think.

What Is an Agent Harness? [explainer]. The rules, checks and memory around the model are most of what makes an outbound agent safe to run.

So the simple takeaway is this: if you cannot say why you are writing this week, do not write.