The discipline, written down Book a study →
V.
MMXXVI
Sheet five

Lifecycle and stack

Automation is not a philosophy applied evenly across a lifecycle. Four of the six stages tolerate it completely. Two of them produce worse outcomes when a machine is left alone with them, and knowing which two is a large part of this work.

Lifecycle

Six stages, one measurement layer

Measurement is not a seventh stage. It runs underneath all six, and a stage without instrumentation is a stage nobody can correct.

RefStageDescriptionAutomation
01

Attract

Creating attention on purpose

Outbound, paid, content, events, and the answer layer. The stage everybody photographs and the one that gets blamed when a quarter misses.

Sourcing, enrichment, sequencing, scheduling, distribution, and monitoring what the answer engines say about you. All of it runs at volume without a person in the loop, provided the specification and the tolerance were written first.
Full
02

Capture

Turning attention into a record

Forms, tracking, events, and the routing that decides where a record goes the second it exists.

Automate completely, and never on trust. Every capture path gets a real test event watched from the page to the record before it carries a single live enquiry.
Full, once proven
03

Qualify

Deciding who is worth a person's time

Fit, intent, timing, and the difference between a company that matches your words and a company that matches your business.

This is the first stage that gets worse when a machine is left alone with it. A model scoring fit will confidently qualify people who match the vocabulary and fail the business test, and it will do so at a volume nobody can check by reading. Every scored list gets a false positive hunt using an independent signal, then splits into confirmed and flagged. Only confirmed rows get acted on.
Supervised only
04

Nurture

Staying useful until the buyer is ready

Owned, permission based email and the content that earns the next reply. The compounding stage, and the one companies skip because it pays late.

Sequencing, segmentation, timing and suppression all automate cleanly. The judgement that does not automate is what the message is worth reading for, which belongs upstream in the specification.
Full
05

Convert

The handover to a person

The first human reply, the call, the proposal, and the follow-up that either happens on the day or never.

The second stage that breaks unattended. Speed of first human reply is a measured variable and automating the moment somebody raises a hand is where trust is spent. Automate the routing, the reminder and the record. Never the reply.
Supervised only
06

Expand

Retention, reactivation, referral

The database the company already owns. Usually its least used asset, and the reason a demand problem is often not a demand problem at all.

Reactivation is the cheapest qualified contact a company can get, and it is largely unclaimed as a proposition. Automate the identification and the sequencing, gate the offer.
Full
Exceptions

Qualify and convert

Both failures are quiet, which is what makes them expensive.

An unsupervised qualification step does not error. It produces a clean, plausible, confidently scored list, and the wrong people on it are indistinguishable from the right people until somebody with commercial knowledge reads a sample. By then the list has been mailed from a warmed domain, and the deliverability that took a month to build has been spent on people who were never buyers.

An unsupervised conversion step does not error either. The routing works, the reminder fires, the record updates. What is missing is that a person who raised their hand waited three days for a reply from somebody who cared, and no field in any system records the moment they stopped caring back.

Working rule

Automate the routing, the reminder and the record. Never the reply

Plate V  ·  Profiles and mouldings, measured  ·  Public domain
The stack

Four layers, bought in the wrong order

A marketing stack is not a shopping list. It exists to make the lifecycle measurable and repeatable, and every layer below solves a problem that no layer above it can reach.

RefLayerWhat it is
L1

Capture

Forms, tracking, event definitions, consent, and routing.

Every path is proven with a live test event before it carries anything. A configured connector and a working one are different states, and only one of them has been observed.

L2

Record

The customer relationship system, the data model, and the written definition behind every field.

Fields without written definitions produce reports that two people read differently. The definitions are the asset. The software is replaceable.

L3

Activate

Email, advertising, content publishing, sequences, and the answer layer.

The activation tools get bought first and diagnosed last. Nothing in this layer fixes a problem that lives in the layer above it.

L4

Read

Reporting, attribution, and monitoring what the answer engines say about the brand.

Attribution is contested and acceptable uncertainty is a real engineering answer. What is not acceptable is a figure whose method nobody can describe.

Buying discipline

Diagnose the stage before buying the tool

A recurring purchase in business to business marketing is nurture software, bought by a company that has no audience to nurture. A few hundred marketable contacts is an audience problem, and no amount of sequencing software has ever solved an audience problem.

The order that works is unglamorous. Diagnose which lifecycle stage is actually failing. Confirm the failure with a reading rather than a feeling. Then buy the smallest thing that addresses that stage, prove it end to end, and only then widen.

A stack assembled before the specification exists becomes an annual subscription line that nobody can cancel because nobody can prove it is doing nothing.

What the machine does

Genuinely, at volume

  • Reads at a volume no team can. Whole sites, whole competitor sets, whole back catalogues, every week.
  • Drafts, at a standard that gets through an editor rather than replacing one.
  • Checks mechanically. If a script can detect a fault, a script should, because human review under load misses the same things every time.
  • Watches the answer engines across several models on a fixed cadence, which is manual work nobody was ever going to do consistently.
  • Enriches and classifies at scale, and produces candidates for a person to rule on.
What it does not

However well it is prompted

  • Decide. A model produces candidates, and a named person turns a candidate into a finding.
  • Verify its own work. A model checking a draft it wrote is checking a view it already formed.
  • Replace the specification. A system with no written target produces confident output aimed at nothing.
  • Be trusted on a figure it generated. Every number traces to a source a reader could check, or it comes out.
Next

Every rule was written after something went wrong.

Thirty of them, across measurement, system order, automation, cost and review, published in full.