The AI Marketing Playbook for B2B SaaS

Four operating principles

• Build a system, not a stack

  Tool-first adoption produces a Frankenstack where quality depends on who did the work that day. System-first adoption compounds, because every fix improves every future output.

• Delegate by verifiability × blast radius

  Not by difficulty. Two questions decide what AI may own: how easily can a human verify the output, and how much damage does an error do before it is caught?

• Checkpoints are structural, not optional

  In the BCG/Harvard study of 758 professionals, accuracy fell from 84% to 60–70% because people stopped checking the model. The fix is architectural: enforced approval gates, plus AI critics that run before any human sees a draft.

• Context beats prompts

  The difference between mediocre and excellent output is rarely the prompt — it is what the model knows about your ICP, positioning, voice and proof when it starts.

The workflow library

Seven workflows, ordered by strength of evidence, each with the same anatomy and a quality gate:

• W1 — Content & SEO production

  The best-documented workflow — and the one where doing it lazily now carries a measurable penalty.

• W2 — Account research & ABM

  The best documented economics: research depth per account that was never affordable with humans, at pennies per account.

• W3 — Outbound personalisation

  The strongest results and the ugliest documented blowback. The difference is volume discipline.

• W4 — Product marketing & competitive intel

  The quiet compounding win: positioning, win/loss and competitive work as standing capabilities instead of annual projects.

• W5 — Repurposing & social

  The easiest workflow to start with and the easiest to do badly. One dense asset becomes twelve; the voice file decides whether the twelve are worth reading.

• W6 — Analytics & experimentation

  The least glamorous workflow and possibly the highest-leverage one: AI as tireless analyst, human as decision-maker.

• W7 — Lifecycle email

  Last in the library because the public evidence is thinnest — which is itself worth knowing before a vendor tells you otherwise.

Governance that actually holds

The AI usage policy is one page per component, six components, amnesty-first — roughly 78% of people who use AI at work bring unapproved tools, so the policy's first job is visibility, not restriction. Fact-checking is a named-owner table built on the two-list trick: one file is the only pool of statistics AI may cite, the other lists the claims it must never make. And the one number to watch above all is edit depth per AI draft — falling edit depth means either quality rose or reviewers stopped reading, and only a human who still reads carefully can tell you which.

The 24-question self-assessment

Four themes — system and context, workflows and human-in-the-loop, governance and risk, measurement and capability — scored 0–3 for a total out of 72. Bands run from ad hoc (0–24) through emerging and operating to compounding (61–72), which the document itself tells you to treat with suspicion. Score blind, compare with two colleagues, and treat any two-point spread as an alignment finding.