On this page
In brief
- We mapped our own marketing the way we map a client's operations, and the map surfaced steps nobody would have listed from memory.
- Every step was scored on judgement, repeatability, impact and data to verify against, giving four verdicts, including what stays human.
- The same questions apply to onboarding, fund launches, breach reporting and board MI.
The challenge we shared with our clients
Typically, regulated firms now hold an AI mandate and a budget to match. A clear method for deciding where AI should go is much harder to find. Two patterns tend to follow: a long pause while options are weighed, or automating whatever a vendor demo made look easy.
We had the same challenge in miniature. A founder-led startup, eight marketing channels, competitor and regulatory news to monitor, content to draft, publish, measure and learn from, and a founder whose highest-value hours belong to clients and product. Run manually, that is roughly a 40 hour week of marketing work.
A marketing suite at around £3,000 per month would have imposed someone else's processes on us. Handing everything to AI would have put a language model in charge of our voice and our credibility. Somewhere in the middle was a set of right answers, and we needed a way to find them that would stand up to scrutiny.
Stage one: map how the work actually happens
We mapped our marketing the way we map a client's operations: by sitting with the work itself. Seven high-level processes emerged, from intelligence and research through content, publishing, engagement and performance. Each expanded into detail; the performance process alone held 11 distinct nodes once it was drawn in full.
The value showed up immediately in the nodes nobody would have listed from memory. None of these appear on a marketing suite's feature list, and all of them turned out to matter.
Stage two: score every step, then decide deliberately
Every node was scored on a lens simple enough to survive contact with a board. Four verdicts came out of the scoring.
| Process node | Verdict |
|---|---|
| Competitor and news sweeps | Automate Scheduled research with structured, sourced findings |
| Drafting for eight channels | AI drafts, a person approves Nothing publishes without sign-off |
| Capturing performance numbers | Assisted The platforms expose little, so every figure records its source |
| First-hour comment engagement | Keep manual, by design High judgement, poor automation candidate |
The rows that surprise people are the useful ones. The scoring showed what we could automate and, just as clearly, what is better kept with a person. For a regulated firm, that second list is often the one the board values most.
Three of our first-pass scores changed on contact with reality, and the case study names them. A method stays useful when it keeps revising its own scores against what it finds.
The full case study is free to download
22 pages: the complete 24 row decision table, the operating model the scores produced, how the platform was built in 7 working days under mockup-first governance, the guardrail engine refusing an invented statistic, what stayed human on purpose, and the real numbers, including what is deliberately unclaimed.
Download the case studyWhat this means beyond marketing
Marketing was simply the process we could publish, because it is our own. Swap the word for client onboarding, fund launches, breach reporting or board MI and the questions are identical: where does AI belong in this process, what stays human, and how do you evidence the decision?
That is what a four-week engagement answers, and the method is published in full.