What goes wrong when marketing adopts AI informally
Marketing teams were among the first to use AI assistants every day. Research for a sector report, a first draft of a launch page, ten variants of an email subject line: all faster than before. The cost shows up at the editor’s desk. A statistic in paragraph two has no source. A quote attributed to an analyst firm cannot be found. A competitor comparison is a year out of date. The editor now spends the time saved on drafting doing detective work, and occasionally something slips through to publication.
The fix is not to ban the tools or to add another approval layer at the end. It is to record provenance while the research is done, and to put review gates where each kind of error is cheapest to catch.
What the work involves
Research provenance. Writers research with AI assistance, but every claim that might appear in a piece goes into a provenance template: the claim, the primary source, the date, the exact figure or quote, and who checked it. Secondary summaries, including the AI’s own summary, are not sources. Claims without a primary source are dropped or labelled as opinion.
Editorial review gates. We define four gates for the chosen campaign, each with an owner:
- Brief approval — audience, message and sources to use, agreed before drafting.
- Fact-check — every figure and quote in the draft compared with the provenance template.
- Claims review — comparative, product or regulated claims checked by the person who owns them, including legal where needed.
- Final sign-off — the named editor approves publication.
One campaign, end to end. We run a real campaign through the workflow, not a demonstration piece. That shows where the gates are too heavy, where prompts need better examples, and which content types can move faster than others.
The approach comes from the same practice as my book AI for Everyday Automation, which covers research and report workflows with assistants, adapted here for editorial teams with publication risk.
The signature deliverable
You end with a research provenance template, editorial review gates and a campaign workflow pilot. Illustrative example of a provenance template row:
| Claim used | Primary source | Date | Exact figure or quote | Verified by | Status |
|---|---|---|---|---|---|
| Share of buyers researching vendors before contact | Named industry survey, methodology page linked | Publication date | Figure and sample size as published | Writer, then fact-checker | Verified |
| Competitor launched feature X | Competitor release notes | Release date | Feature name as written | Writer | Verified |
| ”Most teams struggle with…” | None | — | — | — | Removed or rewritten as opinion |
Illustrative example. Rows show the format, not real research.
How acceptance is judged
Before the pilot, we sample recent pieces and record time from brief to approved draft, the number of fact-check corrections, rework rounds, and claims published without a primary source. During the pilot we record the same measures on the campaign. The content lead accepts the pilot. The adoption measure is whether writers fill in the provenance template on the next piece without being reminded, and whether editors trust it enough to stop re-researching.
Ownership and handover
Editorial ownership never moves. The content lead owns the template, the gates and the prompt library. Writers are coached to research with provenance as they go. The handover note explains how to adapt the gates for new content types, and which ones can safely be lighter.
Boundaries
This pilot is for teams that publish. If sales needs sourced account research and CRM hygiene, see AI enablement for revenue operations. If product managers are synthesising customer interviews and feedback, the evidence rules are different: see AI enablement for product teams. For a broader view across several business teams, start at practical AI enablement for business teams.